Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix_quan_hang

test=develop

cmake/external/boost.cmake

@@ -24,7 +24,7 @@ set(BOOST_PROJECT       "extern_boost")
 # So we use 1.41.0 here.
 set(BOOST_VER           "1.41.0")
 set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
-set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
+set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
 
 MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
 

cmake/external/grpc.cmake

@@ -44,7 +44,7 @@ ExternalProject_Add(
     # 3. keep only zlib, cares, protobuf, boringssl under "third_party",
     #    checkout and clean other dirs under third_party
     # 4. remove .git, and package the directory.
-    URL "http://paddlepaddledeps.cdn.bcebos.com/grpc-v1.10.x.tar.gz"
+    URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.10.x.tar.gz"
     URL_MD5  "1f268a2aff6759839dccd256adcc91cf"
     PREFIX          ${GRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""

cmake/external/mklml.cmake

@@ -34,7 +34,7 @@ SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
 SET(TIME_VERSION "2019.0.1.20181227")
 IF(WIN32)
     SET(MKLML_VER "mklml_win_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
+    SET(MKLML_URL "https://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/mklml.lib)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5md.lib)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/mklml.dll)
@@ -43,7 +43,7 @@ ELSE()
     #TODO(intel-huying):
     #  Now enable Erf function in mklml library temporarily, it will be updated as offical version later.
     SET(MKLML_VER "Glibc225_vsErf_mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
+    SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/libmklml_intel.so)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5.so)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/libmklml_intel.so)

paddle/fluid/API.spec

@@ -10,6 +10,9 @@ paddle.fluid.default_startup_program (ArgSpec(args=[], varargs=None, keywords=No
 paddle.fluid.default_main_program (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', '5430f54ab4895f9f47db6bebbaf71659'))
 paddle.fluid.program_guard (ArgSpec(args=['main_program', 'startup_program'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b54f403e57825a1592aece03afe3afb6'))
 paddle.fluid.name_scope (ArgSpec(args=['prefix'], varargs=None, keywords=None, defaults=(None,)), ('document', '0ef753f5cec69fef9ae6ad8b867b33a2'))
+paddle.fluid.cuda_places (ArgSpec(args=['device_ids'], varargs=None, keywords=None, defaults=(None,)), ('document', '7d9a51fc9cf3c5245b5227080a8064c3'))
+paddle.fluid.cpu_places (ArgSpec(args=['device_count'], varargs=None, keywords=None, defaults=(None,)), ('document', '4c0cd83f0b401fc2ff84c70974e5d210'))
+paddle.fluid.cuda_pinned_places (ArgSpec(args=['device_count'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd0c3ebd813c39958c92b78e3eef7e912'))
 paddle.fluid.Executor.__init__ (ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.Executor.close (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'f5369953dd0c443961cf79f7a00e1a03'))
 paddle.fluid.Executor.run (ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False)), ('document', 'f482e93b38b4018796969a2e1dde479d'))
@@ -44,7 +47,7 @@ paddle.fluid.AsyncExecutor.run (ArgSpec(args=['self', 'program', 'data_feed', 'f
 paddle.fluid.AsyncExecutor.save_model (ArgSpec(args=['self', 'save_path'], varargs=None, keywords=None, defaults=None), ('document', 'c8ac0dfcb3b187aba25d03af7fea56b2'))
 paddle.fluid.AsyncExecutor.stop (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '5f23d043607bb5d55e466ec3f578e093'))
 paddle.fluid.CompiledProgram.__init__ (ArgSpec(args=['self', 'program_or_graph'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.CompiledProgram.with_data_parallel (ArgSpec(args=['self', 'loss_name', 'build_strategy', 'exec_strategy', 'share_vars_from'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'e1af7fd53cf868554f312779fc803864'))
+paddle.fluid.CompiledProgram.with_data_parallel (ArgSpec(args=['self', 'loss_name', 'build_strategy', 'exec_strategy', 'share_vars_from', 'places'], varargs=None, keywords=None, defaults=(None, None, None, None, None)), ('document', 'a8c7793803cf976680d9478e378fa356'))
 paddle.fluid.CompiledProgram.with_inference_optimize (ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=None), ('document', '9e5b009d850191a010e859189c127fd8'))
 paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.ExecutionStrategy) -> None
 paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.GradientScaleStrategy, arg0: int) -> None
@@ -58,6 +61,12 @@ paddle.fluid.io.load_params (ArgSpec(args=['executor', 'dirname', 'main_program'
 paddle.fluid.io.load_persistables (ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)), ('document', '28df5bfe26ca7a077f91156abb0fe6d2'))
 paddle.fluid.io.save_inference_model (ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True)), ('document', '70f4f53f13572436ac72d1c8b5efeb9d'))
 paddle.fluid.io.load_inference_model (ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename', 'pserver_endpoints'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '7a5255386075dac3c75b7058254fcdcb'))
+paddle.fluid.io.PyReader.__init__ (ArgSpec(args=['self', 'feed_list', 'capacity', 'use_double_buffer', 'iterable'], varargs=None, keywords=None, defaults=(True, False)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
+paddle.fluid.io.PyReader.decorate_batch_generator (ArgSpec(args=['self', 'reader', 'places'], varargs=None, keywords=None, defaults=(None,)), ('document', 'a3fefec8bacd6ce83f49906a9d05e779'))
+paddle.fluid.io.PyReader.decorate_sample_generator (ArgSpec(args=['self', 'sample_generator', 'batch_size', 'drop_last', 'places'], varargs=None, keywords=None, defaults=(True, None)), ('document', '7abd9cf7d695bab5bb6cf7ded5903cb2'))
+paddle.fluid.io.PyReader.decorate_sample_list_generator (ArgSpec(args=['self', 'reader', 'places'], varargs=None, keywords=None, defaults=(None,)), ('document', 'faef298f73e91aedcfaf5d184f3109b7'))
+paddle.fluid.io.PyReader.reset (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'ff1cc1e2beb8824d453656c72c28ddfb'))
+paddle.fluid.io.PyReader.start (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'b7ea0a548991924e4cfe61a577b8e56d'))
 paddle.fluid.initializer.ConstantInitializer.__init__ (ArgSpec(args=['self', 'value', 'force_cpu'], varargs=None, keywords=None, defaults=(0.0, False)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.initializer.UniformInitializer.__init__ (ArgSpec(args=['self', 'low', 'high', 'seed'], varargs=None, keywords=None, defaults=(-1.0, 1.0, 0)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.initializer.NormalInitializer.__init__ (ArgSpec(args=['self', 'loc', 'scale', 'seed'], varargs=None, keywords=None, defaults=(0.0, 1.0, 0)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -222,6 +231,7 @@ paddle.fluid.layers.teacher_student_sigmoid_loss (ArgSpec(args=['input', 'label'
 paddle.fluid.layers.huber_loss (ArgSpec(args=['input', 'label', 'delta'], varargs=None, keywords=None, defaults=None), ('document', '431a4301c35032166ec029f7432c80a7'))
 paddle.fluid.layers.tree_conv (ArgSpec(args=['nodes_vector', 'edge_set', 'output_size', 'num_filters', 'max_depth', 'act', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(1, 2, 'tanh', None, None, None)), ('document', '34ea12ac9f10a65dccbc50100d12e607'))
 paddle.fluid.layers.npair_loss (ArgSpec(args=['anchor', 'positive', 'labels', 'l2_reg'], varargs=None, keywords=None, defaults=(0.002,)), ('document', '46994d10276dd4cb803b4062b5d14329'))
+paddle.fluid.layers.fsp_matrix (ArgSpec(args=['x', 'y'], varargs=None, keywords=None, defaults=None), ('document', 'b76ccca3735bea4a58a0dbf0d77c5393'))
 paddle.fluid.layers.data (ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)), ('document', '33bbd42027d872b3818b3d64ec52e139'))
 paddle.fluid.layers.open_files (ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None)), ('document', 'b1ae2e1cc0750e58726374061ea90ecc'))
 paddle.fluid.layers.read_file (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', 'b0a1c2fc51c27a106da28f3308c41f5e'))
@@ -229,7 +239,7 @@ paddle.fluid.layers.shuffle (ArgSpec(args=['reader', 'buffer_size'], varargs=Non
 paddle.fluid.layers.batch (ArgSpec(args=['reader', 'batch_size'], varargs=None, keywords=None, defaults=None), ('document', 'f563d376d35e1a4c4db100fd11b381a0'))
 paddle.fluid.layers.double_buffer (ArgSpec(args=['reader', 'place', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', '07e5b796674796eb1ef3fee9c10d24e3'))
 paddle.fluid.layers.random_data_generator (ArgSpec(args=['low', 'high', 'shapes', 'lod_levels', 'for_parallel'], varargs=None, keywords=None, defaults=(True,)), ('document', '9b7f0f86ec24bbc97643cadcb6499cff'))
-paddle.fluid.layers.py_reader (ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)), ('document', '13dabc57863f62ab3141586784ee356b'))
+paddle.fluid.layers.py_reader (ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)), ('document', '4357643685cfd65454ba5a15f0151709'))
 paddle.fluid.layers.create_py_reader_by_data (ArgSpec(args=['capacity', 'feed_list', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, True)), ('document', '350f74d93fab9adb2ac4950f1c26416b'))
 paddle.fluid.layers.Preprocessor.__init__ (ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.Preprocessor.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -255,6 +265,7 @@ paddle.fluid.layers.reverse (ArgSpec(args=['x', 'axis'], varargs=None, keywords=
 paddle.fluid.layers.has_inf (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '8f8c0306117ea441f20dcbbdba1f0ecc'))
 paddle.fluid.layers.has_nan (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '2e53e83127dbfd86e7098bdfe9a549e8'))
 paddle.fluid.layers.isfinite (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '0a437011c3906079fd8947ed3e52d292'))
+paddle.fluid.layers.range (ArgSpec(args=['start', 'end', 'step', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '2ec937ede953ded2fdff2675883900bb'))
 paddle.fluid.layers.While.__init__ (ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.While.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.Switch.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -376,23 +387,9 @@ paddle.fluid.contrib.Calibrator.__init__ (ArgSpec(args=['self'], varargs='args',
 paddle.fluid.contrib.Calibrator.sample_data (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '3b8c85ca1e2cf753cc8c90a6c6992958'))
 paddle.fluid.contrib.Calibrator.save_int8_model (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.contrib.reader.ctr_reader.ctr_reader (ArgSpec(args=['feed_dict', 'file_type', 'file_format', 'dense_slot_index', 'sparse_slot_index', 'capacity', 'thread_num', 'batch_size', 'file_list', 'slots', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b2ebf3de2a6ef1af2c3b88d2db7591ab'))
-paddle.fluid.contrib.build_compressor (ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.CompressPass.__init__ (ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.CompressPass.add_strategy (ArgSpec(args=['self', 'strategy'], varargs=None, keywords=None, defaults=None), ('document', '3bf6010b6f47d3c86df0ec8957be95e0'))
-paddle.fluid.contrib.CompressPass.apply (ArgSpec(args=['self', 'graph'], varargs=None, keywords=None, defaults=None), ('document', 'a92bf85d4b59bd4f2ac1706d7c4899a6'))
-paddle.fluid.contrib.ImitationGraph.__init__ (ArgSpec(args=['self', 'program'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.ImitationGraph.all_parameters (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.SensitivePruneStrategy.__init__ (ArgSpec(args=['self', 'pruner', 'start_epoch', 'end_epoch', 'delta_rate', 'acc_loss_threshold', 'sensitivities'], varargs=None, keywords=None, defaults=(None, 0, 10, 0.2, 0.2, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.SensitivePruneStrategy.on_batch_begin (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.SensitivePruneStrategy.on_batch_end (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.SensitivePruneStrategy.on_compress_begin (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.SensitivePruneStrategy.on_compress_end (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_begin (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.SensitivePruneStrategy.on_epoch_end (ArgSpec(args=['self', 'context'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.MagnitudePruner.__init__ (ArgSpec(args=['self', 'threshold'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.MagnitudePruner.prune (ArgSpec(args=['self', 'param', 'threshold'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.RatioPruner.__init__ (ArgSpec(args=['self', 'ratios'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e7a81a325b296a9ca502ee5adb4fc85d'))
-paddle.fluid.contrib.RatioPruner.prune (ArgSpec(args=['self', 'param', 'ratio'], varargs=None, keywords=None, defaults=(None,)), ('document', '358cbf2978c91028fb96a195a9884645'))
+paddle.fluid.contrib.Compressor.__init__ (ArgSpec(args=['self', 'place', 'scope', 'train_program', 'train_reader', 'train_feed_list', 'train_fetch_list', 'eval_program', 'eval_reader', 'eval_feed_list', 'eval_fetch_list', 'teacher_programs', 'checkpoint_path', 'train_optimizer', 'distiller_optimizer'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, [], './checkpoints', None, None)), ('document', '31ae143830c9bf6b43547dd546c5ba80'))
+paddle.fluid.contrib.Compressor.config (ArgSpec(args=['self', 'config_file'], varargs=None, keywords=None, defaults=None), ('document', '780d9c007276ccbb95b292400d7807b0'))
+paddle.fluid.contrib.Compressor.run (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', 'c6e43d6a078d307672283c1f36e04fe9'))
 paddle.fluid.contrib.load_persistables_for_increment (ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var', 'lookup_table_var_path'], varargs=None, keywords=None, defaults=None), ('document', '2ab36d4f7a564f5f65e455807ad06c67'))
 paddle.fluid.contrib.load_persistables_for_inference (ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var_name'], varargs=None, keywords=None, defaults=None), ('document', '59066bac9db0ac6ce414d05780b7333f'))
 paddle.fluid.contrib.convert_dist_to_sparse_program (ArgSpec(args=['program'], varargs=None, keywords=None, defaults=None), ('document', '74c39c595dc70d6be2f16d8e462d282b'))
@@ -432,48 +429,59 @@ paddle.fluid.nets.img_conv_group (ArgSpec(args=['input', 'conv_num_filter', 'poo
 paddle.fluid.optimizer.SGDOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'regularization', 'name'], varargs=None, keywords=None, defaults=(None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.SGDOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.SGDOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.SGDOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.SGDOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.MomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'use_nesterov', 'regularization', 'name'], varargs=None, keywords=None, defaults=(False, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.MomentumOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.MomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.MomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.MomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.AdagradOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name', 'initial_accumulator_value'], varargs=None, keywords=None, defaults=(1e-06, None, None, 0.0)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdagradOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.AdagradOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.AdagradOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdagradOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.AdamOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name', 'lazy_mode'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None, False)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdamOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.AdamOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.AdamOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdamOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.AdamaxOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdamaxOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.AdamaxOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.AdamaxOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdamaxOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'decay', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.95, 1e-06, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.DecayedAdagradOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.DecayedAdagradOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.FtrlOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'l1', 'l2', 'lr_power', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.0, 0.0, -0.5, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.FtrlOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.FtrlOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.FtrlOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.FtrlOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.RMSPropOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'rho', 'epsilon', 'momentum', 'centered', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.95, 1e-06, 0.0, False, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.RMSPropOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.RMSPropOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.RMSPropOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.RMSPropOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.AdadeltaOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'epsilon', 'rho', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, 0.95, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdadeltaOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.AdadeltaOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.AdadeltaOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.AdadeltaOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.ModelAverage.__init__ (ArgSpec(args=['self', 'average_window_rate', 'min_average_window', 'max_average_window', 'regularization', 'name'], varargs=None, keywords=None, defaults=(10000, 10000, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.ModelAverage.apply (ArgSpec(args=['self', 'executor', 'need_restore'], varargs=None, keywords=None, defaults=(True,)), ('document', '46234a5470590feb336346f70a3db715'))
 paddle.fluid.optimizer.ModelAverage.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.ModelAverage.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.ModelAverage.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.ModelAverage.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.optimizer.ModelAverage.restore (ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None), ('document', '18db9c70be9c4dd466f9844457b21bfe'))
 paddle.fluid.optimizer.LarsMomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'lars_coeff', 'lars_weight_decay', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.0005, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.LarsMomentumOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
 paddle.fluid.optimizer.LarsMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
+paddle.fluid.optimizer.LarsMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.optimizer.LarsMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
 paddle.fluid.backward.append_backward (ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '1a79bd7d10ae54ca763ec81bca36ba24'))
 paddle.fluid.regularizer.L1DecayRegularizer.__init__ (ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@@ -512,6 +520,7 @@ paddle.fluid.unique_name.guard (ArgSpec(args=['new_generator'], varargs=None, ke
 paddle.fluid.recordio_writer.convert_reader_to_recordio_file (ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', '65c7523e86f0c50bb729b01667f36310'))
 paddle.fluid.recordio_writer.convert_reader_to_recordio_files (ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', 'bc643f0f5f1b9db57ff0d8a57d379bd7'))
 paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope
+paddle.reader.cache (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', '1676886070eb607cb608f7ba47be0d3c'))
 paddle.reader.map_readers (ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None), ('document', '77cbadb09df588e21e5cc0819b69c87d'))
 paddle.reader.buffered (ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None), ('document', '0d6186f109feceb99f60ec50a0a624cb'))
 paddle.reader.compose (ArgSpec(args=[], varargs='readers', keywords='kwargs', defaults=None), ('document', '884291104e1c3f37f33aae44b7deeb0d'))

paddle/fluid/imperative/layer.cc

@@ -315,6 +315,9 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
       for (size_t i = 0; i < outputs.size(); ++i) {
         framework::Variable* grad = outputs[i]->var_;
         framework::Variable* orig_grad = origin_outputs[i]->var_;
+        VLOG(3) << "AddTo Called with orig_grad is: "
+                << origin_outputs[i]->name_ << " Grad to be added is "
+                << outputs[i]->name_;
         AddTo(grad, orig_grad, place_);
         delete grad;
       }

paddle/fluid/imperative/tracer.cc

@@ -277,6 +277,7 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
           VarBase* var = current_vars_map[var_it->second];
           InitGrad(var, prepared_op.GetDeviceContext());
           grad_out_vars.push_back(var->grads_);
+          VLOG(3) << "grads output var name: " << var->name_;
         }
       }
     }

paddle/fluid/inference/api/demo_ci/run.sh

@@ -27,7 +27,7 @@ if [ -d "$TENSORRT_INCLUDE_DIR" -a -d "$TENSORRT_LIB_DIR" ]; then
 fi
 
 PREFIX=inference-vis-demos%2F
-URL_ROOT=http://paddlemodels.cdn.bcebos.com/${PREFIX}
+URL_ROOT=http://paddlemodels.bj.bcebos.com/${PREFIX}
 
 # download vis_demo data
 function download() {

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -115,14 +115,14 @@ inference_analysis_test(test_analyzer_transformer SRCS analyzer_transformer_test
 # ocr
 set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr")
 if (NOT EXISTS ${OCR_INSTALL_DIR})
-    inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz")
+    inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.bj.bcebos.com/" "inference-vis-demos%2Focr.tar.gz")
 endif()
 inference_analysis_api_test_with_refer_result(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc SERIAL)
 
 # mobilenet with transpose op
 set(MOBILENET_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet")
 if (NOT EXISTS ${MOBILENET_INSTALL_DIR})
-    inference_download_and_uncompress(${MOBILENET_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Fmobilenet.tar.gz")
+    inference_download_and_uncompress(${MOBILENET_INSTALL_DIR} "http://paddlemodels.bj.bcebos.com/" "inference-vis-demos%2Fmobilenet.tar.gz")
 endif()
 inference_analysis_api_test_with_refer_result(test_analyzer_mobilenet_transpose ${MOBILENET_INSTALL_DIR} analyzer_vis_tester.cc SERIAL)
 

paddle/fluid/inference/tests/test.cmake

 include(ExternalProject)
-set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url")
+set(INFERENCE_URL "http://paddle-inference-dist.bj.bcebos.com" CACHE STRING "inference download url")
 set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
     "A path setting inference demo download directories.")
 

paddle/fluid/memory/allocation/CMakeLists.txt

@@ -61,4 +61,6 @@ nv_test(allocation_and_eigen_test SRCS allocation_and_eigen_test.cu DEPS allocat
 
 cc_test(retry_allocator_test SRCS retry_allocator_test.cc DEPS retry_allocator best_fit_allocator locked_allocator cpu_allocator)
 
-cc_test(allocator_facade_test SRCS allocator_facade_test.cc DEPS allocator_facade)
+cc_test(allocator_facade_abs_flags_test SRCS allocator_facade_abs_flags_test.cc DEPS allocator_facade)
+
+cc_test(allocator_facade_frac_flags_test SRCS allocator_facade_frac_flags_test.cc DEPS allocator_facade)

paddle/fluid/memory/allocation/allocator_facade_abs_flags_test.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/memory/allocation/allocator_facade.h"
+#include <gflags/gflags.h>
+#include <gtest/gtest.h>
+
+#ifdef PADDLE_WITH_CUDA
+DECLARE_double(fraction_of_gpu_memory_to_use);
+DECLARE_double(fraction_of_cuda_pinned_memory_to_use);
+DECLARE_uint64(initial_gpu_memory_in_mb);
+DECLARE_uint64(reallocate_gpu_memory_in_mb);
+DECLARE_int64(gpu_allocator_retry_time);
+#endif
+
+namespace paddle {
+namespace memory {
+namespace allocation {
+
+//! Run allocate test cases for different places
+void AllocateTestCases() {
+  auto &instance = AllocatorFacade::Instance();
+  platform::Place place;
+  size_t size = 1024;
+
+  {
+    place = platform::CPUPlace();
+    size = 1024;
+    auto cpu_allocation = instance.Alloc(place, size);
+    ASSERT_NE(cpu_allocation, nullptr);
+    ASSERT_NE(cpu_allocation->ptr(), nullptr);
+    ASSERT_EQ(cpu_allocation->place(), place);
+    ASSERT_EQ(cpu_allocation->size(), size);
+  }
+
+#ifdef PADDLE_WITH_CUDA
+  {
+    place = platform::CUDAPlace(0);
+    size = 1024;
+    auto gpu_allocation = instance.Alloc(place, size);
+    ASSERT_NE(gpu_allocation, nullptr);
+    ASSERT_NE(gpu_allocation->ptr(), nullptr);
+    ASSERT_EQ(gpu_allocation->place(), place);
+    ASSERT_GE(gpu_allocation->size(), size);
+  }
+
+  {
+    // Allocate 2GB gpu memory
+    place = platform::CUDAPlace(0);
+    size = 2 * static_cast<size_t>(1 << 30);
+    auto gpu_allocation = instance.Alloc(place, size);
+    ASSERT_NE(gpu_allocation, nullptr);
+    ASSERT_NE(gpu_allocation->ptr(), nullptr);
+    ASSERT_EQ(gpu_allocation->place(), place);
+    ASSERT_GE(gpu_allocation->size(), size);
+  }
+
+  {
+    place = platform::CUDAPinnedPlace();
+    size = (1 << 20);
+    auto cuda_pinned_allocation =
+        instance.Alloc(platform::CUDAPinnedPlace(), 1 << 20);
+    ASSERT_NE(cuda_pinned_allocation, nullptr);
+    ASSERT_NE(cuda_pinned_allocation->ptr(), nullptr);
+    ASSERT_EQ(cuda_pinned_allocation->place(), place);
+    ASSERT_GE(cuda_pinned_allocation->size(), size);
+  }
+#endif
+}
+
+TEST(Allocator, SpecifyGpuMemory) {
+#ifdef PADDLE_WITH_CUDA
+  // Set to 0.0 to test FLAGS_initial_gpu_memory_in_mb and
+  // FLAGS_reallocate_gpu_memory_in_mb
+  FLAGS_fraction_of_gpu_memory_to_use = 0.0;
+  // 512 MB
+  FLAGS_initial_gpu_memory_in_mb = 512;
+  // 4 MB
+  FLAGS_reallocate_gpu_memory_in_mb = 4;
+  FLAGS_gpu_allocator_retry_time = 500;
+  FLAGS_fraction_of_cuda_pinned_memory_to_use = 0.5;
+#endif
+
+  AllocateTestCases();
+}
+
+}  // namespace allocation
+}  // namespace memory
+}  // namespace paddle

paddle/fluid/memory/allocation/allocator_facade_test.cc → paddle/fluid/memory/allocation/allocator_facade_frac_flags_test.cc

@@ -19,6 +19,8 @@
 #ifdef PADDLE_WITH_CUDA
 DECLARE_double(fraction_of_gpu_memory_to_use);
 DECLARE_double(fraction_of_cuda_pinned_memory_to_use);
+DECLARE_uint64(initial_gpu_memory_in_mb);
+DECLARE_uint64(reallocate_gpu_memory_in_mb);
 DECLARE_int64(gpu_allocator_retry_time);
 #endif
 
@@ -26,13 +28,8 @@ namespace paddle {
 namespace memory {
 namespace allocation {
 
-TEST(allocator, allocator) {
-#ifdef PADDLE_WITH_CUDA
-  FLAGS_fraction_of_gpu_memory_to_use = 0.01;
-  FLAGS_gpu_allocator_retry_time = 500;
-  FLAGS_fraction_of_cuda_pinned_memory_to_use = 0.5;
-#endif
-
+//! Run allocate test cases for different places
+void AllocateTestCases() {
   auto &instance = AllocatorFacade::Instance();
   platform::Place place;
   size_t size = 1024;
@@ -82,6 +79,16 @@ TEST(allocator, allocator) {
 #endif
 }
 
+TEST(Allocator, Allocator) {
+#ifdef PADDLE_WITH_CUDA
+  FLAGS_fraction_of_gpu_memory_to_use = 0.01;
+  FLAGS_gpu_allocator_retry_time = 500;
+  FLAGS_fraction_of_cuda_pinned_memory_to_use = 0.5;
+#endif
+
+  AllocateTestCases();
+}
+
 }  // namespace allocation
 }  // namespace memory
 }  // namespace paddle

paddle/fluid/memory/allocation/legacy_allocator.cc

@@ -37,6 +37,8 @@ DEFINE_bool(init_allocated_mem, false,
             "that initializing the allocated memory with a small value "
             "during unit testing.");
 DECLARE_double(fraction_of_gpu_memory_to_use);
+DECLARE_uint64(initial_gpu_memory_in_mb);
+DECLARE_uint64(reallocate_gpu_memory_in_mb);
 DECLARE_bool(benchmark);
 
 namespace paddle {
@@ -153,12 +155,18 @@ BuddyAllocator *GetGPUBuddyAllocator(int gpu_id) {
                                     platform::GpuMinChunkSize(),
                                     platform::GpuMaxChunkSize());
 
-      VLOG(10) << "\n\nNOTE: each GPU device use "
-               << FLAGS_fraction_of_gpu_memory_to_use * 100
-               << "% of GPU memory.\n"
-               << "You can set GFlags environment variable '"
-               << "FLAGS_fraction_of_gpu_memory_to_use"
-               << "' to change the fraction of GPU usage.\n\n";
+      VLOG(10) << "\n\nNOTE:\n"
+               << "You can set GFlags environment variable "
+               << "'FLAGS_fraction_of_gpu_memory_to_use' "
+               << "or 'FLAGS_initial_gpu_memory_in_mb' "
+               << "or 'FLAGS_reallocate_gpu_memory_in_mb' "
+               << "to change the memory size for GPU usage.\n"
+               << "Current 'FLAGS_fraction_of_gpu_memory_to_use' value is "
+               << FLAGS_fraction_of_gpu_memory_to_use
+               << ". Current 'FLAGS_initial_gpu_memory_in_mb' value is "
+               << FLAGS_initial_gpu_memory_in_mb
+               << ". Current 'FLAGS_reallocate_gpu_memory_in_mb' value is "
+               << FLAGS_reallocate_gpu_memory_in_mb << "\n\n";
     }
   });
 

paddle/fluid/memory/detail/CMakeLists.txt

@@ -9,3 +9,5 @@ endif(${WITH_GPU})
 cc_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator)
 
 cc_library(buddy_allocator SRCS buddy_allocator.cc DEPS memory_block system_allocator glog)
+
+cc_test(buddy_allocator_test SRCS buddy_allocator_test.cc DEPS buddy_allocator)

paddle/fluid/memory/detail/buddy_allocator.cc

@@ -13,6 +13,10 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/memory/detail/buddy_allocator.h"
+
+#include <algorithm>
+#include <utility>
+
 #include "glog/logging.h"
 
 DEFINE_bool(free_idle_memory, false,
@@ -36,9 +40,10 @@ BuddyAllocator::~BuddyAllocator() {
               "have actually been freed";
   while (!pool_.empty()) {
     auto block = static_cast<MemoryBlock*>(std::get<2>(*pool_.begin()));
-    VLOG(10) << "Free from block (" << block << ", " << max_chunk_size_ << ")";
+    VLOG(10) << "Free from block (" << block << ", " << block->size(cache_)
+             << ")";
 
-    system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
+    system_allocator_->Free(block, block->size(cache_), block->index(cache_));
     cache_.invalidate(block);
     pool_.erase(pool_.begin());
   }
@@ -71,7 +76,7 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) {
 
   // refill the pool if failure
   if (it == pool_.end()) {
-    it = RefillPool();
+    it = RefillPool(size);
     // if still failure, fail fatally
     if (it == pool_.end()) {
       return nullptr;
@@ -184,19 +189,28 @@ void* BuddyAllocator::SystemAlloc(size_t size) {
   return static_cast<MemoryBlock*>(p)->data();
 }
 
-BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
+BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool(
+    size_t request_bytes) {
+  size_t allocate_bytes = max_chunk_size_;
+  size_t index = 0;
+
 #ifdef PADDLE_WITH_CUDA
   if (system_allocator_->UseGpu()) {
     if ((total_used_ + total_free_) == 0) {
-      // Compute the maximum allocation size for the first allocation.
-      max_chunk_size_ = platform::GpuMaxChunkSize();
+      // Compute the allocation size for gpu for the first allocation.
+      allocate_bytes = std::max(platform::GpuInitAllocSize(), request_bytes);
+    } else {
+      // Reallocation size
+      if (realloc_size_ == 0) {
+        realloc_size_ = platform::GpuReallocSize();
+      }
+      allocate_bytes = std::max(realloc_size_, request_bytes);
     }
   }
 #endif
 
-  // Allocate a new maximum sized block
-  size_t index = 0;
-  void* p = system_allocator_->Alloc(&index, max_chunk_size_);
+  // Allocate a new block
+  void* p = system_allocator_->Alloc(&index, allocate_bytes);
 
   if (p == nullptr) return pool_.end();
 
@@ -204,7 +218,7 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
            << " from system allocator";
 
   static_cast<MemoryBlock*>(p)->init(&cache_, MemoryBlock::FREE_CHUNK, index,
-                                     max_chunk_size_, nullptr, nullptr);
+                                     allocate_bytes, nullptr, nullptr);
 
   // gpu fallback allocation
   if (system_allocator_->UseGpu() &&
@@ -212,10 +226,10 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
     fallback_alloc_count_++;
   }
 
-  total_free_ += max_chunk_size_;
+  total_free_ += allocate_bytes;
 
   // dump the block into pool
-  return pool_.insert(IndexSizeAddress(index, max_chunk_size_, p)).first;
+  return pool_.insert(IndexSizeAddress(index, allocate_bytes, p)).first;
 }
 
 BuddyAllocator::PoolSet::iterator BuddyAllocator::FindExistChunk(size_t size) {
@@ -286,12 +300,12 @@ void BuddyAllocator::CleanIdleFallBackAlloc() {
 
     VLOG(10) << "Return block " << block << " to fallback allocator.";
 
-    system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
+    system_allocator_->Free(block, block->size(cache_), block->index(cache_));
     cache_.invalidate(block);
 
     pool = PoolSet::reverse_iterator(pool_.erase(std::next(pool).base()));
 
-    total_free_ -= max_chunk_size_;
+    total_free_ -= block->size(cache_);
     fallback_alloc_count_--;
 
     // If no fall allocation exists, return directly
@@ -322,12 +336,12 @@ void BuddyAllocator::CleanIdleNormalAlloc() {
 
     VLOG(10) << "Return block " << block << " to base allocator.";
 
-    system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
+    system_allocator_->Free(block, block->size(cache_), block->index(cache_));
     cache_.invalidate(block);
 
     pool = PoolSet::reverse_iterator(pool_.erase(std::next(pool).base()));
 
-    total_free_ -= max_chunk_size_;
+    total_free_ -= block->size(cache_);
 
     if (!shall_free_alloc()) return;
   }

paddle/fluid/memory/detail/buddy_allocator.h

@@ -60,7 +60,7 @@ class BuddyAllocator {
   void* SystemAlloc(size_t size);
 
   /*! \brief If existing chunks are not suitable, refill pool */
-  PoolSet::iterator RefillPool();
+  PoolSet::iterator RefillPool(size_t request_bytes);
 
   /**
    *  \brief   Find the suitable chunk from existing pool and split
@@ -89,6 +89,8 @@ class BuddyAllocator {
   size_t min_chunk_size_;  // the minimum size of each chunk
   size_t max_chunk_size_;  // the maximum size of each chunk
 
+  size_t realloc_size_ = 0;  // the size of re-allocated chunk
+
  private:
   /**
    * \brief A list of free allocation

paddle/fluid/memory/detail/buddy_allocator_test.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/memory/detail/buddy_allocator.h"
+
+#include <memory>
+
+#include "gflags/gflags.h"
+#include "gtest/gtest.h"
+#include "paddle/fluid/memory/detail/system_allocator.h"
+#include "paddle/fluid/platform/gpu_info.h"
+
+#ifdef PADDLE_WITH_CUDA
+DECLARE_double(fraction_of_gpu_memory_to_use);
+DECLARE_uint64(initial_gpu_memory_in_mb);
+DECLARE_uint64(reallocate_gpu_memory_in_mb);
+#endif
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+constexpr static int test_gpu_id = 0;
+
+void TestBuddyAllocator(BuddyAllocator* allocator, size_t size_bytes) {
+  bool freed = false;
+  size_t used_bytes = allocator->Used();
+
+  if (size_bytes > 0) {
+    void* p = allocator->Alloc(size_bytes);
+
+    EXPECT_NE(p, nullptr);
+#ifdef PADDLE_WITH_CUDA
+    if (size_bytes < platform::GpuMaxChunkSize()) {
+#else
+    if (size_bytes < platform::CpuMaxChunkSize()) {
+#endif
+      // Not allocate from SystemAllocator
+      EXPECT_GE(allocator->Used(), used_bytes + size_bytes);
+    } else {
+      // Allocate from SystemAllocator doesn't count in Used()
+      EXPECT_EQ(allocator->Used(), used_bytes);
+    }
+
+    int* intp = static_cast<int*>(p);
+    std::shared_ptr<int> ptr(intp, [&](void* p) {
+      allocator->Free(intp);
+      freed = true;
+    });
+  } else {
+    freed = true;
+  }
+
+  EXPECT_EQ(used_bytes, allocator->Used());
+  EXPECT_TRUE(freed);
+}
+
+#ifdef PADDLE_WITH_CUDA
+TEST(BuddyAllocator, GpuFraction) {
+  FLAGS_fraction_of_gpu_memory_to_use = 0.01;
+
+  BuddyAllocator buddy_allocator(
+      std::unique_ptr<SystemAllocator>(new GPUAllocator(test_gpu_id)),
+      platform::GpuMinChunkSize(), platform::GpuMaxChunkSize());
+
+  TestBuddyAllocator(&buddy_allocator, 10);
+  TestBuddyAllocator(&buddy_allocator, 10 << 10);
+  TestBuddyAllocator(&buddy_allocator, 10 << 20);
+  TestBuddyAllocator(&buddy_allocator, 2 * static_cast<size_t>(1 << 30));
+}
+
+TEST(BuddyAllocator, InitRealloc) {
+  FLAGS_initial_gpu_memory_in_mb = 100;
+  FLAGS_reallocate_gpu_memory_in_mb = 50;
+
+  EXPECT_EQ(platform::GpuMaxChunkSize(), static_cast<size_t>(100 << 20));
+
+  BuddyAllocator buddy_allocator(
+      std::unique_ptr<SystemAllocator>(new GPUAllocator(test_gpu_id)),
+      platform::GpuMinChunkSize(), platform::GpuMaxChunkSize());
+
+  // Less then initial size and reallocate size
+  TestBuddyAllocator(&buddy_allocator, 10 << 20);
+  // Between initial size and reallocate size and not exceed pool
+  TestBuddyAllocator(&buddy_allocator, 80 << 20);
+  // Less then reallocate size and exceed pool
+  TestBuddyAllocator(&buddy_allocator, 40 << 20);
+  // Greater then reallocate size and exceed pool
+  TestBuddyAllocator(&buddy_allocator, 80 << 20);
+  // Greater then initial size and reallocate size
+  TestBuddyAllocator(&buddy_allocator, 2 * static_cast<size_t>(1 << 30));
+}
+
+TEST(BuddyAllocator, ReallocSizeGreaterThanInit) {
+  FLAGS_initial_gpu_memory_in_mb = 5;
+  FLAGS_reallocate_gpu_memory_in_mb = 10;
+
+  EXPECT_EQ(platform::GpuMaxChunkSize(), static_cast<size_t>(10 << 20));
+
+  BuddyAllocator buddy_allocator(
+      std::unique_ptr<SystemAllocator>(new GPUAllocator(test_gpu_id)),
+      platform::GpuMinChunkSize(), platform::GpuMaxChunkSize());
+
+  // Less then initial size and reallocate size
+  TestBuddyAllocator(&buddy_allocator, 1 << 20);
+  // Between initial size and reallocate size and not exceed pool
+  TestBuddyAllocator(&buddy_allocator, 3 << 20);
+  // Less then initial size and exceed pool
+  TestBuddyAllocator(&buddy_allocator, 3 << 20);
+  // Less then reallocate size and not exceed pool (now pool is 15 MB, used 7
+  // MB)
+  TestBuddyAllocator(&buddy_allocator, 7 << 20);
+  // Less then reallocate size and exceed pool
+  TestBuddyAllocator(&buddy_allocator, 8 << 20);
+  // Greater then initial size and reallocate size
+  TestBuddyAllocator(&buddy_allocator, 2 * static_cast<size_t>(1 << 30));
+}
+#endif
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/fluid/memory/detail/system_allocator.cc

@@ -32,6 +32,9 @@ limitations under the License. */
 
 DECLARE_bool(use_pinned_memory);
 DECLARE_double(fraction_of_gpu_memory_to_use);
+DECLARE_uint64(initial_gpu_memory_in_mb);
+DECLARE_uint64(reallocate_gpu_memory_in_mb);
+
 namespace paddle {
 namespace memory {
 namespace detail {
@@ -119,11 +122,18 @@ void* GPUAllocator::Alloc(size_t* index, size_t size) {
     gpu_alloc_size_ += size;
     return p;
   } else {
-    LOG(WARNING)
-        << "Cannot malloc " << size / 1024.0 / 1024.0
-        << " MB GPU memory. Please shrink FLAGS_fraction_of_gpu_memory_to_use "
-           "environment variable to a lower value. Current value is "
-        << FLAGS_fraction_of_gpu_memory_to_use;
+    LOG(WARNING) << "Cannot malloc " << size / 1024.0 / 1024.0
+                 << " MB GPU memory. Please shrink "
+                    "FLAGS_fraction_of_gpu_memory_to_use or "
+                    "FLAGS_initial_gpu_memory_in_mb or "
+                    "FLAGS_reallocate_gpu_memory_in_mb"
+                    "environment variable to a lower value. "
+                 << "Current FLAGS_fraction_of_gpu_memory_to_use value is "
+                 << FLAGS_fraction_of_gpu_memory_to_use
+                 << ". Current FLAGS_initial_gpu_memory_in_mb value is "
+                 << FLAGS_initial_gpu_memory_in_mb
+                 << ". Current FLAGS_reallocate_gpu_memory_in_mb value is "
+                 << FLAGS_reallocate_gpu_memory_in_mb;
     return nullptr;
   }
 }

paddle/fluid/operators/fsp_op.cc

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/fsp_op.h"
+
+namespace paddle {
+namespace operators {
+
+class FSPOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of FSPOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of FSPOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of FSPOp should not be null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+
+    PADDLE_ENFORCE(
+        x_dims.size() == 4,
+        "The Input(X) must have shape [batch_size, channel, height, width].");
+    PADDLE_ENFORCE(
+        y_dims.size() == 4,
+        "The Input(Y) must have shape [batch_size, channel, height, width].");
+    PADDLE_ENFORCE(
+        (x_dims[2] == y_dims[2]) && (x_dims[3] == y_dims[3]),
+        "The Input(X) and Input(Y) should have the same height and width.");
+
+    ctx->SetOutputDim("Out", {x_dims[0], x_dims[1], y_dims[1]});
+    ctx->ShareLoD("X", "Out");
+  }
+
+ protected:
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    framework::LibraryType library_{framework::LibraryType::kPlain};
+    framework::DataLayout layout_ = framework::DataLayout::kAnyLayout;
+    return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
+                                   ctx.device_context(), layout_, library_);
+  }
+};
+
+class FSPOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor) The input of FSP op with shape [batch_size, x_channel, "
+             "height, width]");
+    AddInput("Y",
+             "(Tensor) The input of FSP op with shape"
+             "[batch_size, y_channel, height, width]."
+             "The y_channel can be different with the x_channel of Input(X)"
+             " while the other dimensions must be the same with Input(X)'s.");
+    AddOutput(
+        "Out",
+        "(Tensor) The output of FSP op with shape "
+        "[batch_size, x_channel, y_channel]. The x_channel is the channel "
+        "of Input(X) and the y_channel is the channel of Input(Y).");
+    AddComment(R"DOC(
+    This op is used to calculate the flow of solution procedure (FSP) matrix of two feature maps.
+    Given feature map x with shape [x_channel, h, w] and feature map y with shape
+    [y_channel, h, w], we can get the fsp matrix of x and y in two steps:
+
+        step 1: reshape x into matrix with shape [x_channel, h * w] and reshape and
+                transpose y into matrix with shape [h * w, y_channel]
+        step 2: multiply x and y to get fsp matrix with shape [x_channel, y_channel]
+
+    The output is a batch of fsp matrices.
+    )DOC");
+  }
+};
+
+class FSPOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+    auto x_grad_name = framework::GradVarName("X");
+    auto y_grad_name = framework::GradVarName("Y");
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+    if (ctx->HasOutput(y_grad_name)) {
+      ctx->SetOutputDim(y_grad_name, y_dims);
+    }
+  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->type(),
+        ctx.device_context());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(fsp, ops::FSPOp, ops::FSPOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<true>);
+REGISTER_OPERATOR(fsp_grad, ops::FSPOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    fsp, ops::FSPOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::FSPOpKernel<paddle::platform::CPUDeviceContext, double>);
+REGISTER_OP_CPU_KERNEL(
+    fsp_grad, ops::FSPGradOpKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::FSPGradOpKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/fsp_op.cu

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/fsp_op.h"
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+REGISTER_OP_CUDA_KERNEL(fsp, ops::FSPOpKernel<plat::CUDADeviceContext, float>,
+                        ops::FSPOpKernel<plat::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(fsp_grad,
+                        ops::FSPGradOpKernel<plat::CUDADeviceContext, float>,
+                        ops::FSPGradOpKernel<plat::CUDADeviceContext, double>);

paddle/fluid/operators/fsp_op.h

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename DeviceContext, typename T>
+class FSPOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<Tensor>("X");
+    auto* y = context.Input<Tensor>("Y");
+    auto* output = context.Output<Tensor>("Out");
+    output->mutable_data<T>(context.GetPlace());
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+
+    auto batch_size = x_dims[0];
+    auto x_channel = x_dims[1];
+    auto y_channel = y_dims[1];
+    auto height = x_dims[2];
+    auto width = x_dims[3];
+
+    auto blas = math::GetBlas<DeviceContext, T>(context);
+
+    math::MatDescriptor x_mat_desc;
+    x_mat_desc.height_ = x_channel;
+    x_mat_desc.width_ = height * width;
+    x_mat_desc.batch_size_ = batch_size;
+    x_mat_desc.stride_ = x_channel * height * width;
+
+    math::MatDescriptor y_mat_desc;
+    y_mat_desc.height_ = height * width;
+    y_mat_desc.width_ = y_channel;
+    y_mat_desc.batch_size_ = batch_size;
+    y_mat_desc.stride_ = y_channel * height * width;
+    y_mat_desc.trans_ = true;
+
+    blas.MatMul(*x, x_mat_desc, *y, y_mat_desc,
+                static_cast<T>(1.0 / (height * width)), output,
+                static_cast<T>(0.0));
+  }
+};
+
+template <typename DeviceContext, typename T>
+class FSPGradOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* d_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* d_y = context.Output<Tensor>(framework::GradVarName("Y"));
+    if (d_x == nullptr && d_y == nullptr) {
+      return;
+    }
+    auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto d_out_dims = d_out->dims();
+    auto batch_size = d_out_dims[0];
+    auto x_channel = d_out_dims[1];
+    auto y_channel = d_out_dims[2];
+    int64_t h = 0;
+    int64_t w = 0;
+
+    auto blas = math::GetBlas<DeviceContext, T>(context);
+    math::SetConstant<DeviceContext, T> set_zero;
+    if (d_x != nullptr) {
+      d_x->mutable_data<T>(context.GetPlace());
+      set_zero(context.template device_context<DeviceContext>(), d_x,
+               static_cast<T>(0));
+      auto* y = context.Input<Tensor>("Y");
+      auto y_dims = y->dims();
+      h = y_dims[2];
+      w = y_dims[3];
+
+      math::MatDescriptor d_out_mat_desc;
+      d_out_mat_desc.height_ = x_channel;
+      d_out_mat_desc.width_ = y_channel;
+      d_out_mat_desc.batch_size_ = batch_size;
+      d_out_mat_desc.stride_ = x_channel * y_channel;
+
+      math::MatDescriptor y_mat_desc;
+      y_mat_desc.height_ = y_channel;
+      y_mat_desc.width_ = h * w;
+      y_mat_desc.batch_size_ = batch_size;
+      y_mat_desc.stride_ = y_channel * h * w;
+
+      blas.MatMul(*d_out, d_out_mat_desc, *y, y_mat_desc,
+                  static_cast<T>(1.0 / (h * w)), d_x, static_cast<T>(0.0));
+    }
+
+    if (d_y != nullptr) {
+      d_y->mutable_data<T>(context.GetPlace());
+      set_zero(context.template device_context<DeviceContext>(), d_y,
+               static_cast<T>(0));
+      auto* x = context.Input<Tensor>("X");
+      auto x_dims = x->dims();
+      h = x_dims[2];
+      w = x_dims[3];
+
+      math::MatDescriptor d_out_mat_desc;
+      d_out_mat_desc.height_ = y_channel;
+      d_out_mat_desc.width_ = x_channel;
+      d_out_mat_desc.batch_size_ = batch_size;
+      d_out_mat_desc.stride_ = x_channel * y_channel;
+      d_out_mat_desc.trans_ = true;
+
+      math::MatDescriptor x_mat_desc;
+      x_mat_desc.height_ = x_channel;
+      x_mat_desc.width_ = h * w;
+      x_mat_desc.batch_size_ = batch_size;
+      x_mat_desc.stride_ = x_channel * h * w;
+
+      blas.MatMul(*d_out, d_out_mat_desc, *x, x_mat_desc,
+                  static_cast<T>(1.0 / (h * w)), d_y, static_cast<T>(0.0));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/range_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/range_op.h"
+
+namespace paddle {
+namespace operators {
+
+class RangeOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    if (ctx->HasInput("Start")) {
+      auto s_dims = ctx->GetInputDim("Start");
+      PADDLE_ENFORCE((s_dims.size() == 1) && (s_dims[0] == 1),
+                     "The shape of Input(Start) should be [1].");
+    }
+    if (ctx->HasInput("End")) {
+      auto e_dims = ctx->GetInputDim("End");
+      PADDLE_ENFORCE((e_dims.size() == 1) && (e_dims[0] == 1),
+                     "The shape of Input(End) should be [1].");
+    }
+    if (ctx->HasInput("Step")) {
+      auto step_dims = ctx->GetInputDim("Step");
+      PADDLE_ENFORCE((step_dims.size() == 1) && (step_dims[0] == 1),
+                     "The shape of Input(Step) should be [1].");
+    }
+    ctx->SetOutputDim("Out", {-1});
+  }
+};
+
+class RangeOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("Start",
+             "Start of interval. The interval includes this value. It is a "
+             "tensor with shape=[1].");
+    AddInput("End",
+             "End of interval. The interval does not include this value, "
+             "except in some cases where step is not an integer and floating "
+             "point round-off affects the length of out. It is a tensor with "
+             "shape=[1].");
+    AddInput("Step", "Spacing between values. It is a tensor with shape=[1].");
+    AddOutput("Out", "A sequence of numbers.");
+    AddComment(R"DOC(
+    Return evenly spaced values within a given interval. Values are generated within the half-open interval [start, stop) (in other words, the interval including start but excluding stop). Like arange function of numpy.
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(range, ops::RangeOp, ops::RangeOpMaker);
+REGISTER_OP_CPU_KERNEL(range, ops::CPURangeKernel<int>,
+                       ops::CPURangeKernel<float>, ops::CPURangeKernel<double>,
+                       ops::CPURangeKernel<int64_t>);

paddle/fluid/operators/range_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/range_op.h"
+#include "paddle/fluid/platform/cuda_primitives.h"
+
+namespace paddle {
+namespace operators {
+
+#define CUDA_1D_KERNEL_LOOP(i, n)                              \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
+       i += blockDim.x * gridDim.x)
+
+template <typename T>
+__global__ void RangeKernel(T start, T step, int64_t size, T* out) {
+  CUDA_1D_KERNEL_LOOP(index, size) { out[index] = start + step * index; }
+}
+
+template <typename T>
+class CUDARangeKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* start_t = context.Input<framework::Tensor>("Start");
+    auto* end_t = context.Input<framework::Tensor>("End");
+    auto* step_t = context.Input<framework::Tensor>("Step");
+    auto* out = context.Output<framework::Tensor>("Out");
+
+    framework::Tensor n;
+    framework::TensorCopy(*start_t, platform::CPUPlace(), &n);
+    T start = n.data<T>()[0];
+    framework::TensorCopy(*end_t, platform::CPUPlace(), &n);
+    T end = n.data<T>()[0];
+    framework::TensorCopy(*step_t, platform::CPUPlace(), &n);
+    T step = n.data<T>()[0];
+
+    int64_t size = 0;
+    GetSize(start, end, step, &size);
+    out->Resize(framework::make_ddim({size}));
+    T* out_data = out->mutable_data<T>(context.GetPlace());
+
+    auto stream = context.cuda_device_context().stream();
+    int block = 512;
+    int grid = (size + block - 1) / block;
+    RangeKernel<T><<<grid, block, 0, stream>>>(start, step, size, out_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(range, ops::CUDARangeKernel<int>,
+                        ops::CUDARangeKernel<int64_t>,
+                        ops::CUDARangeKernel<float>,
+                        ops::CUDARangeKernel<double>);

paddle/fluid/operators/range_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <functional>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+void GetSize(T start, T end, T step, int64_t* size) {
+  PADDLE_ENFORCE(!std::equal_to<T>()(step, 0),
+                 "The step of range op should not be 0.");
+  PADDLE_ENFORCE(((start < end) && (step > 0)) || ((start > end) && (step < 0)),
+                 "The step should be greater than 0 while start < end. And the "
+                 "step should be less than 0 while start > end.");
+  *size = std::is_integral<T>::value
+              ? ((std::abs(end - start) + std::abs(step) - 1) / std::abs(step))
+              : std::ceil(std::abs((end - start) / step));
+}
+
+template <typename T>
+class CPURangeKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    T start = context.Input<framework::Tensor>("Start")->data<T>()[0];
+    T end = context.Input<framework::Tensor>("End")->data<T>()[0];
+    T step = context.Input<framework::Tensor>("Step")->data<T>()[0];
+    auto* out = context.Output<framework::Tensor>("Out");
+    int64_t size = 0;
+    GetSize(start, end, step, &size);
+    out->Resize(framework::make_ddim({size}));
+    T* out_data = out->mutable_data<T>(context.GetPlace());
+    T value = start;
+    for (int64_t i = 0; i < size; ++i) {
+      out_data[i] = value;
+      value += step;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/reader/CMakeLists.txt

@@ -17,7 +17,9 @@ function(reader_library TARGET_NAME)
         PARENT_SCOPE)
 endfunction()
 
+cc_library(py_reader SRCS py_reader.cc DEPS reader)
 cc_library(buffered_reader SRCS buffered_reader.cc DEPS reader simple_threadpool)
+
 reader_library(open_files_op SRCS open_files_op.cc DEPS buffered_reader)
 reader_library(create_random_data_generator_op SRCS create_random_data_generator_op.cc)
 reader_library(create_shuffle_reader_op SRCS create_shuffle_reader_op.cc)
@@ -26,7 +28,7 @@ reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_o
 reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc DEPS buffered_reader)
 reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc)
 reader_library(create_custom_reader_op SRCS create_custom_reader_op.cc)
-reader_library(create_py_reader_op SRCS create_py_reader_op.cc)
+reader_library(create_py_reader_op SRCS create_py_reader_op.cc DEPS py_reader)
 
 if (NOT WIN32 AND NOT ON_INFER)
     cc_library(ctr_reader SRCS ctr_reader.cc DEPS gzstream reader zlib)
@@ -38,7 +40,7 @@ cc_test(reader_blocking_queue_test SRCS reader_blocking_queue_test.cc)
 # Export local libraries to parent
 # set(READER_LIBRARY ${LOCAL_READER_LIBS} PARENT_SCOPE)
 
-op_library(read_op)
+op_library(read_op DEPS py_reader buffered_reader)
 
 foreach(src ${LOCAL_READER_LIBS})
     set(OP_LIBRARY ${src} ${OP_LIBRARY} CACHE INTERNAL "op libs")

paddle/fluid/operators/reader/blocking_queue.h

@@ -16,6 +16,7 @@
 
 #include <condition_variable>  // NOLINT
 #include <deque>
+#include <utility>
 
 #include "paddle/fluid/platform/enforce.h"
 
@@ -34,7 +35,7 @@ class BlockingQueue {
   explicit BlockingQueue(size_t capacity, bool speed_test_mode = false)
       : capacity_(capacity), speed_test_mode_(speed_test_mode), closed_(false) {
     PADDLE_ENFORCE_GT(
-        capacity_, 0,
+        capacity_, static_cast<size_t>(0),
         "The capacity of a reader::BlockingQueue must be greater than 0.");
   }
 

paddle/fluid/operators/reader/buffered_reader.cc

@@ -30,8 +30,10 @@ BufferedReader::~BufferedReader() {
 #ifdef PADDLE_WITH_CUDA
   if (platform::is_gpu_place(place_)) {
     platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
-    PADDLE_ENFORCE(cudaStreamDestroy(stream));
-    for (auto &event : events) PADDLE_ENFORCE(cudaEventDestroy(event));
+    PADDLE_ENFORCE(cudaStreamDestroy(stream_));
+    for (auto &event : events_) {
+      PADDLE_ENFORCE(cudaEventDestroy(event));
+    }
   }
 #endif
 }
@@ -46,15 +48,15 @@ BufferedReader::BufferedReader(
 #ifdef PADDLE_WITH_CUDA
   if (platform::is_gpu_place(place_)) {
     platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
-    compute_stream =
+    compute_stream_ =
         ((platform::CUDADeviceContext *)(platform::DeviceContextPool::Instance()
                                              .Get(place_)))
             ->stream();
-    events.resize(buffer_size);
-    for (auto &event : events) {
+    events_.resize(buffer_size);
+    for (auto &event : events_) {
       PADDLE_ENFORCE(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
     }
-    PADDLE_ENFORCE(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
+    PADDLE_ENFORCE(cudaStreamCreateWithFlags(&stream_, cudaStreamNonBlocking));
   }
 #endif
   cpu_buffer_.resize(buffer_size);
@@ -73,7 +75,7 @@ void BufferedReader::ReadAsync(size_t i) {
 #ifdef PADDLE_WITH_CUDA
   if (platform::is_gpu_place(place_)) {
     platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
-    PADDLE_ENFORCE(cudaEventRecord(events[i], compute_stream));
+    PADDLE_ENFORCE(cudaEventRecord(events_[i], compute_stream_));
   }
 #endif
   position_.emplace(thread_pool_.enqueue([this, i]() -> size_t {
@@ -91,7 +93,7 @@ void BufferedReader::ReadAsync(size_t i) {
     // commands from different streams cannot run concurrently.
     if (platform::is_gpu_place(place_)) {
       platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
-      PADDLE_ENFORCE(cudaStreamWaitEvent(stream, events[i], 0));
+      PADDLE_ENFORCE(cudaStreamWaitEvent(stream_, events_[i], 0));
       TensorVec &gpu = gpu_buffer_[i];
       gpu.resize(cpu.size());
       platform::RecordEvent record_event("BufferedReader:MemoryCopy");
@@ -106,12 +108,14 @@ void BufferedReader::ReadAsync(size_t i) {
         if (platform::is_cuda_pinned_place(cpu_place)) {
           memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
                        boost::get<platform::CUDAPinnedPlace>(cpu_place),
-                       cpu_ptr, size, stream);
+                       cpu_ptr, size, stream_);
         } else if ((platform::is_gpu_place(cpu_place))) {
           memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
                        boost::get<platform::CUDAPlace>(cpu_place), cpu_ptr,
-                       size, stream);
+                       size, stream_);
         } else {
+          // if cpu place is not pinned, async copy is slower than sync copy,
+          // so we use sync copy instead.
           // TODO(zcd): The default stream should not be used here.
           memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
                        boost::get<platform::CPUPlace>(cpu_place), cpu_ptr, size,
@@ -119,7 +123,7 @@ void BufferedReader::ReadAsync(size_t i) {
         }
         gpu[i].set_lod(cpu[i].lod());
       }
-      PADDLE_ENFORCE(cudaStreamSynchronize(stream));
+      PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
     }
 #endif
     return i;

paddle/fluid/operators/reader/buffered_reader.h

@@ -15,6 +15,7 @@
 #pragma once
 
 #include <list>
+#include <memory>
 #include <queue>
 #include <vector>
 #include "ThreadPool.h"
@@ -63,9 +64,9 @@ class BufferedReader : public framework::DecoratedReader {
   std::vector<TensorVec> gpu_buffer_;
   size_t prev_pos_{-1UL};
 #ifdef PADDLE_WITH_CUDA
-  cudaStream_t stream;
-  cudaStream_t compute_stream;
-  std::vector<cudaEvent_t> events;
+  cudaStream_t stream_;
+  cudaStream_t compute_stream_;
+  std::vector<cudaEvent_t> events_;
 #endif
 };
 

paddle/fluid/operators/reader/create_py_reader_op.cc

@@ -12,37 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
+#include "paddle/fluid/operators/reader/py_reader.h"
 #include "paddle/fluid/operators/reader/reader_op_registry.h"
 
 namespace paddle {
 namespace operators {
 namespace reader {
 
-class PyReader : public framework::FileReader {
- public:
-  explicit PyReader(const std::shared_ptr<LoDTensorBlockingQueue>& queue)
-      : framework::FileReader() {
-    PADDLE_ENFORCE(queue != nullptr, "LoDTensorBlockingQueue must not be null");
-    queue_ = queue;
-  }
-
-  void ReadNext(std::vector<framework::LoDTensor>* out) override {
-    bool success;
-    *out = queue_->Pop(&success);
-    if (!success) out->clear();
-  }
-
-  ~PyReader() { queue_->Close(); }
-
-  void Shutdown() override { queue_->Close(); }
-
-  void Start() override { queue_->ReOpen(); }
-
- private:
-  std::shared_ptr<LoDTensorBlockingQueue> queue_;
-};
-
 class CreatePyReaderOp : public framework::OperatorBase {
  public:
   using framework::OperatorBase::OperatorBase;

paddle/fluid/operators/reader/py_reader.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/operators/reader/py_reader.h"
+#include <memory>
+
+namespace paddle {
+namespace operators {
+namespace reader {
+
+PyReader::PyReader(const std::shared_ptr<LoDTensorBlockingQueue>& queue)
+    : framework::FileReader() {
+  PADDLE_ENFORCE(queue != nullptr, "LoDTensorBlockingQueue must not be null");
+  queue_ = queue;
+}
+
+void PyReader::ReadNext(std::vector<framework::LoDTensor>* out) {
+  bool success;
+  *out = queue_->Pop(&success);
+  if (!success) out->clear();
+}
+
+PyReader::~PyReader() { queue_->Close(); }
+
+void PyReader::Shutdown() { queue_->Close(); }
+
+void PyReader::Start() { queue_->ReOpen(); }
+
+}  // namespace reader
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/reader/py_reader.h

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <atomic>
+#include <memory>
+#include <vector>
+#include "paddle/fluid/framework/reader.h"
+#include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
+
+namespace paddle {
+namespace operators {
+namespace reader {
+
+class PyReader : public framework::FileReader {
+ public:
+  explicit PyReader(const std::shared_ptr<LoDTensorBlockingQueue>& queue);
+
+  void ReadNext(std::vector<framework::LoDTensor>* out) override;
+
+  ~PyReader();
+
+  void Shutdown() override;
+
+  void Start() override;
+
+ private:
+  std::shared_ptr<LoDTensorBlockingQueue> queue_;
+};
+
+}  // namespace reader
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/platform/gpu_info.cc

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/platform/gpu_info.h"
-
 #include <algorithm>
 #include <cstdlib>
 #include <string>
@@ -31,6 +30,8 @@ constexpr static float fraction_of_gpu_memory_to_use = 0.92f;
 constexpr static float fraction_of_gpu_memory_to_use = 0.5f;
 #endif
 
+constexpr static float fraction_reserve_gpu_memory = 0.05f;
+
 DEFINE_double(fraction_of_gpu_memory_to_use, fraction_of_gpu_memory_to_use,
               "Allocate a trunk of gpu memory that is this fraction of the "
               "total gpu memory size. Future memory usage will be allocated "
@@ -38,6 +39,24 @@ DEFINE_double(fraction_of_gpu_memory_to_use, fraction_of_gpu_memory_to_use,
               "additional trunks of the same size will be requested from gpu "
               "until the gpu has no memory left for another trunk.");
 
+DEFINE_uint64(
+    initial_gpu_memory_in_mb, 0ul,
+    "Allocate a trunk of gpu memory whose byte size is specified by "
+    "the flag. Future memory usage will be allocated from the "
+    "truck. If the trunk doesn't have enough gpu memory, additional "
+    "trunks of the gpu memory will be requested from gpu with size "
+    "specified by FLAGS_reallocate_gpu_memory_in_mb until the gpu has "
+    "no memory left for the additional trunk. Note: if you set this "
+    "flag, the memory size set by "
+    "FLAGS_fraction_of_gpu_memory_to_use will be overrided by this "
+    "flag. If you don't set this flag, PaddlePaddle will use "
+    "FLAGS_fraction_of_gpu_memory_to_use to allocate gpu memory");
+
+DEFINE_uint64(reallocate_gpu_memory_in_mb, 0ul,
+              "If this flag is set, Paddle will reallocate the gpu memory with "
+              "size specified by this flag. Else Paddle will reallocate by "
+              "FLAGS_fraction_of_gpu_memory_to_use");
+
 DEFINE_bool(
     enable_cublas_tensor_op_math, false,
     "The enable_cublas_tensor_op_math indicate whether to use Tensor Core, "
@@ -180,13 +199,43 @@ void GpuMemoryUsage(size_t *available, size_t *total) {
 }
 
 size_t GpuMaxAllocSize() {
+  return std::max(GpuInitAllocSize(), GpuReallocSize());
+}
+
+size_t GpuInitAllocSize() {
+  if (FLAGS_initial_gpu_memory_in_mb > 0ul) {
+    // Initial memory will be allocated by FLAGS_initial_gpu_memory_in_mb
+    return static_cast<size_t>(FLAGS_initial_gpu_memory_in_mb << 20);
+  }
+
+  // FLAGS_initial_gpu_memory_in_mb is 0, initial memory will be allocated by
+  // fraction
   size_t total = 0;
   size_t available = 0;
 
   GpuMemoryUsage(&available, &total);
+  size_t reserving = static_cast<size_t>(fraction_reserve_gpu_memory * total);
 
-  // Reserve the rest for page tables, etc.
-  return static_cast<size_t>(total * FLAGS_fraction_of_gpu_memory_to_use);
+  return static_cast<size_t>((total - reserving) *
+                             FLAGS_fraction_of_gpu_memory_to_use);
+}
+
+size_t GpuReallocSize() {
+  if (FLAGS_reallocate_gpu_memory_in_mb > 0ul) {
+    // Additional memory will be allocated by FLAGS_reallocate_gpu_memory_in_mb
+    return static_cast<size_t>(FLAGS_reallocate_gpu_memory_in_mb << 20);
+  }
+
+  // FLAGS_reallocate_gpu_memory_in_mb is 0, additional memory will be allocated
+  // by fraction
+  size_t total = 0;
+  size_t available = 0;
+
+  GpuMemoryUsage(&available, &total);
+  size_t reserving = static_cast<size_t>(fraction_reserve_gpu_memory * total);
+
+  return static_cast<size_t>((total - reserving) *
+                             FLAGS_fraction_of_gpu_memory_to_use);
 }
 
 size_t GpuMinChunkSize() {
@@ -201,16 +250,13 @@ size_t GpuMaxChunkSize() {
   GpuMemoryUsage(&available, &total);
   VLOG(10) << "GPU Usage " << available / 1024 / 1024 << "M/"
            << total / 1024 / 1024 << "M";
-  size_t reserving = static_cast<size_t>(0.05 * total);
+  size_t reserving = static_cast<size_t>(fraction_reserve_gpu_memory * total);
   // If available less than minimum chunk size, no usable memory exists.
   available =
       std::min(std::max(available, GpuMinChunkSize()) - GpuMinChunkSize(),
                total - reserving);
 
-  // Reserving the rest memory for page tables, etc.
-
-  size_t allocating = static_cast<size_t>(FLAGS_fraction_of_gpu_memory_to_use *
-                                          (total - reserving));
+  size_t allocating = GpuMaxAllocSize();
 
   PADDLE_ENFORCE_LE(allocating, available,
                     "Insufficient GPU memory to allocation.");

paddle/fluid/platform/gpu_info.h

@@ -60,6 +60,12 @@ void GpuMemoryUsage(size_t *available, size_t *total);
 //! Get the maximum allocation size of current GPU device.
 size_t GpuMaxAllocSize();
 
+//! Get the initial allocation size of current GPU device.
+size_t GpuInitAllocSize();
+
+//! Get the re-allocation size of current GPU device.
+size_t GpuReallocSize();
+
 //! Get the minimum chunk size for GPU buddy allocator.
 size_t GpuMinChunkSize();
 

paddle/fluid/pybind/CMakeLists.txt

@@ -5,7 +5,7 @@ set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune
 if(WITH_PYTHON)
   list(APPEND PYBIND_DEPS py_func_op)
 endif()
-set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc imperative.cc ir.cc inference_api.cc)
+set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc reader_py.cc async_executor_py.cc imperative.cc ir.cc inference_api.cc)
 
 if(WITH_PYTHON)
   if(WITH_AMD_GPU)

paddle/fluid/pybind/pybind.cc

@@ -55,6 +55,7 @@ limitations under the License. */
 #include "paddle/fluid/pybind/ir.h"
 #include "paddle/fluid/pybind/protobuf.h"
 #include "paddle/fluid/pybind/pybind.h"  // NOLINT
+#include "paddle/fluid/pybind/reader_py.h"
 #include "paddle/fluid/pybind/recordio.h"
 #include "paddle/fluid/pybind/tensor_py.h"
 
@@ -128,6 +129,11 @@ static inline bool IsSamePlace(const PlaceType1 &p1, const PlaceType2 &p2) {
   return paddle::platform::Place(p1) == paddle::platform::Place(p2);
 }
 
+template <typename PlaceType>
+static inline int PlaceIndex(const PlaceType &p) {
+  return static_cast<int>(paddle::platform::Place(p).which());
+}
+
 PYBIND11_MODULE(core, m) {
   // Not used, just make sure cpu_info.cc is linked.
   paddle::platform::CpuTotalPhysicalMemory();
@@ -531,6 +537,7 @@ PYBIND11_MODULE(core, m) {
 
 All parameter, weight, gradient are variables in Paddle.
 )DOC")
+      .def(py::init<>())
       .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
       .def("set_int",
            [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
@@ -572,14 +579,13 @@ All parameter, weight, gradient are variables in Paddle.
            },
            py::return_value_policy::reference);
 
-  py::class_<framework::ReaderHolder>(m, "Reader", "")
-      .def("start", &framework::ReaderHolder::Start)
-      .def("reset", &framework::ReaderHolder::ResetAll);
+  BindReader(&m);
 
   using LoDTensorBlockingQueue =
       ::paddle::operators::reader::LoDTensorBlockingQueue;
   using LoDTensorBlockingQueueHolder =
       ::paddle::operators::reader::LoDTensorBlockingQueueHolder;
+
   py::class_<LoDTensorBlockingQueue, std::shared_ptr<LoDTensorBlockingQueue>>(
       m, "LoDTensorBlockingQueue", "")
       .def("push",
@@ -776,6 +782,7 @@ All parameter, weight, gradient are variables in Paddle.
              PADDLE_THROW("Cannot use CUDAPlace in CPU only version");
 #endif
            })
+      .def("_type", &PlaceIndex<platform::CUDAPlace>)
       .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::Place>)
       .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::CUDAPlace>)
       .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::CPUPlace>)
@@ -785,6 +792,7 @@ All parameter, weight, gradient are variables in Paddle.
 
   py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
       .def(py::init<>())
+      .def("_type", &PlaceIndex<platform::CPUPlace>)
       .def("_equals", &IsSamePlace<platform::CPUPlace, platform::Place>)
       .def("_equals", &IsSamePlace<platform::CPUPlace, platform::CUDAPlace>)
       .def("_equals", &IsSamePlace<platform::CPUPlace, platform::CPUPlace>)
@@ -800,6 +808,7 @@ All parameter, weight, gradient are variables in Paddle.
 #endif
              new (&self) platform::CUDAPinnedPlace();
            })
+      .def("_type", &PlaceIndex<platform::CUDAPinnedPlace>)
       .def("_equals", &IsSamePlace<platform::CUDAPinnedPlace, platform::Place>)
       .def("_equals",
            &IsSamePlace<platform::CUDAPinnedPlace, platform::CUDAPlace>)
@@ -811,16 +820,25 @@ All parameter, weight, gradient are variables in Paddle.
 
   py::class_<platform::Place>(m, "Place")
       .def(py::init<>())
+      .def("_type", &PlaceIndex<platform::Place>)
       .def("_equals", &IsSamePlace<platform::Place, platform::Place>)
       .def("_equals", &IsSamePlace<platform::Place, platform::CUDAPlace>)
       .def("_equals", &IsSamePlace<platform::Place, platform::CPUPlace>)
       .def("_equals", &IsSamePlace<platform::Place, platform::CUDAPinnedPlace>)
       .def("is_gpu_place",
            [](platform::Place &self) { return platform::is_gpu_place(self); })
+      .def("is_cpu_place",
+           [](platform::Place &self) { return platform::is_cpu_place(self); })
+      .def("is_cuda_pinned_place",
+           [](platform::Place &self) {
+             return platform::is_cuda_pinned_place(self);
+           })
       .def("gpu_device_id",
            [](platform::Place &self) {
              return boost::get<platform::CUDAPlace>(self).device;
            })
+      .def("set_place", [](platform::Place &self,
+                           const platform::Place &other) { self = other; })
       .def("set_place",
            [](platform::Place &self, const platform::CPUPlace &cpu_place) {
              self = cpu_place;

paddle/fluid/pybind/reader_py.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/pybind/reader_py.h"
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include "paddle/fluid/framework/reader.h"
+#include "paddle/fluid/operators/reader/buffered_reader.h"
+#include "paddle/fluid/operators/reader/py_reader.h"
+#include "paddle/fluid/platform/place.h"
+#include "pybind11/stl.h"
+
+namespace paddle {
+namespace pybind {
+
+class MultiDeviceFeedReader {
+ public:
+  using ResultDictList =
+      std::vector<std::unordered_map<std::string, framework::LoDTensor>>;
+
+  MultiDeviceFeedReader(
+      const std::shared_ptr<operators::reader::LoDTensorBlockingQueue> &queue,
+      const std::vector<std::string> &names,
+      const std::vector<platform::Place> &dst_places, bool use_double_buffer)
+      : queue_(queue),
+        names_(names),
+        pool_(new ::ThreadPool(dst_places.size())) {
+    std::shared_ptr<framework::ReaderBase> reader(
+        new operators::reader::PyReader(queue));
+
+    readers_.reserve(dst_places.size());
+    for (auto &p : dst_places) {
+      auto *holder = new framework::ReaderHolder();
+      if (use_double_buffer) {
+        holder->Reset(
+            framework::MakeDecoratedReader<operators::reader::BufferedReader>(
+                reader, p, 2));
+      } else {
+        if (platform::is_gpu_place(p)) {
+          PADDLE_THROW(
+              "Place cannot be CUDAPlace when use_double_buffer is False");
+        }
+        holder->Reset(reader);
+      }
+      readers_.emplace_back(holder);
+    }
+
+    futures_.resize(dst_places.size());
+    ret_.resize(dst_places.size());
+    ReadAsync();
+  }
+
+  ResultDictList ReadNext() {
+    bool success = WaitFutures();
+
+    if (!success) {
+      return {};
+    }
+
+    ResultDictList result(ret_.size());
+    for (size_t i = 0; i < ret_.size(); ++i) {
+      for (size_t j = 0; j < names_.size(); ++j) {
+        result[i].emplace(names_[j], std::move(ret_[i][j]));
+      }
+    }
+    ReadAsync();
+    return result;
+  }
+
+  void Reset() {
+    Shutdown();
+    Start();
+    ReadAsync();
+  }
+
+  ~MultiDeviceFeedReader() {
+    queue_->Close();
+    pool_.reset();
+  }
+
+ private:
+  bool WaitFutures() {
+    bool success = true;
+    for (auto &f : futures_) {
+      success &= f.get();
+    }
+    return success;
+  }
+
+  void Shutdown() {
+    for (auto &r : readers_) r->Shutdown();
+  }
+
+  void Start() {
+    for (auto &r : readers_) r->Start();
+  }
+
+  void ReadAsync() {
+    for (size_t i = 0; i < readers_.size(); ++i) {
+      futures_[i] = pool_->enqueue([this, i] {
+        readers_[i]->ReadNext(&ret_[i]);
+        return !ret_[i].empty();
+      });
+    }
+  }
+
+  std::shared_ptr<operators::reader::LoDTensorBlockingQueue> queue_;
+  std::vector<std::string> names_;
+  std::unique_ptr<::ThreadPool> pool_;
+
+  std::vector<std::unique_ptr<framework::ReaderHolder>> readers_;
+
+  std::vector<std::future<bool>> futures_;
+  std::vector<std::vector<framework::LoDTensor>> ret_;
+};
+
+namespace py = pybind11;
+
+void BindReader(py::module *module) {
+  auto &m = *module;
+
+  namespace reader = ::paddle::operators::reader;
+
+  py::class_<framework::ReaderHolder>(m, "Reader", "")
+      .def("start", &framework::ReaderHolder::Start)
+      .def("reset", &framework::ReaderHolder::ResetAll);
+
+  py::class_<MultiDeviceFeedReader>(m, "MultiDeviceFeedReader", "")
+      .def("read_next", &MultiDeviceFeedReader::ReadNext,
+           py::call_guard<py::gil_scoped_release>())
+      .def("reset", &MultiDeviceFeedReader::Reset,
+           py::call_guard<py::gil_scoped_release>());
+
+  m.def("create_py_reader",
+        [](const std::shared_ptr<operators::reader::LoDTensorBlockingQueue>
+               &queue,
+           const std::vector<std::string> &names,
+           const std::vector<platform::Place> &dst_places,
+           bool use_double_buffer) {
+          return new MultiDeviceFeedReader(queue, names, dst_places,
+                                           use_double_buffer);
+        },
+        py::return_value_policy::take_ownership);
+}
+
+}  // namespace pybind
+}  // namespace paddle

paddle/fluid/pybind/reader_py.h

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "pybind11/pybind11.h"
+
+namespace paddle {
+namespace pybind {
+
+void BindReader(pybind11::module *module);
+
+}  // namespace pybind
+}  // namespace paddle

paddle/testing/paddle_gtest_main.cc

@@ -41,6 +41,8 @@ int main(int argc, char** argv) {
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   envs.push_back("fraction_of_gpu_memory_to_use");
+  envs.push_back("initial_gpu_memory_in_mb");
+  envs.push_back("reallocate_gpu_memory_in_mb");
   envs.push_back("allocator_strategy");
 #elif __clang__
   envs.push_back("use_mkldnn");

python/paddle/dataset/flowers.py

@@ -46,9 +46,9 @@ import six
 from six.moves import cPickle as pickle
 __all__ = ['train', 'test', 'valid']
 
-DATA_URL = 'http://paddlemodels.cdn.bcebos.com/flowers/102flowers.tgz'
-LABEL_URL = 'http://paddlemodels.cdn.bcebos.com/flowers/imagelabels.mat'
-SETID_URL = 'http://paddlemodels.cdn.bcebos.com/flowers/setid.mat'
+DATA_URL = 'http://paddlemodels.bj.bcebos.com/flowers/102flowers.tgz'
+LABEL_URL = 'http://paddlemodels.bj.bcebos.com/flowers/imagelabels.mat'
+SETID_URL = 'http://paddlemodels.bj.bcebos.com/flowers/setid.mat'
 DATA_MD5 = '52808999861908f626f3c1f4e79d11fa'
 LABEL_MD5 = 'e0620be6f572b9609742df49c70aed4d'
 SETID_MD5 = 'a5357ecc9cb78c4bef273ce3793fc85c'

python/paddle/dataset/wmt14.py

@@ -15,7 +15,7 @@
 WMT14 dataset.
 The original WMT14 dataset is too large and a small set of data for set is
 provided. This module will download dataset from
-http://paddlepaddle.cdn.bcebos.com/demo/wmt_shrinked_data/wmt14.tgz and
+http://paddlepaddle.bj.bcebos.com/demo/wmt_shrinked_data/wmt14.tgz and
 parse training set and test set into paddle reader creators.
 
 """

python/paddle/fluid/__init__.py

@@ -163,7 +163,8 @@ def __bootstrap__():
 
     if core.is_compiled_with_cuda():
         read_env_flags += [
-            'fraction_of_gpu_memory_to_use', 'cudnn_deterministic',
+            'fraction_of_gpu_memory_to_use', 'initial_gpu_memory_in_mb',
+            'reallocate_gpu_memory_in_mb', 'cudnn_deterministic',
             'enable_cublas_tensor_op_math', 'conv_workspace_size_limit',
             'cudnn_exhaustive_search', 'memory_optimize_debug', 'selected_gpus',
             'sync_nccl_allreduce', 'limit_of_tmp_allocation',

python/paddle/fluid/compiler.py

@@ -17,9 +17,10 @@ import os
 import six
 import sys
 from .. import compat as cpt
+from . import framework
+from .framework import cuda_places, cpu_places
 
 from . import core
-from . import framework
 
 __all__ = ['CompiledProgram', 'ExecutionStrategy', 'BuildStrategy']
 
@@ -44,21 +45,6 @@ def _is_pserver_mode(main_program):
     return False
 
 
-def get_available_places(use_cuda):
-    if use_cuda:
-        gpus_env = os.getenv("FLAGS_selected_gpus")
-        if gpus_env:
-            gpus = [int(s) for s in gpus_env.split(",")]
-        else:
-            gpus = [i for i in six.moves.range(core.get_cuda_device_count())]
-        places = [core.CUDAPlace(i) for i in gpus]
-    else:
-        cpu_num = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
-        places = [core.CPUPlace() for _ in six.moves.range(cpu_num)]
-    assert places, "no place for execution"
-    return places
-
-
 class CompiledProgram(object):
     """
     Compiles to Graph for execution.
@@ -117,7 +103,8 @@ class CompiledProgram(object):
                            loss_name=None,
                            build_strategy=None,
                            exec_strategy=None,
-                           share_vars_from=None):
+                           share_vars_from=None,
+                           places=None):
         """Configs the program to run in data parallel way.
 
         Args:
@@ -132,10 +119,18 @@ class CompiledProgram(object):
                 threads are used, how many iterations to clean up the temp
                 variables. For more information, please refer
                 to fluid.ExecutionStrategy. Default None.
-            share_vars_from(CompiledProgram): If provide, this CompiledProgram
+            share_vars_from(CompiledProgram): If provided, this CompiledProgram
                 will share variables from `share_vars_from`. `share_vars_from`
                 must be run by the executor before this CompiledProgram so that
                 vars are ready.
+            places(list(CUDAPlace)|list(CPUPlace)|None): If provided, only compile
+                program in the given places. Otherwise, the places used when compiled 
+                is determined by the Executor, and the places used are controlled 
+                by environment variables: FLAGS_selected_gpus or CUDA_VISIBLE_DEVICES
+                if using GPU; or CPU_NUM if using CPU. For example, if you want to 
+                run on GPU 0 and 1, set places=[fluid.CUDAPlace(0), fluid.CUDAPlace(1)].
+                If you want to run on 2 CPU cores, set places=[fluid.CPUPlace()]*2.  
+
         Returns:
             self
         """
@@ -150,6 +145,12 @@ class CompiledProgram(object):
             self._exec_strategy = ExecutionStrategy()
         if self._build_strategy is None:
             self._build_strategy = BuildStrategy()
+        if places is not None:
+            if not isinstance(places, (list, tuple)):
+                places = [places]
+            self._places = places
+        else:
+            self._places = None
         self._build_strategy.is_distribution = _is_pserver_mode(self._program)
         return self
 
@@ -192,7 +193,15 @@ class CompiledProgram(object):
             self._local_scopes = []
 
         self._exec_strategy.use_cuda = use_cuda
-        self._places = get_available_places(self._exec_strategy.use_cuda)
+        has_set_place = (self._places is not None)
+        if has_set_place:
+            for p in self._places:
+                assert p._type() == self._place._type(), \
+                    "Place type not match. You may set the wrong type of places"
+        else:
+            self._places = cuda_places(
+            ) if self._exec_strategy.use_cuda else cpu_places()
+        assert self._places, "no place for execution"
 
         if self._exec_strategy.num_threads == 0:
             if self._exec_strategy.use_cuda:
@@ -200,9 +209,7 @@ class CompiledProgram(object):
                 # performance. Worth tunning for other models in the future.
                 self._exec_strategy.num_threads = len(self._places) * 4
             else:
-                cpu_num = int(
-                    os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
-                self._exec_strategy.num_threads = cpu_num * 2
+                self._exec_strategy.num_threads = len(self._places) * 2
 
         # FIXME(dzhwinter): enable_inplace should be after memory_optimize
         # if turn on python memory optimize, turn off the inplace_pass.

python/paddle/fluid/contrib/int8_inference/README.md

@@ -45,28 +45,41 @@ You can load INT8 model by load_inference_model [API](https://github.com/PaddleP
 ```
 
 ## 3. Result
-We provide the results of accuracy measurd on [Intel® Xeon® Platinum Gold Processor](https://ark.intel.com/products/120489/Intel-Xeon-Gold-6148-Processor-27-5M-Cache-2-40-GHz- "Intel® Xeon® Gold 6148 Processor") (also known as Intel® Xeon® Skylake6148).
+We provide the results of accuracy and performance measured on Intel(R) Xeon(R) Gold 6271 (single core).
+
+**I. Top-1 Accuracy on Intel(R) Xeon(R) Gold 6271**
 
 | Model  | Dataset  | FP32 Accuracy  | INT8 Accuracy  | Accuracy Diff  |
-| ------------ | ------------ | ------------ | ------------ | ------------ |
-| ResNet-50  | Small  | 72.00%  | 72.00%  |  0.00% |
-| MobileNet-V1  | Small  | 62.00%  | 62.00%  | 0.00%  |
-| ResNet-50  | Full ImageNet Val  |  76.63%  | 76.17%  | 0.46% |
-| MobileNet-V1 | Full ImageNet Val  | 70.78%  | 70.49%  | 0.29%  |
+| :------------: | :------------: | :------------: | :------------: | :------------: |
+| ResNet-50  | Full ImageNet Val  |  76.63%  | 76.23%  | 0.40% |
+| MobileNet-V1 | Full ImageNet Val  | 70.78%  | 70.47%  | 0.31%  |
+
+**II. Throughput on Intel(R) Xeon(R) Gold 6271 (batch size 1 on single core)**
+
+| Model  | Dataset  | FP32 Throughput  | INT8 Throughput  |  Ratio(INT8/FP32)  |
+| :------------: | :------------: | :------------: | :------------: | :------------: |
+| ResNet-50  | Full ImageNet Val  |  11.54 images/s | 32.2 images/s | 2.79 |
+| MobileNet-V1 | Full ImageNet Val  | 49.21 images/s | 108.37 images/s | 2.2  |
 
-Please note that [Small](http://paddle-inference-dist.cdn.bcebos.com/int8/calibration_test_data.tar.gz "Small") is a subset of [full ImageNet validation dataset](http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_img_val.tar "full ImageNet validation dataset"). 
+Please note that [full ImageNet validation dataset](http://www.image-net.org/challenges/LSVRC/2012/nnoupb/ILSVRC2012_img_val.tar "full ImageNet validation dataset") can be downloaded by script `test_calibration.py` with `DATASET=full`. 
 
 Notes:
 * The accuracy measurement requires the model with `label`.
-* The INT8 theoretical speedup is ~1.33X on Intel® Xeon® Skylake Server (please refer to `This allows for 4x more input at the cost of 3x more instructions or 33.33% more compute` in  [Reference](https://software.intel.com/en-us/articles/lower-numerical-precision-deep-learning-inference-and-training "Reference")).
+* The INT8 theoretical speedup is 4X on Intel® Xeon® Cascadelake Server (please refer to `providing a theoretical peak compute gain of 4x int8 OPS over fp32 OPS` in  [Reference](https://software.intel.com/en-us/articles/lower-numerical-precision-deep-learning-inference-and-training "Reference")). However, the actual test results at the model level will be less than 4X, and in general the average is about 2X. In addition, the calculation library optimization of batch size 1 is not as good as the large batch size.
 
 ## 4. How to reproduce the results
-* Small dataset
+* Small dataset (Single core)
 ```bash
 FLAGS_use_mkldnn=true python python/paddle/fluid/contrib/tests/test_calibration.py
 ```
 
-* Full dataset
+* Full dataset (Single core)
 ```bash
 FLAGS_use_mkldnn=true DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py
 ```
+
+* Full dataset (Multi-core)
+```bash
+FLAGS_use_mkldnn=true OMP_NUM_THREADS=20 DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py
+```
+> Notes: This is an example command with 20 cores by using set `OMP_NUM_THREADS` value.

python/paddle/fluid/contrib/slim/__init__.py

@@ -13,13 +13,4 @@
 # limitations under the License.
 
 from .core import *
-from .graph import *
-from .prune import *
-__all__ = [
-    'build_compressor',
-    'CompressPass',
-    'ImitationGraph',
-    'SensitivePruneStrategy',
-    'MagnitudePruner',
-    'RatioPruner',
-]
+__all__ = ['Compressor', ]

python/paddle/fluid/contrib/slim/core/__init__.py

@@ -14,11 +14,9 @@
 
 from . import config
 from .config import *
-from . import compress_pass
-from .compress_pass import *
+from . import compressor
+from .compressor import *
 from . import strategy
 from .strategy import *
-from . import pass_builder
-from .pass_builder import *
 
-__all__ = config.__all__ + compress_pass.__all__ + strategy.__all__ + pass_builder.__all__
+__all__ = config.__all__ + compressor.__all__ + strategy.__all__

python/paddle/fluid/contrib/slim/core/compress_pass.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from ....core import CPUPlace
-from ..graph import get_executor
-
-__all__ = ['Context', 'CompressPass']
-
-
-class Context(object):
-    """
-    The context in the process of compression.
-    Args:
-        exe: The executor used to execute graph.
-        graph: The graph to be compressed.
-        scope: The scope used to execute graph.
-        program_exe: The program_exe is used to execute the program
-                     created for modifying the variables in scope.
-    """
-
-    def __init__(self, exe, graph, scope, program_exe=None):
-        # The total number of epoches to be trained.
-        self.epoch = 0
-        # Current epoch
-        self.epoch_id = 0
-        # Current batch
-        self.batch_id = 0
-        self.exe = exe
-        self.graph = graph
-        self.scope = scope
-        self.program_exe = program_exe
-
-
-class CompressPass(object):
-    """
-    The pass used to compress model.
-    Args:
-        place: The device used in compression.
-        data_reader: The data_reader used to run graph.
-        data_feeder: The data_feeder used to run graph.
-        scope: The scope used to run graph.
-        metrics: The metrics for evaluating model.
-        epoch: The total epoches of trainning in compression.
-        program_exe: The program_exe is used to execute the program
-                     created for modifying the variables in scope.
-    """
-
-    def __init__(self,
-                 place=None,
-                 data_reader=None,
-                 data_feeder=None,
-                 scope=None,
-                 metrics=None,
-                 epoch=None,
-                 program_exe=None):
-        self.strategies = []
-        self.place = CPUPlace() if place is None else place
-        self.data_reader = data_reader
-        self.data_feeder = data_feeder
-        self.scope = scope
-        self.metrics = metrics
-        self.epoch = epoch
-        self.program_exe = program_exe
-
-    def add_strategy(self, strategy):
-        """
-        Add a strategy to current compress pass.
-        Args:
-            strategy: The strategy to be added into current compress pass.
-        """
-        self.strategies.append(strategy)
-        self.epoch = max(strategy.end_epoch, self.epoch)
-
-    def apply(self, graph):
-        """
-        Compress a model.
-        Args:
-            graph: The target graph to be compressed.
-        """
-        self.executor = get_executor(graph, self.place)
-        context = Context(
-            self.executor, graph, self.scope, program_exe=self.program_exe)
-
-        for strategy in self.strategies:
-            strategy.on_compress_begin(context)
-
-        for epoch in range(self.epoch):
-
-            for strategy in self.strategies:
-                strategy.on_epoch_begin(context)
-
-            for data in self.data_reader():
-
-                for strategy in self.strategies:
-                    strategy.on_batch_begin(context)
-                fetches = None
-                if self.metrics:
-                    fetches = self.metrics.values()
-                feed = None
-                if self.data_feeder:
-                    feed = self.data_feeder.feed(data)
-                results = self.executor.run(graph,
-                                            fetches=fetches,
-                                            scope=self.scope,
-                                            feed=feed)
-                if results:
-                    print("results: {}".format(
-                        zip(self.metrics.keys(), results)))
-                for strategy in self.strategies:
-                    strategy.on_batch_end(context)
-                context.batch_id += 1
-
-            for strategy in self.strategies:
-                strategy.on_epoch_end(context)
-            context.epoch_id += 1
-
-        for strategy in self.strategies:
-            strategy.on_compress_end(context)

python/paddle/fluid/contrib/slim/core/compressor.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ....core import CPUPlace
+from .... import compiler
+from .... import io
+from .... import profiler
+from .... import scope_guard
+from ....data_feeder import DataFeeder
+from ..graph import *
+from .config import ConfigFactory
+import numpy as np
+from collections import Iterable
+import time
+import os
+import logging
+import sys
+import pickle
+import functools
+
+__all__ = ['Context', 'Compressor']
+
+logging.basicConfig(format='%(asctime)s-%(levelname)s: %(message)s')
+_logger = logging.getLogger(__name__)
+_logger.setLevel(logging.INFO)
+
+
+def cached_reader(reader, sampled_rate, cache_path, cached_id):
+    """
+    Sample partial data from reader and cache them into local file system.
+    Args:
+        reader: Iterative data source.
+        sampled_rate(float): The sampled rate used to sample partial data for evaluation. None means using all data in eval_reader. default: None.
+        cache_path(str): The path to cache the sampled data.
+        cached_id(int): The id of dataset sampled. Evaluations with same cached_id use the same sampled dataset. default: 0.
+    """
+    np.random.seed(cached_id)
+    cache_path = os.path.join(cache_path, str(cached_id))
+    _logger.debug('read data from: {}'.format(cache_path))
+
+    def s_reader():
+        if os.path.isdir(cache_path):
+            for file_name in open(os.path.join(cache_path, "list")):
+                yield np.load(os.path.join(cache_path, file_name.strip()))
+        else:
+            os.makedirs(cache_path)
+            list_file = open(os.path.join(cache_path, "list"), 'w')
+            batch = 0
+            dtype = None
+            for data in reader():
+                if batch == 0 or (np.random.uniform() < sampled_rate):
+                    np.save(
+                        os.path.join(cache_path, 'batch' + str(batch)), data)
+                    list_file.write('batch' + str(batch) + '.npy\n')
+                    batch += 1
+                    yield data
+
+    return s_reader
+
+
+class Context(object):
+    """
+    The context in the process of compression.
+    """
+
+    def __init__(self,
+                 place,
+                 scope,
+                 train_graph=None,
+                 train_reader=None,
+                 eval_graph=None,
+                 eval_reader=None,
+                 teacher_graphs=None,
+                 train_optimizer=None,
+                 distiller_optimizer=None):
+        """
+        Args:
+            place: The device place where the compression job running.
+            scope: The scope used in compression job.
+            train_graph: The graph with loss as output node.
+            eval_graph: The graph used for evaluation.
+            eval_reader: The data reader used for evaluation.
+            teacher_graphs: The teacher graphs used in distillation strategies.
+            train_optimizer: The optimizer used to append backward ops and
+                             optimization ops into train_graph.
+            distiller_optimizer: The optimizer used by distillation strategies.
+        """
+        # The total number of epoches to be trained.
+        self.epoch = 0
+        # Current epoch
+        self.epoch_id = 0
+        # Current batch
+        self.batch_id = 0
+
+        self.k_v = {}
+
+        self.place = place
+        self.scope = scope
+        self.train_graph = train_graph
+        self.train_reader = train_reader
+        self.eval_graph = eval_graph
+        self.eval_reader = eval_reader
+        self.executor = None
+        self.teacher_graphs = teacher_graphs
+        self.train_optimizer = train_optimizer
+        self.distiller_optimizer = distiller_optimizer
+        self.optimize_graph = None
+        self.cache_path = './eval_cache'
+        self.eval_results = {}
+
+    def to_file(self, file_name):
+        """
+        Save the context into file.
+        """
+        data = {}
+        data['epoch_id'] = self.epoch_id
+        data['eval_results'] = self.eval_results
+        with open(file_name, 'wb') as context_file:
+            pickle.dump(data, context_file)
+
+    def from_file(self, file_name):
+        """
+        Load the context from file.
+        """
+        with open(file_name) as context_file:
+            if sys.version_info < (3, 0):
+                data = pickle.load(context_file)
+            else:
+                data = pickle.load(context_file, encoding='bytes')
+            self.epoch_id = data['epoch_id']
+            self.eval_results = data['eval_results']
+
+    def eval_converged(self, metric_name, delta=0.001):
+        """
+        Check whether the training has been converged.
+        Args:
+            metric_name(str): The metric used to check convergence.
+            delta(float): '(metric[k] - metric[k-1] / metric[k-1]) < delta'
+                          means that the training has been converged.
+        Returns:
+            bool: True means the training has been converged.
+        """
+        # TODO(wanghaoshuang@baidu.com): enhence this method.
+        if (metric_name not in self.eval_results
+            ) or len(self.eval_results[metric_name]) < 2:
+            return False
+        results = self.eval_results[metric_name][-2:]
+        _logger.info('Latest evaluations: {}'.format(results))
+        return abs(results[1] - results[0]) / results[0] < delta
+
+    def run_eval_graph(self, sampled_rate=None, cached_id=0):
+        """
+        Evaluate the current mode in context.
+        Args:
+            sampled_rate(float): The sampled rate used to sample partial data
+            for evaluation. None means using all data in eval_reader. default: None.
+            cached_id(int): The id of dataset sampled. Evaluations with same
+                            cached_id use the same sampled dataset. default: 0.
+        """
+        _logger.info('Running evaluation')
+        assert self.eval_graph is not None
+        assert self.eval_reader is not None
+        eval_graph = self.eval_graph.clone(for_test=True)
+
+        executor = SlimGraphExecutor(self.place)
+        results = []
+        batch_id = 0
+        s_time = time.time()
+        reader = self.eval_reader
+        if sampled_rate:
+            reader = cached_reader(reader, sampled_rate, self.cache_path,
+                                   cached_id)
+        for data in reader():
+            result = executor.run(eval_graph, self.scope, data=data)
+            result = [np.mean(r) for r in result]
+            results.append(result)
+            if batch_id % 20 == 0:
+                _logger.info("batch-{}; {}={}".format(
+                    batch_id, eval_graph.out_nodes.keys(), result))
+            batch_id += 1
+        result = np.mean(np.array(results), axis=0)
+        _logger.info("Final eval result: {}={}".format(
+            eval_graph.out_nodes.keys(), result))
+        if not isinstance(result, Iterable):
+            result = [result]
+        _logger.info('Finish evaluation')
+        return result, eval_graph.out_nodes.keys()
+
+    def put(self, key, value):
+        self.k_v[key] = value
+
+    def get(self, key):
+        return self.k_v.get(key)
+
+
+class Compressor(object):
+    """
+    The pass used to compress model.
+    """
+
+    def __init__(self,
+                 place,
+                 scope,
+                 train_program,
+                 train_reader=None,
+                 train_feed_list=None,
+                 train_fetch_list=None,
+                 eval_program=None,
+                 eval_reader=None,
+                 eval_feed_list=None,
+                 eval_fetch_list=None,
+                 teacher_programs=[],
+                 checkpoint_path='./checkpoints',
+                 train_optimizer=None,
+                 distiller_optimizer=None):
+        """
+        Args:
+            place(fluid.Place): The device place where the compression job running.
+            scope(fluid.core.Scope): The scope used to run graph.
+            train_program(Program): The main program to be compressed. It must have loss op.
+            train_reader: The data reader used for training.
+            train_feed_list(dict): A dict to indicate the input variable of the training program.
+                                   The key is user-defined and human-readable name.
+                                   The value is the name of Variable.
+            train_fetch_list(dict): A dict to indicate the output variable of the training program.
+                                   The key is user-defined and human-readable name.
+                                   The value is the name of Variable.
+            eval_program(Program): The program used for evaluation.
+            eval_reader: The data reader used for evaluation.
+            eval_feed_list(dict): A dict to indicate the input variable of the evaluation program.
+                                   The key is user-defined and human-readable name.
+                                   The value is the name of Variable.
+            eval_fetch_list(dict): A dict to indicate the output variable of the evaluation program.
+                                   The key is user-defined and human-readable name.
+                                   The value is the name of Variable.
+            teacher_programs: The teacher graphs used in distillation strategies.
+            train_optimizer: The optimizer used to append backward ops and
+                             optimization ops into train_graph.
+            distiller_optimizer: The optimizer used by distillation strategies. In distillation strategy,
+                                 this optimizer is used to minimize the combined loss of student-net and
+                                 teacher-net while train_optimizer is used to minimize loss of
+                                 student-net in fine-tune stage. 
+
+        """
+        assert isinstance(
+            train_feed_list, list
+        ), "train_feed_list should be a list of tuple, such as [('image', image.name), ('label', gt.name)]"
+        assert isinstance(
+            eval_feed_list, list
+        ), "eval_feed_list should be a list of tuple, such as [('image', image.name), ('label', gt.name)]"
+        self.strategies = []
+        self.epoch = 0
+        self.place = CPUPlace() if place is None else place
+        self.scope = scope
+        self.train_graph = GraphWrapper(
+            train_program, in_nodes=train_feed_list, out_nodes=train_fetch_list)
+        self.eval_graph = GraphWrapper(
+            eval_program, in_nodes=eval_feed_list, out_nodes=eval_fetch_list)
+        self.train_reader = train_reader
+        self.eval_reader = eval_reader
+        self.teacher_graphs = []
+        for teacher in teacher_programs:
+            self.teacher_graphs.append(GraphWrapper(teacher))
+
+        self.checkpoint = None
+        self.checkpoint_path = checkpoint_path
+        self.eval_epoch = 1
+
+        self.train_optimizer = train_optimizer
+        self.distiller_optimizer = distiller_optimizer
+        self.init_model = None
+
+    def _add_strategy(self, strategy):
+        """
+        Add a strategy to current compress pass.
+        Args:
+            strategy: The strategy to be added into current compress pass.
+        """
+        self.strategies.append(strategy)
+        self.epoch = max(strategy.end_epoch, self.epoch)
+
+    def config(self, config_file):
+        """
+        Configure the compress pass from file with yaml format.
+        Args:
+            config_file(str): The config file in local file system.
+        """
+        factory = ConfigFactory(config_file)
+        self.epoch = factory.compressor['epoch']
+        for strategy in factory.compressor['strategies']:
+            self._add_strategy(strategy)
+        if 'checkpoint_path' in factory.compressor:
+            self.checkpoint_path = factory.compressor['checkpoint_path']
+
+        if 'init_model' in factory.compressor:
+            self.init_model = factory.compressor['init_model']
+
+    def _init_model(self, context):
+        """
+        Load model that has been compressed. 
+        """
+        if self.init_model and os.path.exists(self.init_model):
+            exe = SlimGraphExecutor(context.place)
+            with scope_guard(context.scope):
+                context.train_graph.load_persistables(self.init_model, exe)
+            flops = context.eval_graph.flops()
+            conv_flops = context.eval_graph.flops(only_conv=True)
+            context.eval_graph.update_param_shape(context.scope)
+            context.eval_graph.update_groups_of_conv()
+            _logger.info("conv flops: -{}".format(1 - float(
+                context.eval_graph.flops(only_conv=True)) / conv_flops))
+            _logger.info("total flops: -{}".format(1 - float(
+                context.eval_graph.flops()) / flops))
+            context.train_graph.update_param_shape(context.scope)
+            context.train_graph.update_groups_of_conv()
+            context.train_graph.infer_shape()
+            _logger.info("Init model from: {}".format(self.init_model))
+
+    def _load_checkpoint(self, context):
+        """
+        Load checkpoints from file.
+        """
+        _logger.debug('_load_checkpoint')
+        strategies = self.strategies
+        if self.checkpoint_path:
+            if not os.path.exists(self.checkpoint_path):
+                _logger.warning("Checkpints path doesn't exist: [{}]".format(
+                    self.checkpoint_path))
+                return context, strategies
+            checkpoints = [
+                dir for dir in os.listdir(self.checkpoint_path)
+                if os.path.isdir(os.path.join(self.checkpoint_path, dir))
+            ]
+            _logger.debug('self.checkpoint_path: {}'.format(
+                self.checkpoint_path))
+            _logger.info('checkpoints: {}'.format(checkpoints))
+            if len(checkpoints) > 0:
+                latest = max([int(ck) for ck in checkpoints])
+                latest_ck_path = os.path.join(self.checkpoint_path, str(latest))
+
+                model_path = os.path.join(latest_ck_path, 'model')
+                context_path = os.path.join(latest_ck_path, 'context')
+                strategy_path = os.path.join(latest_ck_path, 'strategies')
+                if os.path.exists(context_path):
+                    context.from_file(context_path)
+                    context.epoch_id += 1
+                if os.path.exists(strategy_path):
+                    with open(strategy_path, 'rb') as strategy_file:
+                        if sys.version_info < (3, 0):
+                            strategies = pickle.load(strategy_file)
+                        else:
+                            strategies = pickle.load(
+                                strategy_file, encoding='bytes')
+
+                if os.path.exists(model_path):
+                    exe = SlimGraphExecutor(context.place)
+                    with scope_guard(context.scope):
+                        context.optimize_graph.load_persistables(model_path,
+                                                                 exe)
+                    context.optimize_graph.update_param_shape(context.scope)
+                    context.optimize_graph.update_groups_of_conv()
+                    context.eval_graph.update_param_shape(context.scope)
+                    context.eval_graph.update_groups_of_conv()
+                    _logger.info("Loaded params from: {}".format(model_path))
+        return context, strategies
+
+    def _save_checkpoint(self, context):
+        """
+        Save checkpoints to file.
+        """
+        if context.epoch_id % 1 == 0 and self.checkpoint_path:
+            checkpoint_path = os.path.join(self.checkpoint_path,
+                                           str(context.epoch_id))
+            model_path = os.path.join(checkpoint_path, 'model')
+            context_path = os.path.join(checkpoint_path, 'context')
+            strategy_path = os.path.join(checkpoint_path, 'strategies')
+            if not os.path.isdir(model_path):
+                os.makedirs(model_path)
+            exe = SlimGraphExecutor(context.place)
+            with scope_guard(context.scope):
+                context.optimize_graph.save_persistables(model_path, exe)
+            context.to_file(context_path)
+            with open(strategy_path, 'wb') as strategy_file:
+                pickle.dump(self.strategies, strategy_file)
+            _logger.info('Saved checkpoint to: {}'.format(checkpoint_path))
+
+    def _train_one_epoch(self, context):
+        """
+        Train one epoch.
+        """
+
+        executor = SlimGraphExecutor(self.place)
+
+        if context.optimize_graph.compiled_graph is None:
+            context.optimize_graph.compiled_graph = compiler.CompiledProgram(
+                context.optimize_graph.program).with_data_parallel(
+                    loss_name=context.optimize_graph.out_nodes['loss'])
+
+        for data in context.train_reader():
+            for strategy in self.strategies:
+                strategy.on_batch_begin(context)
+            results = executor.run(context.optimize_graph,
+                                   context.scope,
+                                   data=data)
+            results = [float(np.mean(result)) for result in results]
+            if context.batch_id % 20 == 0:
+                _logger.info("epoch:{}; batch_id:{}; {} = {}".format(
+                    context.epoch_id, context.batch_id,
+                    context.optimize_graph.out_nodes.keys(
+                    ), [round(r, 3) for r in results]))
+            for strategy in self.strategies:
+                strategy.on_batch_end(context)
+            context.batch_id += 1
+        context.batch_id = 0
+
+    def _eval(self, context):
+        """
+        Runing evaluation.
+        """
+        results, names = context.run_eval_graph()
+        for name, result in zip(names, results):
+            if name not in context.eval_results:
+                context.eval_results[name] = []
+            context.eval_results[name].append(result)
+
+    def run(self):
+        """
+        Execute compressiong pass.
+        """
+        context = Context(
+            place=self.place,
+            scope=self.scope,
+            train_graph=self.train_graph,
+            train_reader=self.train_reader,
+            eval_graph=self.eval_graph,
+            eval_reader=self.eval_reader,
+            teacher_graphs=self.teacher_graphs,
+            train_optimizer=self.train_optimizer,
+            distiller_optimizer=self.distiller_optimizer)
+        self.context = context
+        if self.teacher_graphs:
+            context.put('teachers', self.teacher_graphs)
+        self._init_model(context)
+        if not context.optimize_graph:
+            if context.train_optimizer:
+                context.train_optimizer._name = 'train_opt'
+                context.optimize_graph = context.train_graph.get_optimize_graph(
+                    context.train_optimizer, context.place, context.scope)
+            else:
+                context.optimize_graph = context.train_graph
+
+        context, self.strategies = self._load_checkpoint(context)
+
+        for strategy in self.strategies:
+            strategy.on_compression_begin(context)
+        start = context.epoch_id
+        self._eval(context)
+        for epoch in range(start, self.epoch):
+            context.epoch_id = epoch
+            for strategy in self.strategies:
+                strategy.on_epoch_begin(context)
+            self._train_one_epoch(context)
+            for strategy in self.strategies:
+                strategy.on_epoch_end(context)
+            if self.eval_epoch and epoch % self.eval_epoch == 0:
+                self._eval(context)
+            self._save_checkpoint(context)
+        for strategy in self.strategies:
+            strategy.on_compression_end(context)
+        return context.eval_graph

python/paddle/fluid/contrib/slim/core/config.py

@@ -17,8 +17,9 @@ import funcsigs
 import yaml
 from collections import OrderedDict
 from ..prune import *
-from .compress_pass import *
+from ..quantization import *
 from .strategy import *
+from ..distillation import *
 
 __all__ = ['ConfigFactory']
 """This factory is used to create instances by loading and parsing configure file with yaml format.
@@ -29,15 +30,10 @@ class ConfigFactory(object):
     def __init__(self, config):
         """Init a factory from configure file."""
         self.instances = {}
+        self.compressor = {}
         self.version = None
         self._parse_config(config)
 
-    def get_compress_pass(self):
-        """
-        Get compress pass from factory.
-        """
-        return self.instance('compress_pass')
-
     def instance(self, name):
         """
         Get instance from factory.
@@ -59,8 +55,16 @@ class ConfigFactory(object):
             args = {}
             for key in keys:
                 value = attrs[key]
+                if isinstance(value, str) and value.lower() == 'none':
+                    value = None
                 if isinstance(value, str) and value in self.instances:
                     value = self.instances[value]
+                if isinstance(value, list):
+                    for i in range(len(value)):
+                        if isinstance(value[i],
+                                      str) and value[i] in self.instances:
+                            value[i] = self.instances[value[i]]
+
                 args[key] = value
             self.instances[name] = class_(**args)
         return self.instances.get(name)
@@ -76,16 +80,23 @@ class ConfigFactory(object):
                     assert self.version == int(key_values['version'])
 
                 # parse pruners
-                if key == 'pruners' or key == 'strategies':
+                if key == 'distillers' or key == 'pruners' or key == 'quantizers' or key == 'strategies':
                     instances = key_values[key]
                     for name in instances:
                         self._new_instance(name, instances[name])
 
-                if key == 'compress_pass':
-                    compress_pass = self._new_instance(key, key_values[key])
-                    for name in key_values[key]['strategies']:
-                        strategy = self.instance(name)
-                        compress_pass.add_strategy(strategy)
+                if key == 'compressor':
+                    self.compressor['strategies'] = []
+                    self.compressor['epoch'] = key_values[key]['epoch']
+                    if 'init_model' in key_values[key]:
+                        self.compressor['init_model'] = key_values[key][
+                            'init_model']
+                    self.compressor['checkpoint_path'] = key_values[key][
+                        'checkpoint_path']
+                    if 'strategies' in key_values[key]:
+                        for name in key_values[key]['strategies']:
+                            strategy = self.instance(name)
+                            self.compressor['strategies'].append(strategy)
 
                 if key == 'include':
                     for config_file in key_values[key]:

python/paddle/fluid/contrib/slim/core/pass_builder.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from .compress_pass import CompressPass
-from .config import ConfigFactory
-
-__all__ = ['build_compressor']
-
-
-def build_compressor(place=None,
-                     data_reader=None,
-                     data_feeder=None,
-                     scope=None,
-                     metrics=None,
-                     epoch=None,
-                     config=None):
-    if config is not None:
-        factory = ConfigFactory(config)
-        comp_pass = factory.get_compress_pass()
-    else:
-        comp_pass = CompressPass()
-    comp_pass.place = place
-    comp_pass.data_reader = data_reader
-    comp_pass.data_feeder = data_feeder
-    comp_pass.scope = scope
-    comp_pass.metrics = metrics
-    comp_pass.epoch = epoch
-    return comp_pass

python/paddle/fluid/contrib/slim/core/strategy.py

@@ -20,7 +20,7 @@ class Strategy(object):
     Base class for all strategies.
     """
 
-    def __init__(self, start_epoch=0, end_epoch=10):
+    def __init__(self, start_epoch=0, end_epoch=0):
         """
         Args:
             start_epoch: The first epoch to apply the strategy.
@@ -29,7 +29,7 @@ class Strategy(object):
         self.start_epoch = start_epoch
         self.end_epoch = end_epoch
 
-    def on_compress_begin(self, context):
+    def on_compression_begin(self, context):
         pass
 
     def on_epoch_begin(self, context):
@@ -44,5 +44,5 @@ class Strategy(object):
     def on_batch_end(self, context):
         pass
 
-    def on_compress_end(self, context):
+    def on_compression_end(self, context):
         pass

python/paddle/fluid/contrib/slim/demo/filter_prune/config.yaml

-version: 1.0
-pruners:
-    pruner_1:
-        class: 'RatioPruner'
-        ratios:
-            'conv1_1.w': 0.3
-            'conv1_2.w': 0.4
-            '*': 0.9
-        group_dims:
-            '*': [1, 2, 3]
-        criterions:
-            '*': 'l1-norm'
-strategies:
-    strategy_1:
-        class: 'SensitivePruneStrategy'
-        pruner: 'pruner_1'
-        start_epoch: 0
-        end_epoch: 10
-        delta_rate: 0.20
-        acc_loss_threshold: 0.2
-        sensitivities:
-            'conv1_1.w': 0.4
-
-compress_pass:
-    class: 'CompressPass'
-    epoch: 100
-    strategies:
-        - strategy_1

python/paddle/fluid/contrib/slim/demo/filter_prune/demo.py

-# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import paddle.fluid as fluid
-import paddle
-import os
-import sys
-from paddle.fluid.contrib.slim import CompressPass
-from paddle.fluid.contrib.slim import build_compressor
-from paddle.fluid.contrib.slim import ImitationGraph
-
-
-class LinearModel(object):
-    def __init__(slef):
-        pass
-
-    def train(self):
-        train_program = fluid.Program()
-        startup_program = fluid.Program()
-        startup_program.random_seed = 10
-        with fluid.program_guard(train_program, startup_program):
-            x = fluid.layers.data(name='x', shape=[13], dtype='float32')
-            y = fluid.layers.data(name='y', shape=[1], dtype='float32')
-            predict = fluid.layers.fc(input=x, size=1, act=None)
-            cost = fluid.layers.square_error_cost(input=predict, label=y)
-            avg_cost = fluid.layers.mean(cost)
-            eval_program = train_program.clone()
-            sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-            sgd_optimizer.minimize(avg_cost)
-
-        train_reader = paddle.batch(
-            paddle.dataset.uci_housing.train(), batch_size=1)
-        eval_reader = paddle.batch(
-            paddle.dataset.uci_housing.test(), batch_size=1)
-        place = fluid.CPUPlace()
-        train_feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
-        eval_feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
-        exe = fluid.Executor(place)
-        exe.run(startup_program)
-        train_metrics = {"loss": avg_cost.name}
-        eval_metrics = {"loss": avg_cost.name}
-
-        graph = ImitationGraph(train_program)
-        config = './config.yaml'
-        comp_pass = build_compressor(
-            place,
-            data_reader=train_reader,
-            data_feeder=train_feeder,
-            scope=fluid.global_scope(),
-            metrics=train_metrics,
-            epoch=1,
-            config=config)
-        comp_pass.apply(graph)
-
-
-if __name__ == "__main__":
-    model = LinearModel()
-    model.train()

python/paddle/fluid/contrib/slim/graph/graph_pass.py → python/paddle/fluid/contrib/slim/distillation/__init__.py

-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -12,31 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__all__ = ['GraphPass', 'PruneParameterPass']
+from . import distiller
+from .distiller import *
+from . import distillation_strategy
+from .distillation_strategy import *
 
-
-class GraphPass(object):
-    """
-    Base class for all graph pass.
-    """
-
-    def __init__(self):
-        pass
-
-    def apply(self, graph):
-        pass
-
-
-class PruneParameterPass(GraphPass):
-    """
-    Generate a graph for pruning parameters from target graph.
-    """
-
-    def __init__(self, pruned_params, thresholds):
-        super(PruneParameterPass, self).__init__()
-        self.pruned_params = pruned_params
-        self.thresholds = thresholds
-        self.default_threshold = thresholds['*']
-
-    def apply(self, graph):
-        pass
+__all__ = distiller.__all__
+__all__ += distillation_strategy.__all__

python/paddle/fluid/contrib/slim/distillation/distillation_strategy.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..core.strategy import Strategy
+from ....framework import Program, program_guard
+from .... import Executor
+import logging
+
+__all__ = ['DistillationStrategy']
+
+logging.basicConfig(format='%(asctime)s-%(levelname)s: %(message)s')
+_logger = logging.getLogger(__name__)
+_logger.setLevel(logging.INFO)
+
+
+class DistillationStrategy(Strategy):
+    def __init__(self, distillers=None, start_epoch=0, end_epoch=0):
+        """
+        Args:
+            distillers(list): A list of distiller used to combine student graph and teacher graph
+                              by adding some loss.
+            start_epoch(int): The epoch when to merge student graph and teacher graph for
+                              distillation training. default: 0
+            end_epoch(int): The epoch when to finish distillation training. default: 0
+            
+        """
+        super(DistillationStrategy, self).__init__(start_epoch, end_epoch)
+        self.distillers = distillers
+
+    def on_compression_begin(self, context):
+        # load from checkpoint
+        if context.epoch_id > 0:
+            if context.epoch_id > self.start_epoch and context.epoch_id < self.end_epoch:
+                _logger.info('Restore DistillationStrategy')
+                self._create_distillation_graph(context)
+                _logger.info('Restore DistillationStrategy finish.')
+
+    def on_epoch_begin(self, context):
+        if self.start_epoch == context.epoch_id:
+            _logger.info('DistillationStrategy::on_epoch_begin.')
+            self._create_distillation_graph(context)
+            _logger.info('DistillationStrategy set optimize_graph.')
+
+    def _create_distillation_graph(self, context):
+        """
+        step 1: Merge student graph and teacher graph into distillation graph.
+        step 2: Add loss into distillation graph by distillers.
+        step 3: Append backward ops and optimize ops into distillation graph for training.
+        """
+        # step 1
+        teacher = context.teacher_graphs[0]
+        for var in teacher.program.list_vars():
+            var.stop_gradient = True
+        graph = context.train_graph.clone()
+        graph.merge(teacher)
+        graph.out_nodes['student_loss'] = graph.out_nodes['loss']
+
+        # step 2
+        for distiller in self.distillers:
+            graph = distiller.distiller_loss(graph)
+
+        # step 3
+        startup_program = Program()
+        with program_guard(graph.program, startup_program):
+            context.distiller_optimizer._name = 'distillation_optimizer'
+            context.distiller_optimizer.minimize(
+                graph.var(graph.out_nodes['loss'])._var)
+        exe = Executor(context.place)
+        exe.run(startup_program, scope=context.scope)
+
+        # backup graph for fine-tune after distillation
+        context.put('distillation_backup_optimize_graph',
+                    context.optimize_graph)
+        context.optimize_graph = graph
+
+    def on_epoch_end(self, context):
+        if context.epoch_id == (self.end_epoch - 1):
+            _logger.info('DistillationStrategy::on_epoch_end.')
+            # restore optimize_graph for fine-tune or other strategy in next stage.
+            context.optimize_graph = context.get(
+                'distillation_backup_optimize_graph')
+            _logger.info(
+                'DistillationStrategy set context.optimize_graph to None.')

python/paddle/fluid/contrib/slim/distillation/distiller.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .... import layers
+from .... import optimizer
+from .... import Executor
+from .... import Program
+from .... import program_guard
+from .... import regularizer
+
+__all__ = ['FSPDistiller', 'L2Distiller']
+
+
+class L2Distiller(object):
+    """
+    Combine two layers from student net and teacher net by l2-loss.
+    And add the loss into the total loss using for distillation training.
+    """
+
+    def __init__(self,
+                 student_feature_map,
+                 teacher_feature_map,
+                 distillation_loss_weight=1):
+        """
+        Args:
+            student_feature_map(str): The name of feature map from student network.
+            teacher_feature_map(str): The name of feature map from teacher network.
+                                      It's shape should be the same with student network.
+            distillation_loss_weight(float): The weight of the l2-loss.
+        """
+        self.student_feature_map = student_feature_map
+        self.teacher_feature_map = teacher_feature_map
+        self.distillation_loss_weight = distillation_loss_weight
+
+    def distiller_loss(self, graph):
+        """
+        Modify graph inplace to add l2-loss.
+        Args: 
+            graph(GraphWrapper): The graph to be modified.
+        Returns:
+            GraphWrapper: The modified graph.
+        """
+        distiller_pass = L2DistillerPass(self.student_feature_map,
+                                         self.teacher_feature_map,
+                                         self.distillation_loss_weight)
+        dis_graph = distiller_pass.apply(graph)
+        return dis_graph
+
+
+class L2DistillerPass(object):
+    """
+    The pass used to add l2-loss.
+    """
+
+    def __init__(self,
+                 student_feature_map,
+                 teacher_feature_map,
+                 distillation_loss_weight=1):
+        """
+        Args:
+            student_feature_map(str): The name of feature map from student network.
+            teacher_feature_map(str): The name of feature map from teacher network.
+                                      It's shape should be the same with student network.
+            distillation_loss_weight(float): The weight of the l2-loss.
+        """
+        self.student_feature_map = student_feature_map
+        self.teacher_feature_map = teacher_feature_map
+        self.distillation_loss_weight = distillation_loss_weight
+
+    def apply(self, graph):
+        ret_graph = graph
+        with program_guard(ret_graph.program):
+
+            student_feature_map = ret_graph.var(self.student_feature_map)._var
+            teacher_feature_map = ret_graph.var(self.teacher_feature_map)._var
+            l2loss = layers.reduce_mean(
+                layers.square(student_feature_map - teacher_feature_map))
+
+            distillation_loss = l2loss * self.distillation_loss_weight
+            student_loss = ret_graph.var(ret_graph.out_nodes['loss'])._var
+            loss = distillation_loss + student_loss
+
+            ret_graph.out_nodes[
+                'l2loss_' + self.student_feature_map + "_" +
+                self.teacher_feature_map] = distillation_loss.name
+            ret_graph.out_nodes['loss'] = loss.name
+        return ret_graph
+
+
+class FSPDistiller(object):
+    """
+    Combine layers from student net and teacher net by fsp-loss.
+    """
+
+    def __init__(self, student_pairs, teacher_pairs,
+                 distillation_loss_weight=1):
+        """
+        Args:
+            student_pairs(list<tuple>): Each tuple, with two variable names, in student_pairs indicates
+                                        a section in student network. The variables in a tuple should
+                                        have the same feature map size.
+            teacher_pairs(list<tuple>): Each tuple, with two variable names, in teacher_pairs indicates
+                                        a section in teacher network. The variables in a tuple should
+                                        have the same feature map size. Varibale named teacher_pairs[i][j]
+                                        should has the save channel number with that of variable named 
+                                        student_pairs[i][j].
+
+            distillation_loss_weight(float): The weight of the fsp-loss. default: 1.
+        """
+        self.student_pairs = student_pairs
+        self.teacher_pairs = teacher_pairs
+        self.distillation_loss_weight = distillation_loss_weight
+
+    def distiller_loss(self, graph):
+        """
+        Modify graph inplace to add fsp-loss.
+        Args: 
+            graph(GraphWrapper): The graph to be modified.
+        Returns:
+            GraphWrapper: The modified graph.
+        """
+        distiller_pass = FSPDistillerPass(self.student_pairs,
+                                          self.teacher_pairs,
+                                          self.distillation_loss_weight)
+        dis_graph = distiller_pass.apply(graph)
+        return dis_graph
+
+
+class FSPDistillerPass(object):
+    '''
+    Combine layers from student net and teacher net by fsp-loss.
+    '''
+
+    def __init__(self, s_pairs, t_pairs, distillation_loss_weight=1):
+        """
+        Args:
+            s_pairs(list<tuple>): Each tuple, with two variable names, in student_pairs indicates
+                                        a section in student network. The variables in a tuple should
+                                        have the same feature map size.
+            t_pairs(list<tuple>): Each tuple, with two variable names, in teacher_pairs indicates
+                                        a section in teacher network. The variables in a tuple should
+                                        have the same feature map size. Varibale named teacher_pairs[i][j]
+                                        should has the save channel number with that of variable named 
+                                        student_pairs[i][j].
+
+            distillation_loss_weight(float): The weight of the fsp-loss. default: 1.
+        """
+        self.s_pairs = s_pairs
+        self.t_pairs = t_pairs
+        self.distillation_loss_weight = distillation_loss_weight
+
+    def apply(self, graph):
+        ret_graph = graph
+        with program_guard(ret_graph.program):
+            losses = []
+            for s_pair, t_pair in zip(self.s_pairs, self.t_pairs):
+                s_pair_start = ret_graph.var(s_pair[0])._var
+                s_pair_end = ret_graph.var(s_pair[1])._var
+                s_fsp_matrix = self._fsp_matrix(s_pair_start, s_pair_end)
+                t_pair_start = ret_graph.var(t_pair[0])._var
+                t_pair_end = ret_graph.var(t_pair[1])._var
+                t_fsp_matrix = self._fsp_matrix(t_pair_start, t_pair_end)
+                l2_loss = layers.reduce_mean(
+                    layers.square(s_fsp_matrix - t_fsp_matrix))
+                losses.append(l2_loss)
+            distillation_loss = layers.sum(
+                losses) * self.distillation_loss_weight
+            student_loss = ret_graph.var(ret_graph.out_nodes['loss'])._var
+            loss = distillation_loss + student_loss
+
+            ret_graph.out_nodes[
+                'fsp_distillation_loss'] = distillation_loss.name
+            ret_graph.out_nodes['loss'] = loss.name
+        return ret_graph
+
+    def _fsp_matrix(self, fea_map_0, fea_map_1):
+        return layers.fsp_matrix(fea_map_0, fea_map_1)

python/paddle/fluid/contrib/slim/graph/__init__.py

@@ -14,10 +14,7 @@
 
 from . import executor
 from .executor import *
-from . import graph
-from .graph import *
-from . import graph_pass
-from .graph_pass import *
+from . import graph_wrapper
+from .graph_wrapper import *
 __all__ = executor.__all__
-__all__ += graph.__all__
-__all__ += graph_pass.__all__
+__all__ += graph_wrapper.__all__

python/paddle/fluid/contrib/slim/graph/executor.py

@@ -12,51 +12,46 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import abc
-from abc import abstractmethod
+from ....compiler import CompiledProgram
+from ....data_feeder import DataFeeder
 from .... import executor
-from .graph import IRGraph, ImitationGraph
+from .graph_wrapper import GraphWrapper
 
-__all__ = ['get_executor']
+__all__ = ['SlimGraphExecutor']
 
 
-class GraphExecutor(object):
-    __metaclass__ = abc.ABCMeta
+class SlimGraphExecutor(object):
+    """
+    Wrapper of executor used to run GraphWrapper.
+    """
 
     def __init__(self, place):
-        self.place = place
-
-    @abstractmethod
-    def run(self, graph, feches=None, feed=None):
-        pass
-
-
-class IRGraphExecutor(GraphExecutor):
-    def run(self, grah, fetches, feed=None):
-        pass
-
-
-class ImitationGraphExecutor(GraphExecutor):
-    def __init__(self, place):
-        super(ImitationGraphExecutor, self).__init__(place)
         self.exe = executor.Executor(place)
+        self.place = place
 
-    def run(self, graph, scope=None, fetches=None, feed=None):
-        assert isinstance(graph, ImitationGraph)
-        fetch_list = None
-        if fetches:
-            fetch_list = [
-                graph.program.global_block().var(name) for name in fetches
-            ]
-        results = self.exe.run(graph.program,
+    def run(self, graph, scope, data=None):
+        """
+        Runing a graph with a batch of data.
+        Args:
+            graph(GraphWrapper): The graph to be executed.
+            scope(fluid.core.Scope): The scope to be used.
+            data(list<tuple>): A batch of data. Each tuple in this list is a sample.
+                               It will feed the items of tuple to the in_nodes of graph.
+        Returns:
+            results(list): A list of result with the same order indicated by graph.out_nodes.
+        """
+        assert isinstance(graph, GraphWrapper)
+        if data is not None:
+            feeder = DataFeeder(
+                feed_list=graph.in_nodes.values(),
+                place=self.place,
+                program=graph.program)
+            feed = feeder.feed(data)
+
+        fetch_list = graph.out_nodes.values()
+        program = graph.compiled_graph if graph.compiled_graph else graph.program
+        results = self.exe.run(program,
                                scope=scope,
                                fetch_list=fetch_list,
                                feed=feed)
         return results
-
-
-def get_executor(graph, place):
-    if isinstance(graph, ImitationGraph):
-        return ImitationGraphExecutor(place)
-    if isinstance(graph, IRGraph):
-        return IRGraphExecutor(place)

python/paddle/fluid/contrib/slim/graph/graph_wrapper.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections import OrderedDict
+from .... import io
+from .... import compiler
+from ....framework import Program
+from ....framework import program_guard
+from ....framework import Parameter
+from ....framework import Variable
+from ....executor import Executor
+import copy
+from collections import Iterable
+from ....io import save_inference_model, load_inference_model, save_persistables
+import numpy as np
+import pickle
+import os
+
+__all__ = ['GraphWrapper', 'VarWrapper', 'OpWrapper']
+
+OPTIMIZER_OPS = [
+    'momentum',
+    'lars_momentum',
+    'adagrad',
+    'adam',
+    'adamax',
+    'decayed_adagrad',
+    'adadelta',
+    'rmsprop',
+]
+
+
+class VarWrapper(object):
+    def __init__(self, var, graph):
+        assert isinstance(var, Variable)
+        assert isinstance(graph, GraphWrapper)
+        self._var = var
+        self._graph = graph
+
+    def __eq__(self, v):
+        """
+        Overwrite this function for ...in... syntax in python.
+        """
+        return self._var.name == v._var.name
+
+    def name(self):
+        """
+        Get the name of the variable.
+        """
+        return self._var.name
+
+    def shape(self):
+        """
+        Get the shape of the varibale.
+        """
+        return self._var.shape
+
+    def set_shape(self, shape):
+        """
+        Set the shape of the variable.
+        """
+        self._var.desc.set_shape(shape)
+
+    def inputs(self):
+        """
+        Get all the operators that use this variable as output.
+        Returns:
+            list<OpWrapper>: A list of operators.
+        """
+        ops = []
+        for op in self._graph.ops():
+            if self in op.all_inputs():
+                ops.append(op)
+        return ops
+
+    def outputs(self):
+        """
+        Get all the operators that use this variable as input.
+        Returns:
+            list<OpWrapper>: A list of operators.
+        """
+        ops = []
+        for op in self._graph.ops():
+            if self in op.all_outputs():
+                ops.append(op)
+        return ops
+
+
+class OpWrapper(object):
+    def __init__(self, op, graph):
+        assert isinstance(graph, GraphWrapper)
+        self._op = op
+        self._graph = graph
+
+    def __eq__(self, op):
+        """
+        Overwrite this function for ...in... syntax in python.
+        """
+        return self.idx() == op.idx()
+
+    def all_inputs(self):
+        """
+        Get all the input variables of this operator.
+        """
+        return [
+            self._graph.var(var_name) for var_name in self._op.input_arg_names
+        ]
+
+    def all_outputs(self):
+        """
+        Get all the output variables of this operator.
+        """
+        return [
+            self._graph.var(var_name) for var_name in self._op.output_arg_names
+        ]
+
+    def idx(self):
+        """
+        Get the id of this operator.
+        """
+        return self._op.idx
+
+    def type(self):
+        """
+        Get the type of this operator.
+        """
+        return self._op.type
+
+    def is_bwd_op(self):
+        """
+        Whether this operator is backward op.
+        """
+        return self.type().endswith('_grad')
+
+    def is_opt_op(self):
+        """
+        Whether this operator is optimizer op.
+        """
+        return self.type() in OPTIMIZER_OPS
+
+    def inputs(self, name):
+        """
+        Get all the varibales by the input name.
+        """
+        return [self._graph.var(var_name) for var_name in self._op.input(name)]
+
+    def outputs(self, name):
+        """
+        Get all the varibales by the output name.
+        """
+        return [self._graph.var(var_name) for var_name in self._op.output(name)]
+
+    def set_attr(self, key, value):
+        """
+        Set the value of attribute by attribute's name.
+
+        Args:
+            key(str): the attribute name.
+            value(bool|int|str|float|list): the value of the attribute.
+        """
+        self._op._set_attr(key, value)
+
+    def attr(self, name):
+        """
+        Get the attribute by name.
+
+        Args:
+            name(str): the attribute name.
+
+        Returns:
+            bool|int|str|float|list: The attribute value. The return value
+            can be any valid attribute type.
+        """
+        return self._op.attr(name)
+
+
+class GraphWrapper(object):
+    """
+    It is a wrapper of paddle.fluid.framework.IrGraph with some special functions
+    for paddle slim framework.
+    """
+
+    def __init__(self, program=None, in_nodes=[], out_nodes=[]):
+        """
+        Args:
+            program(framework.Program): A program with 
+            in_nodes(dict): A dict to indicate the input nodes of the graph.
+                            The key is user-defined and human-readable name.
+                            The value is the name of Variable.
+            out_nodes(dict): A dict to indicate the input nodes of the graph.
+                            The key is user-defined and human-readable name.
+                            The value is the name of Variable.
+        """
+        super(GraphWrapper, self).__init__()
+        self.program = Program() if program is None else program
+        self.compiled_graph = None
+        self.in_nodes = OrderedDict(in_nodes)
+        self.out_nodes = OrderedDict(out_nodes)
+        self._attrs = OrderedDict()
+
+    def all_parameters(self):
+        """
+        Get all the parameters in this graph.
+        Returns:
+            list<VarWrapper>: A list of VarWrapper instances.
+        """
+        params = []
+        for block in self.program.blocks:
+            for param in block.all_parameters():
+                params.append(VarWrapper(param, self))
+        return params
+
+    def is_parameter(self, var):
+        """
+        Whether the given variable is parameter.
+        Args:
+            var(VarWrapper): The given varibale.
+        """
+        return isinstance(var._var, Parameter)
+
+    def is_persistable(self, var):
+        """
+        Whether the given variable is persistable.
+        Args:
+            var(VarWrapper): The given varibale.
+        """
+        return var._var.persistable
+
+    def compile(self, for_parallel=True, for_test=False):
+        """
+        Compile the program in this wrapper to framework.CompiledProgram for next running.
+        This function must be called if the program is modified.
+        Args:
+            for_parallel(bool): Whether the program to run in data parallel way. default: True.
+            for_test(bool): Whether the compiled program is used for test.
+        """
+        target = self.program
+        if for_test:
+            loss = None
+        else:
+            loss = self.out_nodes['loss']
+        if for_parallel:
+            # disable memory optimize for stable training
+            build_strategy = compiler.BuildStrategy()
+            build_strategy.enable_inplace = False
+            build_strategy.memory_optimize = False
+            self.compiled_graph = compiler.CompiledProgram(
+                target).with_data_parallel(
+                    loss_name=loss, build_strategy=build_strategy)
+        else:
+            self.compiled_graph = compiler.CompiledProgram(target)
+
+    def ops(self):
+        """
+        Return all operator nodes included in the graph as a set.
+        """
+        ops = []
+        for block in self.program.blocks:
+            for op in block.ops:
+                ops.append(OpWrapper(op, self))
+        return ops
+
+    def vars(self):
+        """
+        Get all the variables.
+        """
+        return [VarWrapper(var, self) for var in self.program.list_vars()]
+
+    def var(self, name):
+        """
+        Get the variable by variable name.
+        """
+        return VarWrapper(self.program.global_block().var(name), self)
+
+    def clone(self, for_test=False):
+        """
+        Clone a new graph from current graph.
+        Returns:
+            (GraphWrapper): The wrapper of a new graph.
+        """
+        return GraphWrapper(
+            self.program.clone(for_test),
+            copy.deepcopy(self.in_nodes), copy.deepcopy(self.out_nodes))
+
+    def merge(self, graph):
+        """
+        Merge a graph into current graph.
+        Args:
+            graph(GraphWrapper): The graph to be merged by current graph.
+        """
+        for var in graph.program.list_vars():
+            new_var = self.program.global_block()._clone_variable(
+                var, force_persistable=False)
+            new_var.stop_gradient = var.stop_gradient
+            # TODO: parameters should be cloned
+        for op in graph.ops():
+            op = op._op
+            inputs = {}
+            outputs = {}
+            attrs = {}
+            for input_name in op.input_names:
+                inputs[input_name] = [
+                    self.var(in_var_name)._var
+                    for in_var_name in op.input(input_name)
+                ]
+            for output_name in op.output_names:
+                outputs[output_name] = [
+                    self.var(out_var_name)._var
+                    for out_var_name in op.output(output_name)
+                ]
+            for attr_name in op.attr_names:
+                attrs[attr_name] = op.attr(attr_name)
+            self.program.global_block().append_op(
+                type=op.type, inputs=inputs, outputs=outputs, attrs=attrs)
+
+    def program(self):
+        """
+        Get the program in current wrapper.
+        """
+        return self.program
+
+    def pre_ops(self, op):
+        """
+        Get all the previous operators of target operator.
+        Args:
+            op(OpWrapper): Target operator..
+        Returns:
+            list<OpWrapper>: A list of operators.
+        """
+        ops = []
+        for p in self.ops():
+            for in_var in op.all_inputs():
+                if in_var in p.all_outputs():
+                    ops.append(p)
+        return ops
+
+    def next_ops(self, op):
+        """
+        Get all the next operators of target operator.
+        Args:
+            op(OpWrapper): Target operator..
+        Returns:
+            list<OpWrapper>: A list of operators.
+        """
+        ops = []
+        for p in self.ops():
+            for out_var in op.all_outputs():
+                if out_var in p.all_inputs():
+                    ops.append(p)
+        return ops
+
+    def get_param_by_op(self, op):
+        """
+        Get the parameters used by target operator.
+        """
+        assert isinstance(op, OpWrapper)
+        params = []
+        for var in op.all_inputs():
+            if isinstance(var._var, Parameter):
+                params.append(var)
+        assert len(params) > 0
+        return params
+
+    def numel_params(self):
+        """
+        Get the number of elements in all parameters.
+        """
+        ret = 0
+        for param in self.all_parameters():
+            ret += np.product(param.shape())
+        return ret
+
+    def get_optimize_graph(self, optimizer, place, scope, no_grad_var_names=[]):
+        """
+        Get a new graph for training by appending some backward operators and optimization operators.
+        Args:
+            optimizer: The optimzier used to generate training graph.
+            place: The place to run the graph.
+            scope: The scope used to run the graph. Some new variable will be added into this scope.
+            no_grad_var_names(list<str>): Names of variables that should be ignored while computing gradients. default: [].
+        Returns:
+            (GraphWrapper): The wrapper of new graph with backward ops and optimization ops. 
+        """
+        graph = self.clone()
+        startup_program = Program()
+        with program_guard(
+                main_program=graph.program, startup_program=startup_program):
+            target_name = None
+            if 'loss' in graph.out_nodes:
+                target_name = graph.out_nodes['loss']
+            elif 'cost' in graph.out_nodes:
+                target_name = graph.out_nodes['cost']
+            target = graph.var(target_name)._var
+            optimizer.minimize(target, no_grad_set=no_grad_var_names)
+
+        exe = Executor(place)
+        exe.run(program=startup_program, scope=scope)
+        return graph
+
+    def flops(self, only_conv=False):
+        """
+        Get the flops of current graph.
+        Args:
+            only_conv: Only calculating the conv layers. default: False.
+        Returns:
+            int: The flops of current graph.
+        """
+        flops = 0
+        for op in self.ops():
+            if op.type() in ['conv2d', 'depthwise_conv2d']:
+                filter_shape = op.inputs("Filter")[0].shape()
+                input_shape = op.inputs("Input")[0].shape()
+                output_shape = op.outputs("Output")[0].shape()
+                c_out, c_in, k_h, k_w = filter_shape
+                _, _, h_out, w_out = output_shape
+                groups = op.attr("groups")
+                kernel_ops = k_h * k_w * (c_in / groups)
+                if len(op.inputs("Bias")) > 0:
+                    with_bias = 1
+                else:
+                    with_bias = 0
+                flops += 2 * h_out * w_out * c_out * (kernel_ops + with_bias)
+            elif op.type() == 'pool2d' and not only_conv:
+                input_shape = op.inputs("X")[0].shape()
+                output_shape = op.outputs("Out")[0].shape()
+                _, c_out, h_out, w_out = output_shape
+                k_size = op.attr("ksize")
+                flops += h_out * w_out * c_out * (k_size[0]**2)
+
+            elif op.type() == 'mul' and not only_conv:
+                x_shape = list(op.inputs("X")[0].shape())
+                y_shape = op.inputs("Y")[0].shape()
+                if x_shape[0] == -1:
+                    x_shape[0] = 1
+                flops += 2 * x_shape[0] * x_shape[1] * y_shape[1]
+
+            elif op.type() in ['relu', 'sigmoid', 'batch_norm'
+                               ] and not only_conv:
+                input_shape = list(op.inputs("X")[0].shape())
+                if input_shape[0] == -1:
+                    input_shape[0] = 1
+                flops += np.product(input_shape)
+
+        return flops
+
+    def save_persistables(self, path, exe):
+        """
+        Save all the persistable variables into file.
+        Args:
+            path(str): The path to save the persistables.
+            exe(framework.Executor): The executor used to save the persistables.
+        """
+        io.save_persistables(exe.exe, path, main_program=self.program)
+
+    def load_persistables(self, path, exe):
+        """
+        Load the persistable variables from file.
+        Args:
+            path(str): The path to load the persistables.
+            exe(framework.Executor): The executor used to load the persistables.
+        """
+
+        def if_exist(var):
+            return os.path.exists(os.path.join(path, var.name))
+
+        io.load_vars(
+            exe.exe, path, main_program=self.program, predicate=if_exist)
+
+    def update_param_shape(self, scope):
+        """
+        Update the shape of parameters in the graph according to tensors in scope.
+        It is used after loading pruned parameters from file.
+        """
+        for param in self.all_parameters():
+            tensor_shape = np.array(scope.find_var(param.name()).get_tensor(
+            )).shape
+            param.set_shape(tensor_shape)
+
+    def infer_shape(self):
+        """
+        Update the groups of convolution layer according to current filters.
+        It is used after loading pruned parameters from file.
+        """
+        for op in self.ops():
+            if op.type() != 'conditional_block':
+                op._op.desc.infer_shape(op._op.block.desc)
+
+    def update_groups_of_conv(self):
+        for op in self.ops():
+            if op.type() == 'depthwise_conv2d':
+                op.set_attr('groups', op.inputs('Filter')[0].shape()[0])

python/paddle/fluid/contrib/slim/prune/prune_strategy.py

@@ -13,54 +13,919 @@
 # limitations under the License.
 
 from ..core.strategy import Strategy
-from ....framework import Program, program_guard
+from ..graph import VarWrapper, OpWrapper, GraphWrapper
+from ....framework import Program, program_guard, Parameter
 from .... import layers
+import prettytable as pt
 import numpy as np
+from scipy.optimize import leastsq
+import copy
+import re
+import os
+import pickle
+import logging
+import sys
 
-__all__ = ['SensitivePruneStrategy', 'PruneStrategy']
+__all__ = ['SensitivePruneStrategy', 'UniformPruneStrategy']
 
+logging.basicConfig(format='%(asctime)s-%(levelname)s: %(message)s')
+_logger = logging.getLogger(__name__)
+_logger.setLevel(logging.INFO)
+
+
+class PruneStrategy(Strategy):
+    """
+    The base class of all pruning strategies.
+    """
 
-class SensitivePruneStrategy(Strategy):
     def __init__(self,
                  pruner=None,
                  start_epoch=0,
-                 end_epoch=10,
-                 delta_rate=0.20,
-                 acc_loss_threshold=0.2,
-                 sensitivities=None):
-        super(SensitivePruneStrategy, self).__init__(start_epoch, end_epoch)
+                 end_epoch=0,
+                 target_ratio=0.5,
+                 metric_name=None,
+                 pruned_params='conv.*_weights'):
+        """
+        Args:
+            pruner(slim.Pruner): The pruner used to prune the parameters.
+            start_epoch(int): The 'on_epoch_begin' function will be called in start_epoch. default: 0
+            end_epoch(int): The 'on_epoch_end' function will be called in end_epoch. default: 0
+            target_ratio(float): The flops ratio to be pruned from current model.
+            metric_name(str): The metric used to evaluate the model.
+                         It should be one of keys in out_nodes of graph wrapper.
+            pruned_params(str): The pattern str to match the parameter names to be pruned.
+        """
+        super(PruneStrategy, self).__init__(start_epoch, end_epoch)
         self.pruner = pruner
-        self.delta_rate = delta_rate
-        self.acc_loss_threshold = acc_loss_threshold
-        self.sensitivities = sensitivities
+        self.target_ratio = target_ratio
+        self.metric_name = metric_name
+        self.pruned_params = pruned_params
+        self.pruned_list = []
+        self.backup = {}
+        self.param_shape_backup = {}
 
+    def _eval_graph(self, context, sampled_rate=None, cached_id=0):
+        """
+        Evaluate the current mode in context.
+        Args:
+            context(slim.core.Context): The context storing all information used to evaluate the current model.
+            sampled_rate(float): The sampled rate used to sample partial data for evaluation. None means using all data in eval_reader. default: None.
+            cached_id(int): The id of dataset sampled. Evaluations with same cached_id use the same sampled dataset. default: 0.
+        """
+        results, names = context.run_eval_graph(sampled_rate, cached_id)
+        metric = np.mean(results[list(names).index(self.metric_name)])
+        return metric
 
-class PruneStrategy(Strategy):
+    def _prune_filters_by_ratio(self,
+                                scope,
+                                params,
+                                ratio,
+                                place,
+                                lazy=False,
+                                only_graph=False):
+        """
+        Pruning filters by given ratio.
+        Args:
+            scope(fluid.core.Scope): The scope used to pruning filters.
+            params(list<VarWrapper>): A list of filter parameters.
+            ratio(float): The ratio to be pruned.
+            place(fluid.Place): The device place of filter parameters.
+            lazy(bool): True means setting the pruned elements to zero.
+                        False means cutting down the pruned elements.
+            only_graph(bool): True means only modifying the graph.
+                              False means modifying graph and variables in  scope.
+        """
+        if params[0].name() in self.pruned_list[0]:
+            return
+        param_t = scope.find_var(params[0].name()).get_tensor()
+        pruned_idx = self.pruner.cal_pruned_idx(
+            params[0].name(), np.array(param_t), ratio, axis=0)
+        for param in params:
+            assert isinstance(param, VarWrapper)
+            param_t = scope.find_var(param.name()).get_tensor()
+            if lazy:
+                self.backup[param.name()] = copy.deepcopy(np.array(param_t))
+            pruned_param = self.pruner.prune_tensor(
+                np.array(param_t), pruned_idx, pruned_axis=0, lazy=lazy)
+            if not only_graph:
+                param_t.set(pruned_param, place)
+            ori_shape = param.shape()
+            if param.name() not in self.param_shape_backup:
+                self.param_shape_backup[param.name()] = copy.deepcopy(
+                    param.shape())
+            new_shape = list(param.shape())
+            new_shape[0] = pruned_param.shape[0]
+            param.set_shape(new_shape)
+            _logger.debug(
+                '|----------------------------------------+----+------------------------------+------------------------------|'
+            )
+            _logger.debug('|{:^40}|{:^4}|{:^30}|{:^30}|'.format(
+                str(param.name()), str(0), str(ori_shape), str(param.shape())))
+            self.pruned_list[0].append(param.name())
+        return pruned_idx
+
+    def _prune_parameter_by_idx(self,
+                                scope,
+                                params,
+                                pruned_idx,
+                                pruned_axis,
+                                place,
+                                lazy=False,
+                                only_graph=False):
+        """
+        Pruning parameters in given axis.
+        Args:
+            scope(fluid.core.Scope): The scope storing paramaters to be pruned.
+            params(VarWrapper): The parameter to be pruned.
+            pruned_idx(list): The index of elements to be pruned.
+            pruned_axis(int): The pruning axis.
+            place(fluid.Place): The device place of filter parameters.
+            lazy(bool): True means setting the pruned elements to zero.
+                        False means cutting down the pruned elements.
+            only_graph(bool): True means only modifying the graph.
+                              False means modifying graph and variables in  scope.
+        """
+        if params[0].name() in self.pruned_list[pruned_axis]:
+            return
+        for param in params:
+            assert isinstance(param, VarWrapper)
+            param_t = scope.find_var(param.name()).get_tensor()
+            if lazy:
+                self.backup[param.name()] = copy.deepcopy(np.array(param_t))
+            pruned_param = self.pruner.prune_tensor(
+                np.array(param_t), pruned_idx, pruned_axis, lazy=lazy)
+            if not only_graph:
+                param_t.set(pruned_param, place)
+            ori_shape = param.shape()
+            if param.name() not in self.param_shape_backup:
+                self.param_shape_backup[param.name()] = copy.deepcopy(
+                    param.shape())
+            new_shape = list(param.shape())
+            new_shape[pruned_axis] = pruned_param.shape[pruned_axis]
+            param.set_shape(new_shape)
+            _logger.debug(
+                '|----------------------------------------+----+------------------------------+------------------------------|'
+            )
+            _logger.debug('|{:^40}|{:^4}|{:^30}|{:^30}|'.format(
+                str(param.name()),
+                str(pruned_axis), str(ori_shape), str(param.shape())))
+            self.pruned_list[pruned_axis].append(param.name())
+
+    def _forward_search_related_op(self, graph, param):
+        """
+        Forward search operators that will be affected by pruning of param.
+        Args:
+            graph(GraphWrapper): The graph to be searched.
+            param(VarWrapper): The current pruned parameter.
+        Returns:
+            list<OpWrapper>: A list of operators.
+        """
+        assert isinstance(param, VarWrapper)
+        visited = {}
+        for op in graph.ops():
+            visited[op.idx()] = False
+        stack = []
+        for op in graph.ops():
+            if (not op.is_bwd_op()) and (param in op.all_inputs()):
+                stack.append(op)
+        visit_path = []
+        while len(stack) > 0:
+            top_op = stack[len(stack) - 1]
+            if visited[top_op.idx()] == False:
+                visit_path.append(top_op)
+                visited[top_op.idx()] = True
+            next_ops = None
+            if top_op.type() == "conv2d" and param not in top_op.all_inputs():
+                next_ops = None
+            elif top_op.type() == "mul":
+                next_ops = None
+            else:
+                next_ops = self._get_next_unvisited_op(graph, visited, top_op)
+            if next_ops == None:
+                stack.pop()
+            else:
+                stack += next_ops
+        return visit_path
+
+    def _get_next_unvisited_op(self, graph, visited, top_op):
+        """
+        Get next unvisited adjacent operators of given operators.
+        Args:
+            graph(GraphWrapper): The graph used to search. 
+            visited(list): The ids of operators that has been visited.
+            top_op: The given operator.
+        Returns:
+            list<OpWrapper>: A list of operators. 
+        """
+        assert isinstance(top_op, OpWrapper)
+        next_ops = []
+        for op in graph.next_ops(top_op):
+            if (visited[op.idx()] == False) and (not op.is_bwd_op()):
+                next_ops.append(op)
+        return next_ops if len(next_ops) > 0 else None
+
+    def _get_accumulator(self, graph, param):
+        """
+        Get accumulators of given parameter. The accumulator was created by optimizer.
+        Args:
+            graph(GraphWrapper): The graph used to search.
+            param(VarWrapper): The given parameter.
+        Returns:
+            list<VarWrapper>: A list of accumulators which are variables.
+        """
+        assert isinstance(param, VarWrapper)
+        params = []
+        for op in param.outputs():
+            if op.is_opt_op():
+                for out_var in op.all_outputs():
+                    if graph.is_persistable(out_var) and out_var.name(
+                    ) != param.name():
+                        params.append(out_var)
+        return params
+
+    def _forward_pruning_ralated_params(self,
+                                        graph,
+                                        scope,
+                                        param,
+                                        place,
+                                        ratio=None,
+                                        pruned_idxs=None,
+                                        lazy=False,
+                                        only_graph=False):
+        """
+        Pruning all the parameters affected by the pruning of given parameter.
+        Args:
+            graph(GraphWrapper): The graph to be searched.
+            scope(fluid.core.Scope): The scope storing paramaters to be pruned.
+            param(VarWrapper): The given parameter.
+            place(fluid.Place): The device place of filter parameters.
+            ratio(float): The target ratio to be pruned.
+            pruned_idx(list): The index of elements to be pruned.
+            lazy(bool): True means setting the pruned elements to zero.
+                        False means cutting down the pruned elements.
+            only_graph(bool): True means only modifying the graph.
+                              False means modifying graph and variables in  scope.
+        """
+        assert isinstance(
+            graph,
+            GraphWrapper), "graph must be instance of slim.core.GraphWrapper"
+        assert isinstance(
+            param, VarWrapper), "param must be instance of slim.core.VarWrapper"
+
+        if param.name() in self.pruned_list[0]:
+            return
+        related_ops = self._forward_search_related_op(graph, param)
+
+        if ratio is None:
+            assert pruned_idxs is not None
+            self._prune_parameter_by_idx(
+                scope, [param] + self._get_accumulator(graph, param),
+                pruned_idxs,
+                pruned_axis=0,
+                place=place,
+                lazy=lazy,
+                only_graph=only_graph)
+
+        else:
+            pruned_idxs = self._prune_filters_by_ratio(
+                scope, [param] + self._get_accumulator(graph, param),
+                ratio,
+                place,
+                lazy=lazy,
+                only_graph=only_graph)
+        corrected_idxs = pruned_idxs[:]
+
+        for idx, op in enumerate(related_ops):
+            if op.type() == "conv2d" and (param not in op.all_inputs()):
+                for in_var in op.all_inputs():
+                    if graph.is_parameter(in_var):
+                        conv_param = in_var
+                        self._prune_parameter_by_idx(
+                            scope, [conv_param] + self._get_accumulator(
+                                graph, conv_param),
+                            corrected_idxs,
+                            pruned_axis=1,
+                            place=place,
+                            lazy=lazy,
+                            only_graph=only_graph)
+            if op.type() == "depthwise_conv2d":
+                for in_var in op.all_inputs():
+                    if graph.is_parameter(in_var):
+                        conv_param = in_var
+                        self._prune_parameter_by_idx(
+                            scope, [conv_param] + self._get_accumulator(
+                                graph, conv_param),
+                            corrected_idxs,
+                            pruned_axis=0,
+                            place=place,
+                            lazy=lazy,
+                            only_graph=only_graph)
+            elif op.type() == "elementwise_add":
+                # pruning bias
+                for in_var in op.all_inputs():
+                    if graph.is_parameter(in_var):
+                        bias_param = in_var
+                        self._prune_parameter_by_idx(
+                            scope, [bias_param] + self._get_accumulator(
+                                graph, bias_param),
+                            pruned_idxs,
+                            pruned_axis=0,
+                            place=place,
+                            lazy=lazy,
+                            only_graph=only_graph)
+            elif op.type() == "mul":  # pruning fc layer
+                fc_input = None
+                fc_param = None
+                for in_var in op.all_inputs():
+                    if graph.is_parameter(in_var):
+                        fc_param = in_var
+                    else:
+                        fc_input = in_var
+
+                idx = []
+                feature_map_size = fc_input.shape()[2] * fc_input.shape()[3]
+                range_idx = np.array(range(feature_map_size))
+                for i in corrected_idxs:
+                    idx += list(range_idx + i * feature_map_size)
+                corrected_idxs = idx
+                self._prune_parameter_by_idx(
+                    scope, [fc_param] + self._get_accumulator(graph, fc_param),
+                    corrected_idxs,
+                    pruned_axis=0,
+                    place=place,
+                    lazy=lazy,
+                    only_graph=only_graph)
+
+            elif op.type() == "concat":
+                concat_inputs = op.all_inputs()
+                last_op = related_ops[idx - 1]
+                for out_var in last_op.all_outputs():
+                    if out_var in concat_inputs:
+                        concat_idx = concat_inputs.index(out_var)
+                offset = 0
+                for ci in range(concat_idx):
+                    offset += concat_inputs[ci].shape()[1]
+                corrected_idxs = [x + offset for x in pruned_idxs]
+            elif op.type() == "batch_norm":
+                bn_inputs = op.all_inputs()
+                mean = bn_inputs[2]
+                variance = bn_inputs[3]
+                alpha = bn_inputs[0]
+                beta = bn_inputs[1]
+                self._prune_parameter_by_idx(
+                    scope, [mean] + self._get_accumulator(graph, mean),
+                    corrected_idxs,
+                    pruned_axis=0,
+                    place=place,
+                    lazy=lazy,
+                    only_graph=only_graph)
+                self._prune_parameter_by_idx(
+                    scope, [variance] + self._get_accumulator(graph, variance),
+                    corrected_idxs,
+                    pruned_axis=0,
+                    place=place,
+                    lazy=lazy,
+                    only_graph=only_graph)
+                self._prune_parameter_by_idx(
+                    scope, [alpha] + self._get_accumulator(graph, alpha),
+                    corrected_idxs,
+                    pruned_axis=0,
+                    place=place,
+                    lazy=lazy,
+                    only_graph=only_graph)
+                self._prune_parameter_by_idx(
+                    scope, [beta] + self._get_accumulator(graph, beta),
+                    corrected_idxs,
+                    pruned_axis=0,
+                    place=place,
+                    lazy=lazy,
+                    only_graph=only_graph)
+
+    def _prune_parameters(self,
+                          graph,
+                          scope,
+                          params,
+                          ratios,
+                          place,
+                          lazy=False,
+                          only_graph=False):
+        """
+        Pruning the given parameters.
+        Args:
+            graph(GraphWrapper): The graph to be searched.
+            scope(fluid.core.Scope): The scope storing paramaters to be pruned.
+            params(list<str>): A list of parameter names to be pruned.
+            ratios(list<float>): A list of ratios to be used to pruning parameters.
+            place(fluid.Place): The device place of filter parameters.
+            pruned_idx(list): The index of elements to be pruned.
+            lazy(bool): True means setting the pruned elements to zero.
+                        False means cutting down the pruned elements.
+            only_graph(bool): True means only modifying the graph.
+                              False means modifying graph and variables in  scope.
+
+        """
+        _logger.debug('\n################################')
+        _logger.debug('#       pruning parameters       #')
+        _logger.debug('################################\n')
+        _logger.debug(
+            '|----------------------------------------+----+------------------------------+------------------------------|'
+        )
+        _logger.debug('|{:^40}|{:^4}|{:^30}|{:^30}|'.format('parameter', 'axis',
+                                                            'from', 'to'))
+        assert len(params) == len(ratios)
+        self.pruned_list = [[], []]
+        for param, ratio in zip(params, ratios):
+            assert isinstance(param, str) or isinstance(param, unicode)
+            param = graph.var(param)
+            self._forward_pruning_ralated_params(
+                graph,
+                scope,
+                param,
+                place,
+                ratio=ratio,
+                lazy=lazy,
+                only_graph=only_graph)
+            ops = param.outputs()
+            for op in ops:
+                if op.type() == 'conv2d':
+                    brother_ops = self._search_brother_ops(graph, op)
+                    for broher in brother_ops:
+                        for p in graph.get_param_by_op(broher):
+                            self._forward_pruning_ralated_params(
+                                graph,
+                                scope,
+                                p,
+                                place,
+                                ratio=ratio,
+                                lazy=lazy,
+                                only_graph=only_graph)
+        _logger.debug(
+            '|----------------------------------------+----+------------------------------+------------------------------|'
+        )
+
+    def _search_brother_ops(self, graph, op_node):
+        """
+        Search brother operators that was affected by pruning of given operator.
+        Args:
+            graph(GraphWrapper): The graph to be searched.
+            op_node(OpWrapper): The start node for searching.
+        Returns: 
+            list<VarWrapper>: A list of operators.
+        """
+        visited = [op_node.idx()]
+        stack = []
+        brothers = []
+        for op in graph.next_ops(op_node):
+            if (op.type() != 'conv2d') and (op.type() != 'fc') and (
+                    not op._is_bwd_op()):
+                stack.append(op)
+                visited.append(op.idx())
+        while len(stack) > 0:
+            top_op = stack.pop()
+            for parent in graph.pre_ops(top_op):
+                if parent.idx() not in visited and (not parent._is_bwd_op()):
+                    if ((parent.type == 'conv2d') or (parent.type == 'fc')):
+                        brothers.append(parent)
+                    else:
+                        stack.append(parent)
+                    visited.append(parent.idx())
+
+            for child in graph.next_ops(top_op):
+                if (child.type != 'conv2d') and (child.type != 'fc') and (
+                        child.idx() not in visited) and (
+                            not child._is_bwd_op()):
+                    stack.append(child)
+                    visited.append(child.idx())
+        return brothers
+
+    def _prune_graph(self, graph, target_graph):
+        """
+        Pruning parameters of graph according to target graph.
+        Args:
+            graph(GraphWrapper): The graph to be pruned.
+            target_graph(GraphWrapper): The reference graph.
+        Return: None
+        """
+        count = 1
+        _logger.debug(
+            '|----+----------------------------------------+------------------------------+------------------------------|'
+        )
+        _logger.debug('|{:^4}|{:^40}|{:^30}|{:^30}|'.format('id', 'parammeter',
+                                                            'from', 'to'))
+        for param in target_graph.all_parameters():
+            var = graph.var(param.name())
+            ori_shape = var.shape()
+            var.set_shape(param.shape())
+            _logger.debug(
+                '|----+----------------------------------------+------------------------------+------------------------------|'
+            )
+            _logger.debug('|{:^4}|{:^40}|{:^30}|{:^30}|'.format(
+                str(count),
+                str(param.name()), str(ori_shape), str(param.shape())))
+            count += 1
+        _logger.debug(
+            '|----+----------------------------------------+------------------------------+------------------------------|'
+        )
+
+
+class UniformPruneStrategy(PruneStrategy):
     """
-    The strategy that pruning weights by threshold or ratio iteratively.
+    The uniform pruning strategy. The parameters will be pruned by uniform ratio.
     """
 
     def __init__(self,
-                 pruner,
-                 mini_batch_pruning_frequency=1,
+                 pruner=None,
                  start_epoch=0,
-                 end_epoch=10):
-        super(PruneStrategy, self).__init__(start_epoch, end_epoch)
-        self.pruner = pruner
-        self.mini_batch_pruning_frequency = mini_batch_pruning_frequency
-
-    def _triger(self, context):
-        return (context.batch_id % self.mini_batch_pruning_frequency == 0 and
-                self.start_epoch <= context.epoch_id < self.end_epoch)
-
-    def on_batch_end(self, context):
-        if self._triger(context):
-            prune_program = Program()
-            with program_guard(prune_program):
-                for param in context.graph.all_parameters():
-                    prune_program.global_block().clone_variable(param)
-                    p = prune_program.global_block().var(param.name)
-                    zeros_mask = self.pruner.prune(p)
-                    pruned_param = p * zeros_mask
-                    layers.assign(input=pruned_param, output=param)
-            context.program_exe.run(prune_program, scope=context.scope)
+                 end_epoch=0,
+                 target_ratio=0.5,
+                 metric_name=None,
+                 pruned_params='conv.*_weights'):
+        """
+        Args:
+            pruner(slim.Pruner): The pruner used to prune the parameters.
+            start_epoch(int): The 'on_epoch_begin' function will be called in start_epoch. default: 0
+            end_epoch(int): The 'on_epoch_end' function will be called in end_epoch. default: 0
+            target_ratio(float): The flops ratio to be pruned from current model.
+            metric_name(str): The metric used to evaluate the model.
+                         It should be one of keys in out_nodes of graph wrapper.
+            pruned_params(str): The pattern str to match the parameter names to be pruned.
+        """
+        super(UniformPruneStrategy, self).__init__(pruner, start_epoch,
+                                                   end_epoch, target_ratio,
+                                                   metric_name, pruned_params)
+
+    def _get_best_ratios(self, context):
+        """
+        Search a group of ratios for pruning target flops.
+        """
+        _logger.info('_get_best_ratios')
+        pruned_params = []
+        for param in context.eval_graph.all_parameters():
+            if re.match(self.pruned_params, param.name()):
+                pruned_params.append(param.name())
+
+        min_ratio = 0.
+        max_ratio = 1.
+
+        flops = context.eval_graph.flops()
+        model_size = context.eval_graph.numel_params()
+
+        while min_ratio < max_ratio:
+            ratio = (max_ratio + min_ratio) / 2
+            _logger.debug(
+                '-----------Try pruning ratio: {:.2f}-----------'.format(ratio))
+            ratios = [ratio] * len(pruned_params)
+            self._prune_parameters(
+                context.eval_graph,
+                context.scope,
+                pruned_params,
+                ratios,
+                context.place,
+                only_graph=True)
+
+            pruned_flops = 1 - (float(context.eval_graph.flops()) / flops)
+            pruned_size = 1 - (float(context.eval_graph.numel_params()) /
+                               model_size)
+            _logger.debug('Pruned flops: {:.2f}'.format(pruned_flops))
+            _logger.debug('Pruned model size: {:.2f}'.format(pruned_size))
+            for param in self.param_shape_backup.keys():
+                context.eval_graph.var(param).set_shape(self.param_shape_backup[
+                    param])
+            self.param_shape_backup = {}
+
+            if abs(pruned_flops - self.target_ratio) < 1e-2:
+                break
+            if pruned_flops > self.target_ratio:
+                max_ratio = ratio
+            else:
+                min_ratio = ratio
+        _logger.info('Get ratios: {}'.format([round(r, 2) for r in ratios]))
+        return pruned_params, ratios
+
+    def on_epoch_begin(self, context):
+        if context.epoch_id == self.start_epoch:
+            params, ratios = self._get_best_ratios(context)
+
+            self._prune_parameters(context.optimize_graph, context.scope,
+                                   params, ratios, context.place)
+
+            model_size = context.eval_graph.numel_params()
+            flops = context.eval_graph.flops()
+            _logger.debug('\n################################')
+            _logger.debug('#          pruning eval graph    #')
+            _logger.debug('################################\n')
+            self._prune_graph(context.eval_graph, context.optimize_graph)
+            context.optimize_graph.update_groups_of_conv()
+            context.eval_graph.update_groups_of_conv()
+
+            _logger.info(
+                '------------------finish pruning--------------------------------'
+            )
+            _logger.info('Pruned size: {:.2f}'.format(1 - (float(
+                context.eval_graph.numel_params()) / model_size)))
+            _logger.info('Pruned flops: {:.2f}'.format(1 - (float(
+                context.eval_graph.flops()) / flops)))
+            #            metric = self._eval_graph(context)
+            #            _logger.info('Metric after pruning: {:.2f}'.format(metric))
+            _logger.info(
+                '------------------UniformPruneStrategy.on_compression_begin finish--------------------------------'
+            )
+
+
+class SensitivePruneStrategy(PruneStrategy):
+    """
+    Sensitive pruning strategy. Different pruned ratio was applied on each layer.
+    """
+
+    def __init__(self,
+                 pruner=None,
+                 start_epoch=0,
+                 end_epoch=0,
+                 delta_rate=0.20,
+                 target_ratio=0.5,
+                 metric_name='top1_acc',
+                 pruned_params='conv.*_weights',
+                 sensitivities_file='./sensitivities.data',
+                 sensitivities={},
+                 num_steps=1,
+                 eval_rate=None):
+        """
+        Args:
+            pruner(slim.Pruner): The pruner used to prune the parameters.
+            start_epoch(int): The 'on_epoch_begin' function will be called in start_epoch. default: 0.
+            end_epoch(int): The 'on_epoch_end' function will be called in end_epoch. default: 10.
+            delta_rate(float): The delta used to generate ratios when calculating sensitivities. default: 0.2
+            target_ratio(float): The flops ratio to be pruned from current model. default: 0.5
+            metric_name(str): The metric used to evaluate the model.
+                         It should be one of keys in out_nodes of graph wrapper. default: 'top1_acc'
+            pruned_params(str): The pattern str to match the parameter names to be pruned. default: 'conv.*_weights'.
+            sensitivities_file(str): The sensitivities file. default: './sensitivities.data'
+            sensitivities(dict): The user-defined sensitivities. default: {}.
+            num_steps(int): The number of pruning steps. default: 1.
+            eval_rate(float): The rate of sampled data used to calculate sensitivities.
+                              None means using all the data. default: None.
+        """
+        super(SensitivePruneStrategy, self).__init__(pruner, start_epoch,
+                                                     end_epoch, target_ratio,
+                                                     metric_name, pruned_params)
+        self.delta_rate = delta_rate
+        self.pruned_list = []
+        self.sensitivities = sensitivities
+        self.sensitivities_file = sensitivities_file
+        self.backup = {}
+        self.param_shape_backup = {}
+        self.num_steps = num_steps
+        self.eval_rate = eval_rate
+        self.pruning_step = 1 - pow((1 - target_ratio), 1.0 / self.num_steps)
+
+    def _save_sensitivities(self, sensitivities, sensitivities_file):
+        """
+        Save sensitivities into file.
+        """
+        with open(sensitivities_file, 'wb') as f:
+            pickle.dump(sensitivities, f)
+
+    def _load_sensitivities(self, sensitivities_file):
+        """
+        Load sensitivities from file.
+        """
+        sensitivities = {}
+        if sensitivities_file and os.path.exists(sensitivities_file):
+            with open(sensitivities_file, 'rb') as f:
+                if sys.version_info < (3, 0):
+                    sensitivities = pickle.load(f)
+                else:
+                    sensitivities = pickle.load(f, encoding='bytes')
+
+        for param in sensitivities:
+            sensitivities[param]['pruned_percent'] = [
+                round(p, 2) for p in sensitivities[param]['pruned_percent']
+            ]
+        self._format_sensitivities(sensitivities)
+        return sensitivities
+
+    def _format_sensitivities(self, sensitivities):
+        """
+        Print formated sensitivities in debug log level.
+        """
+        tb = pt.PrettyTable()
+        tb.field_names = ["parameter", "size"] + [
+            str(round(i, 2))
+            for i in np.arange(self.delta_rate, 1, self.delta_rate)
+        ]
+        for param in sensitivities:
+            if len(sensitivities[param]['loss']) == (len(tb.field_names) - 2):
+                tb.add_row([param, sensitivities[param]['size']] + [
+                    round(loss, 2) for loss in sensitivities[param]['loss']
+                ])
+        _logger.debug('\n################################')
+        _logger.debug('#      sensitivities table     #')
+        _logger.debug('################################\n')
+        _logger.debug(tb)
+
+    def _compute_sensitivities(self, context):
+        """
+        Computing the sensitivities of all parameters.
+        """
+        _logger.info("calling _compute_sensitivities.")
+        self.param_shape_backup = {}
+        self.backup = {}
+        cached_id = np.random.randint(1000)
+        if self.start_epoch == context.epoch_id:
+            sensitivities_file = self.sensitivities_file
+        else:
+            sensitivities_file = self.sensitivities_file + ".epoch" + str(
+                context.epoch_id)
+        sensitivities = self._load_sensitivities(sensitivities_file)
+
+        for param in context.eval_graph.all_parameters():
+            if not re.match(self.pruned_params, param.name()):
+                continue
+            if param.name() not in sensitivities:
+                sensitivities[param.name()] = {
+                    'pruned_percent': [],
+                    'loss': [],
+                    'size': param.shape()[0]
+                }
+
+        metric = None
+
+        for param in sensitivities.keys():
+            ratio = self.delta_rate
+            while ratio < 1:
+                ratio = round(ratio, 2)
+                if ratio in sensitivities[param]['pruned_percent']:
+                    _logger.debug('{}, {} has computed.'.format(param, ratio))
+                    ratio += self.delta_rate
+                    continue
+                if metric is None:
+                    metric = self._eval_graph(context, self.eval_rate,
+                                              cached_id)
+                # prune parameter by ratio
+                self._prune_parameters(
+                    context.eval_graph,
+                    context.scope, [param], [ratio],
+                    context.place,
+                    lazy=True)
+                self.pruned_list[0]
+                # get accuracy after pruning and update self.sensitivities
+                pruned_metric = self._eval_graph(context, self.eval_rate,
+                                                 cached_id)
+                loss = metric - pruned_metric
+                _logger.info("pruned param: {}; {}; loss={}".format(
+                    param, ratio, loss))
+                for brother in self.pruned_list[0]:
+                    if re.match(self.pruned_params, brother):
+                        if brother not in sensitivities:
+                            sensitivities[brother] = {
+                                'pruned_percent': [],
+                                'loss': []
+                            }
+                        sensitivities[brother]['pruned_percent'].append(ratio)
+                        sensitivities[brother]['loss'].append(loss)
+
+                self._save_sensitivities(sensitivities, sensitivities_file)
+
+                # restore pruned parameters
+                for param_name in self.backup.keys():
+                    param_t = context.scope.find_var(param_name).get_tensor()
+                    param_t.set(self.backup[param_name], context.place)
+
+#                pruned_metric = self._eval_graph(context)
+                self.backup = {}
+
+                ratio += self.delta_rate
+        return sensitivities
+
+    def _get_best_ratios(self, context, sensitivities, target_ratio):
+        """
+        Search a group of ratios for pruning target flops.
+        """
+        _logger.info('_get_best_ratios for pruning ratie: {}'.format(
+            target_ratio))
+        self.param_shape_backup = {}
+        self.backup = {}
+
+        def func(params, x):
+            a, b, c, d = params
+            return a * x * x * x + b * x * x + c * x + d
+
+        def error(params, x, y):
+            return func(params, x) - y
+
+        def slove_coefficient(x, y):
+            init_coefficient = [10, 10, 10, 10]
+            coefficient, loss = leastsq(error, init_coefficient, args=(x, y))
+            return coefficient
+
+        min_loss = 0.
+        max_loss = 0.
+
+        # step 1: fit curve by sensitivities
+        coefficients = {}
+        for param in sensitivities:
+            losses = np.array([0] * 5 + sensitivities[param]['loss'])
+            precents = np.array([0] * 5 + sensitivities[param][
+                'pruned_percent'])
+            coefficients[param] = slove_coefficient(precents, losses)
+            loss = np.max(losses)
+            max_loss = np.max([max_loss, loss])
+
+        # step 2: Find a group of ratios by binary searching.
+        flops = context.eval_graph.flops()
+        model_size = context.eval_graph.numel_params()
+        ratios = []
+        while min_loss < max_loss:
+            loss = (max_loss + min_loss) / 2
+            _logger.info(
+                '-----------Try pruned ratios while acc loss={:.4f}-----------'.
+                format(loss))
+            ratios = []
+            # step 2.1: Get ratios according to current loss
+            for param in sensitivities:
+                coefficient = copy.deepcopy(coefficients[param])
+                coefficient[-1] = coefficient[-1] - loss
+                roots = np.roots(coefficient)
+                for root in roots:
+                    min_root = 1
+                    if np.isreal(root) and root > 0 and root < 1:
+                        selected_root = min(root.real, min_root)
+                ratios.append(selected_root)
+            _logger.info('Pruned ratios={}'.format(
+                [round(ratio, 3) for ratio in ratios]))
+            # step 2.2: Pruning by current ratios
+            self._prune_parameters(
+                context.eval_graph,
+                context.scope,
+                sensitivities.keys(),
+                ratios,
+                context.place,
+                only_graph=True)
+
+            pruned_flops = 1 - (float(context.eval_graph.flops()) / flops)
+            pruned_size = 1 - (float(context.eval_graph.numel_params()) /
+                               model_size)
+            _logger.info('Pruned flops: {:.4f}'.format(pruned_flops))
+            _logger.info('Pruned model size: {:.4f}'.format(pruned_size))
+            for param in self.param_shape_backup.keys():
+                context.eval_graph.var(param).set_shape(self.param_shape_backup[
+                    param])
+            self.param_shape_backup = {}
+
+            # step 2.3: Check whether current ratios is enough
+            if abs(pruned_flops - target_ratio) < 0.015:
+                break
+            if pruned_flops > target_ratio:
+                max_loss = loss
+            else:
+                min_loss = loss
+        return sensitivities.keys(), ratios
+
+    def _current_pruning_target(self, context):
+        '''
+        Get the target pruning rate in current epoch.
+        '''
+        _logger.info('Left number of pruning steps: {}'.format(self.num_steps))
+        if self.num_steps <= 0:
+            return None
+        if (self.start_epoch == context.epoch_id) or context.eval_converged(
+                self.metric_name, 0.005):
+            self.num_steps -= 1
+            return self.pruning_step
+
+    def on_epoch_begin(self, context):
+        current_ratio = self._current_pruning_target(context)
+        if current_ratio is not None:
+            sensitivities = self._compute_sensitivities(context)
+            params, ratios = self._get_best_ratios(context, sensitivities,
+                                                   current_ratio)
+            self._prune_parameters(context.optimize_graph, context.scope,
+                                   params, ratios, context.place)
+
+            self.param_shape_backup = {}
+            self.backup = {}
+
+            model_size = context.eval_graph.numel_params()
+            flops = context.eval_graph.flops()
+            _logger.debug('################################')
+            _logger.debug('#          pruning eval graph    #')
+            _logger.debug('################################')
+            self._prune_graph(context.eval_graph, context.optimize_graph)
+            context.optimize_graph.update_groups_of_conv()
+            context.eval_graph.update_groups_of_conv()
+            context.optimize_graph.compile()  # to update the compiled program
+            context.eval_graph.compile(
+                for_parallel=False,
+                for_test=True)  # to update the compiled program
+            _logger.info(
+                '------------------finish pruning--------------------------------'
+            )
+            _logger.info('Pruned size: {:.3f}'.format(1 - (float(
+                context.eval_graph.numel_params()) / model_size)))
+            _logger.info('Pruned flops: {:.3f}'.format(1 - (float(
+                context.eval_graph.flops()) / flops)))
+            metric = self._eval_graph(context)
+            _logger.info('Metric after pruning: {:.2f}'.format(metric))
+            _logger.info(
+                '------------------SensitivePruneStrategy.on_epoch_begin finish--------------------------------'
+            )

python/paddle/fluid/contrib/slim/prune/pruner.py

@@ -13,9 +13,10 @@
 # limitations under the License.
 
 import numpy as np
+import collections
 from .... import layers
 
-__all__ = ['Pruner', 'MagnitudePruner', 'RatioPruner']
+__all__ = ['Pruner', 'StructurePruner']
 
 
 class Pruner(object):
@@ -30,54 +31,77 @@ class Pruner(object):
         pass
 
 
-class MagnitudePruner(Pruner):
+class StructurePruner(Pruner):
     """
-    Pruner used to pruning a parameter by threshold.
+    Pruner used to pruning parameters by groups.
     """
 
-    def __init__(self, threshold):
-        self.threshold = threshold
-
-    def prune(self, param, threshold=None):
-        if threshold is None:
-            thres = layers.fill_constant(
-                shape=[1], dtype='float32', value=self.threshold)
-        else:
-            thres = threshold
-        zeros_mask = layers.less_than(x=param, y=thres)
-        return zeros_mask
-
-
-class RatioPruner(Pruner):
-    """
-    Pruner used to pruning a parameter by ratio.
-    """
+    def __init__(self, pruning_axis, criterions):
+        """
+        Args:
+            pruning_axis(dict): The key is the name of parameter to be pruned,
+                                '*' means all the parameters.
+                                The value is the axis to be used. Given a parameter
+                                with shape [3, 4], the result of pruning 50% on aixs 1
+                                is a parameter with shape [3, 2].
+            criterions(dict): The key is the name of parameter to be pruned,
+                              '*' means all the parameters.
+                              The value is the criterion used to sort groups for pruning.
+                              It only supports 'l1_norm' currently.
+        """
+        self.pruning_axis = pruning_axis
+        self.criterions = criterions
 
-    def __init__(self, ratios=None):
+    def cal_pruned_idx(self, name, param, ratio, axis=None):
         """
+        Calculate the index to be pruned on axis by given pruning ratio.
         Args:
-            ratios: dict with pair (paramer_name, pruned_ratio). 
+            name(str): The name of parameter to be pruned.
+            param(np.array): The data of parameter to be pruned.
+            ratio(float): The ratio to be pruned.
+            axis(int): The axis to be used for pruning given parameter.
+                       If it is None, the value in self.pruning_axis will be used.
+                       default: None.
+        Returns:
+            list<int>: The indexes to be pruned on axis.
         """
-        self.ratios = ratios
+        criterion = self.criterions[
+            name] if name in self.criterions else self.criterions['*']
+        if axis is None:
+            assert self.pruning_axis is not None, "pruning_axis should set if axis is None."
+            axis = self.pruning_axis[
+                name] if name in self.pruning_axis else self.pruning_axis['*']
+        prune_num = int(round(param.shape[axis] * ratio))
+        reduce_dims = [i for i in range(len(param.shape)) if i != axis]
+        if criterion == 'l1_norm':
+            criterions = np.sum(np.abs(param), axis=tuple(reduce_dims))
+        pruned_idx = criterions.argsort()[:prune_num]
+        return pruned_idx
 
-    def prune(self, param, ratio=None):
+    def prune_tensor(self, tensor, pruned_idx, pruned_axis, lazy=False):
         """
+        Pruning a array by indexes on given axis.
         Args:
-            ratio: `ratio=40%` means pruning (1 - 40%) weights to zero.
+            tensor(numpy.array): The target array to be pruned.
+            pruned_idx(list<int>): The indexes to be pruned.
+            pruned_axis(int): The axis of given array to be pruned on. 
+            lazy(bool): True means setting the pruned elements to zero.
+                        False means remove the pruned elements from memory.
+                        default: False.
+        Returns:
+            numpy.array: The pruned array.
         """
-        if ratio is None:
-            rat = self.ratios[
-                param.name] if param.name in self.ratios else self.ratios['*']
-        else:
-            rat = ratio
-        if rat < 1.0:
-            k = max(int(rat * np.prod(param.shape)), 1)
-            param_vec = layers.reshape(x=param, shape=[1, -1])
-            param_topk, _ = layers.topk(param_vec, k=k)
-            threshold = layers.slice(
-                param_topk, axes=[1], starts=[-1], ends=[k])
-            threshold = layers.reshape(x=threshold, shape=[1])
-            zeros_mask = layers.less_than(x=param, y=threshold)
+        mask = np.zeros(tensor.shape[pruned_axis], dtype=bool)
+        mask[pruned_idx] = True
+
+        def func(data):
+            return data[~mask]
+
+        def lazy_func(data):
+            data[mask] = 0
+            return data
+
+        if lazy:
+            return np.apply_along_axis(lazy_func, pruned_axis, tensor)
         else:
-            zeros_mask = layers.ones(param.shape)
-        return zeros_mask
+            return np.apply_along_axis(func, pruned_axis, tensor)

python/paddle/fluid/contrib/slim/quantization/__init__.py

@@ -16,5 +16,7 @@ from __future__ import print_function
 
 from . import quantization_pass
 from .quantization_pass import *
+from . import quantization_strategy
+from .quantization_strategy import *
 
-__all__ = quantization_pass.__all__
+__all__ = quantization_pass.__all__ + quantization_strategy.__all__

python/paddle/fluid/contrib/slim/quantization/quantization_strategy.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import sys
+import numpy as np
+from .... import Executor
+from .... import io
+from .... import core
+from ....compiler import CompiledProgram
+from ....compiler import BuildStrategy
+from ....framework import IrGraph
+from ..core.strategy import Strategy
+from .quantization_pass import *
+
+__all__ = ['QuantizationStrategy']
+
+logging.basicConfig(format='%(asctime)s-%(levelname)s: %(message)s')
+_logger = logging.getLogger(__name__)
+_logger.setLevel(logging.INFO)
+
+
+class QuantizationStrategy(Strategy):
+    """
+    The strategy for Quantization.
+    """
+
+    def __init__(self,
+                 start_epoch=0,
+                 end_epoch=0,
+                 float_model_save_path=None,
+                 mobile_model_save_path=None,
+                 int8_model_save_path=None,
+                 activation_bits=8,
+                 weight_bits=8,
+                 activation_quantize_type='abs_max',
+                 save_in_nodes=None,
+                 save_out_nodes=None):
+        """
+        Args:
+            start_epoch(int): The 'on_epoch_begin' function will be called in start_epoch. default: 0
+            end_epoch(int): The 'on_epoch_end' function will be called in end_epoch. default: 0
+            float_model_save_path(str): The path to save model with float weights. 
+                            None means it doesn't save float model. defalut: None.
+            mobile_model_save_path(str): The path to save model for paddle-mobile execution.
+                            None means it doesn't save mobile model. defalut: None.
+            int8_model_save_path(str): The path to save model with int8_t weight.
+                            None means it doesn't save int8 model. defalut: None.
+            activation_bits(int): quantization bit number for activation. default: 8.
+            weight_bits(int): quantization bit number for weights. The bias is not quantized.
+                              default: 8.
+            activation_quantize_type(str): quantization type for activation,
+                now support 'abs_max', 'range_abs_max' and 'moving_average_abs_max'.
+                If use 'abs_max' mode, the quantization scale will be calculated
+                dynamically each step in both training and testing period. If use
+                'range_abs_max', a static quantization scale will be calculated
+                during training and used in inference.
+            save_in_nodes(list<str>): A list of variable names used to prune graph 
+                                      for saving inference model.
+            save_out_nodes(list<str>): A list of variable names used to prune graph 
+                                      for saving inference model.
+
+        """
+        super(QuantizationStrategy, self).__init__(start_epoch, end_epoch)
+        self.start_epoch = start_epoch
+        self.end_epoch = end_epoch
+        self.float_model_save_path = float_model_save_path
+        self.mobile_model_save_path = mobile_model_save_path
+        self.int8_model_save_path = int8_model_save_path
+        self.activation_bits = activation_bits
+        self.weight_bits = weight_bits
+        self.activation_quantize_type = activation_quantize_type
+        self.save_out_nodes = save_out_nodes
+        self.save_in_nodes = save_in_nodes
+
+    def on_epoch_begin(self, context):
+        """
+        Insert fake_quantize_op and fake_dequantize_op before trainging and testing.
+        """
+        super(QuantizationStrategy, self).on_compression_begin(context)
+        if self.start_epoch == context.epoch_id:
+            _logger.info('QuantizationStrategy::on_epoch_begin')
+            train_ir_graph = IrGraph(
+                core.Graph(context.optimize_graph.program.desc), for_test=False)
+            test_ir_graph = IrGraph(
+                core.Graph(context.eval_graph.program.desc), for_test=True)
+            transform_pass = QuantizationTransformPass(
+                scope=context.scope,
+                place=context.place,
+                weight_bits=self.weight_bits,
+                activation_bits=self.activation_bits,
+                activation_quantize_type=self.activation_quantize_type)
+            transform_pass.apply(train_ir_graph)
+            transform_pass.apply(test_ir_graph)
+
+            build_strategy = BuildStrategy()
+            build_strategy.enable_inplace = False
+            build_strategy.memory_optimize = False
+            # for quantization training
+            context.optimize_graph.compiled_graph = CompiledProgram(
+                train_ir_graph.graph).with_data_parallel(
+                    loss_name=context.optimize_graph.out_nodes['loss'],
+                    build_strategy=build_strategy)
+            # for evaluation. And program compiled from ir graph must be with data parallel.
+            context.eval_graph.compiled_graph = CompiledProgram(
+                test_ir_graph.graph).with_data_parallel(
+                    build_strategy=build_strategy)
+            # for saving inference model after training
+            context.put('quantization_test_ir_graph_backup', test_ir_graph)
+            _logger.info('Finish QuantizationStrategy::on_epoch_begin')
+
+    def on_epoch_end(self, context):
+        """
+        Free and save inference model.
+        """
+        super(QuantizationStrategy, self).on_compression_end(context)
+
+        if context.epoch_id == self.end_epoch:
+            _logger.info('QuantizationStrategy::on_epoch_end')
+            test_ir_graph = context.get('quantization_test_ir_graph_backup')
+            # freeze the graph after training
+            freeze_pass = QuantizationFreezePass(
+                scope=context.scope,
+                place=context.place,
+                weight_bits=self.weight_bits,
+                activation_bits=self.activation_bits)
+            freeze_pass.apply(test_ir_graph)
+
+            # for other strategies
+            context.eval_graph.program = test_ir_graph.to_program()
+
+            if self.save_out_nodes == None:
+                out_vars = [
+                    context.eval_graph.var(var_name)._var
+                    for var_name in context.eval_graph.out_nodes.values()
+                ]
+            else:
+                out_vars = [
+                    context.eval_graph.var(var_name)._var
+                    for var_name in self.save_out_nodes
+                ]
+
+            if self.save_in_nodes == None:
+                in_vars = list(context.eval_graph.out_nodes.values())
+            else:
+                in_vars = self.save_in_nodes
+
+            # save float model
+            if self.float_model_save_path:
+                executor = Executor(context.place)
+                io.save_inference_model(
+                    self.float_model_save_path,
+                    in_vars,
+                    out_vars,
+                    executor,
+                    main_program=test_ir_graph.to_program(),
+                    model_filename='model',
+                    params_filename='weights',
+                    export_for_deployment=True)
+
+            # save int8 model
+            if self.int8_model_save_path:
+                convert_int8_pass = ConvertToInt8Pass(
+                    scope=context.scope, place=context.place)
+                convert_int8_pass.apply(test_ir_graph)
+
+                executor = Executor(context.place)
+                io.save_inference_model(
+                    self.int8_model_save_path,
+                    in_vars,
+                    out_vars,
+                    executor,
+                    main_program=test_ir_graph.to_program(),
+                    model_filename='model',
+                    params_filename='weights',
+                    export_for_deployment=True)
+
+            # save mobile model
+            if self.mobile_model_save_path:
+                if not self.int8_model_save_path:
+                    # convert the weights as int8_t type
+                    convert_int8_pass = ConvertToInt8Pass(
+                        scope=context.scope, place=context.place)
+                    convert_int8_pass.apply(test_ir_graph)
+                # make some changes on the graph for the mobile inference
+                mobile_pass = TransformForMobilePass()
+                mobile_pass.apply(test_ir_graph)
+                executor = Executor(context.place)
+                io.save_inference_model(
+                    self.mobile_model_save_path,
+                    in_vars,
+                    out_vars,
+                    executor,
+                    main_program=test_ir_graph.to_program(),
+                    model_filename='model',
+                    params_filename='weights',
+                    export_for_deployment=True)
+            _logger.info('Finish QuantizationStrategy::on_epoch_end')

python/paddle/fluid/contrib/slim/tests/configs/config.yaml

-version: 1.0
-include: ["./configs/pruners.yaml", "./configs/pruners_0.yaml"]
-pruners:
-    pruner_1:
-        class: 'RatioPruner'
-        ratios:
-            'conv1_1.w': 0.3
-            'conv1_2.w': 0.4
-            '*': 0.9
-        group_dims:
-            '*': [1, 2, 3]
-        criterions:
-            '*': 'l1-norm'
-strategies:
-    strategy_1:
-        class: 'SensitivePruneStrategy'
-        pruner: 'pruner_2'
-        start_epoch: 0
-        end_epoch: 10
-        delta_rate: 0.20
-        acc_loss_threshold: 0.2
-        sensitivities:
-            'conv1_1.w': 0.4
-
-compress_pass:
-    class: 'CompressPass'
-    epoch: 100
-    strategies:
-        - strategy_1

python/paddle/fluid/contrib/slim/tests/configs/filter_pruning.yaml

+#start_epoch:         The 'on_epoch_begin' function will be called in start_epoch. default: 0.
+#end_epoch:           The 'on_epoch_end' function will be called in end_epoch. default: 10.
+#delta_rate:          The delta used to generate ratios when calculating sensitivities.
+#target_ratio:        The flops ratio to be pruned from current model.
+#metric_name:         The metric used to evaluate the model.
+#pruned_params:       The pattern str to match the parameter names to be pruned.
+#sensitivities_file:  The sensitivities file.
+#num_steps:           The number of pruning steps.
+#eval_rate:           The rate of sampled data used to calculate sensitivities.
+version: 1.0
+pruners:
+    pruner_1:
+        class: 'StructurePruner'
+        pruning_axis:
+            '*': 0
+        criterions:
+            '*': 'l1_norm'
+strategies:
+    sensitive_pruning_strategy:
+        class: 'SensitivePruneStrategy'
+        pruner: 'pruner_1'
+        start_epoch: 0
+        delta_rate: 0.1
+        target_ratio: 0.3
+        num_steps: 1
+        eval_rate: 0.5
+        pruned_params: '.*_sep_weights'
+        sensitivities_file: 'mobilenet_acc_top1_sensitive.data'
+        metric_name: 'acc_top1'
+compressor:
+    epoch: 120
+    checkpoint_path: './checkpoints/'
+    strategies:
+        - sensitive_pruning_strategy

python/paddle/fluid/contrib/slim/tests/configs/pruners.yaml

-version: 1.0
-pruners:
-    pruner_2:
-        class: 'RatioPruner'
-        ratios:
-            'conv1_1.w': 0.5
-            'conv1_2.w': 0.2
-            '*': 0.7
-        group_dims:
-            '*': [1, 2, 3]
-        criterions:
-            '*': 'l1-norm'

python/paddle/fluid/contrib/slim/tests/configs/pruners_0.yaml

-version: 1.0
-pruners:
-    pruner_3:
-        class: 'RatioPruner'
-        ratios:
-            'conv1_1.w': 0.5
-            'conv1_2.w': 0.2
-            '*': 0.7
-        group_dims:
-            '*': [1, 2, 3]
-        criterions:
-            '*': 'l1-norm'

python/paddle/fluid/contrib/slim/tests/distillation/compress.yaml

+#start_epoch(int): The epoch when to merge student graph and teacher graph for
+#                  distillation training. default: 0
+#
+#end_epoch(int): The epoch when to finish distillation training. default: 0
+#
+#student_feature_map(str): The name of feature map from student network.
+#
+#teacher_feature_map(str): The name of feature map from teacher network.
+#                          It's shape should be the same with student network.
+#
+#student_pairs(list<tuple>): Each tuple, with two variable names, in student_pairs indicates
+#                            a section in student network. The variables in a tuple should
+#                            have the same feature map size.
+#
+#teacher_pairs(list<tuple>): Each tuple, with two variable names, in teacher_pairs indicates
+#                            a section in teacher network. The variables in a tuple should
+#                            have the same feature map size. Varibale named teacher_pairs[i][j]
+#                            should has the save channel number with that of variable named 
+#                            student_pairs[i][j].
+#
+#distillation_loss_weight(float): The weight of the loss.
+version: 1.0
+distillers:
+    fsp_distiller:
+        class: 'FSPDistiller'
+#        teacher_pairs: [['teacher_depthwise_conv2d_1.tmp_0', 'teacher_conv2d_3.tmp_0']]
+#        student_pairs: [['student_depthwise_conv2d_1.tmp_0', 'student_conv2d_3.tmp_0']]
+        teacher_pairs: [['teacher_conv2_1_dw.tmp_0', 'teacher_conv1.tmp_0']]
+        student_pairs: [['student_conv2_1_dw.tmp_0', 'student_conv1.tmp_0']]
+        distillation_loss_weight: 1
+    l2_distiller:
+        class: 'L2Distiller'
+        teacher_feature_map: 'teacher.tmp_2'
+        student_feature_map: 'student.tmp_2'
+        distillation_loss_weight: 1
+strategies:
+    distillation_strategy:
+        class: 'DistillationStrategy'
+        distillers: ['fsp_distiller', 'l2_distiller']
+        start_epoch: 0
+        end_epoch: 1
+compressor:
+    epoch: 1
+    checkpoint_path: './distillation_checkpoints/'
+    strategies:
+        - distillation_strategy

python/paddle/fluid/contrib/slim/tests/filter_pruning/compress.yaml

+#start_epoch:         The 'on_epoch_begin' function will be called in start_epoch. default: 0.
+#end_epoch:           The 'on_epoch_end' function will be called in end_epoch. default: 10.
+#delta_rate:          The delta used to generate ratios when calculating sensitivities.
+#target_ratio:        The flops ratio to be pruned from current model.
+#metric_name:         The metric used to evaluate the model.
+#pruned_params:       The pattern str to match the parameter names to be pruned.
+#sensitivities_file:  The sensitivities file.
+#num_steps:           The number of pruning steps.
+#eval_rate:           The rate of sampled data used to calculate sensitivities.
+version: 1.0
+pruners:
+    pruner_1:
+        class: 'StructurePruner'
+        pruning_axis:
+            '*': 0
+        criterions:
+            '*': 'l1_norm'
+strategies:
+    sensitive_pruning_strategy:
+        class: 'SensitivePruneStrategy'
+        pruner: 'pruner_1'
+        start_epoch: 1
+        delta_rate: 0.2
+        target_ratio: 0.08
+        num_steps: 1
+        eval_rate: 0.5
+        pruned_params: 'conv6_sep_weights'
+        sensitivities_file: 'mobilenet_acc_top1_sensitive.data'
+        metric_name: 'acc_top1'
+compressor:
+    epoch: 2
+    checkpoint_path: './checkpoints_pruning/'
+    strategies:
+        - sensitive_pruning_strategy

python/paddle/fluid/contrib/slim/tests/mobilenet.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import paddle.fluid as fluid
+from paddle.fluid.initializer import MSRA
+from paddle.fluid.param_attr import ParamAttr
+
+__all__ = ['MobileNet']
+
+train_parameters = {
+    "input_size": [3, 224, 224],
+    "input_mean": [0.485, 0.456, 0.406],
+    "input_std": [0.229, 0.224, 0.225],
+    "learning_strategy": {
+        "name": "piecewise_decay",
+        "batch_size": 256,
+        "epochs": [30, 60, 90],
+        "steps": [0.1, 0.01, 0.001, 0.0001]
+    }
+}
+
+
+class MobileNet():
+    def __init__(self, name=""):
+        self.params = train_parameters
+        self.name = name
+
+    def net(self, input, class_dim=1000, scale=1.0):
+        # conv1: 112x112
+        input = self.conv_bn_layer(
+            input,
+            filter_size=3,
+            channels=3,
+            num_filters=int(32 * scale),
+            stride=2,
+            padding=1,
+            name=self.name + "_conv1")
+
+        # 56x56
+        input = self.depthwise_separable(
+            input,
+            num_filters1=32,
+            num_filters2=64,
+            num_groups=32,
+            stride=1,
+            scale=scale,
+            name=self.name + "_conv2_1")
+
+        input = self.depthwise_separable(
+            input,
+            num_filters1=64,
+            num_filters2=128,
+            num_groups=64,
+            stride=2,
+            scale=scale,
+            name=self.name + "_conv2_2")
+
+        # 28x28
+        input = self.depthwise_separable(
+            input,
+            num_filters1=128,
+            num_filters2=128,
+            num_groups=128,
+            stride=1,
+            scale=scale,
+            name=self.name + "_conv3_1")
+
+        input = self.depthwise_separable(
+            input,
+            num_filters1=128,
+            num_filters2=256,
+            num_groups=128,
+            stride=2,
+            scale=scale,
+            name=self.name + "_conv3_2")
+
+        # 14x14
+        input = self.depthwise_separable(
+            input,
+            num_filters1=256,
+            num_filters2=256,
+            num_groups=256,
+            stride=1,
+            scale=scale,
+            name=self.name + "_conv4_1")
+
+        input = self.depthwise_separable(
+            input,
+            num_filters1=256,
+            num_filters2=512,
+            num_groups=256,
+            stride=2,
+            scale=scale,
+            name=self.name + "_conv4_2")
+
+        # 14x14
+        for i in range(5):
+            input = self.depthwise_separable(
+                input,
+                num_filters1=512,
+                num_filters2=512,
+                num_groups=512,
+                stride=1,
+                scale=scale,
+                name=self.name + "_conv5" + "_" + str(i + 1))
+        # 7x7
+        input = self.depthwise_separable(
+            input,
+            num_filters1=512,
+            num_filters2=1024,
+            num_groups=512,
+            stride=2,
+            scale=scale,
+            name=self.name + "_conv5_6")
+
+        input = self.depthwise_separable(
+            input,
+            num_filters1=1024,
+            num_filters2=1024,
+            num_groups=1024,
+            stride=1,
+            scale=scale,
+            name=self.name + "_conv6")
+
+        input = fluid.layers.pool2d(
+            input=input,
+            pool_size=0,
+            pool_stride=1,
+            pool_type='avg',
+            global_pooling=True)
+
+        output = fluid.layers.fc(
+            input=input,
+            size=class_dim,
+            act='softmax',
+            param_attr=ParamAttr(
+                initializer=MSRA(), name=self.name + "_fc7_weights"),
+            bias_attr=ParamAttr(name=self.name + "_fc7_offset"),
+            name=self.name)
+        return output
+
+    def conv_bn_layer(self,
+                      input,
+                      filter_size,
+                      num_filters,
+                      stride,
+                      padding,
+                      channels=None,
+                      num_groups=1,
+                      act='relu',
+                      use_cudnn=True,
+                      name=None):
+        conv = fluid.layers.conv2d(
+            input=input,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=num_groups,
+            act=None,
+            use_cudnn=use_cudnn,
+            param_attr=ParamAttr(
+                initializer=MSRA(), name=name + "_weights"),
+            name=name,
+            bias_attr=False)
+        bn_name = name + "_bn"
+        return fluid.layers.batch_norm(
+            input=conv,
+            act=act,
+            name=name,
+            param_attr=ParamAttr(name=bn_name + "_scale"),
+            bias_attr=ParamAttr(name=bn_name + "_offset"),
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance')
+
+    def depthwise_separable(self,
+                            input,
+                            num_filters1,
+                            num_filters2,
+                            num_groups,
+                            stride,
+                            scale,
+                            name=None):
+        depthwise_conv = self.conv_bn_layer(
+            input=input,
+            filter_size=3,
+            num_filters=int(num_filters1 * scale),
+            stride=stride,
+            padding=1,
+            num_groups=int(num_groups * scale),
+            use_cudnn=False,
+            name=name + "_dw")
+
+        pointwise_conv = self.conv_bn_layer(
+            input=depthwise_conv,
+            filter_size=1,
+            num_filters=int(num_filters2 * scale),
+            stride=1,
+            padding=0,
+            name=name + "_sep")
+        return pointwise_conv

python/paddle/fluid/contrib/slim/tests/quantization/compress.yaml

+#start_epoch(int): The epoch to insert quantization operators. default: 0
+#
+#end_epoch(int): The epoch to save inferecne model. default: 0
+#
+#float_model_save_path(str): The path to save model with float weights.
+#                None means it doesn't save float model. defalut: None.
+#
+#mobile_model_save_path(str): The path to save model for paddle-mobile execution.
+#                None means it doesn't save mobile model. defalut: None.
+#
+#int8_model_save_path(str): The path to save model with int8_t weight.
+#                None means it doesn't save int8 model. defalut: None.
+#
+#activation_bits(int): quantization bit number for activation. default: 8.
+#
+#weight_bits(int): quantization bit number for weights. The bias is not quantized.
+#                  default: 8.
+#
+#activation_quantize_type(str): quantization type for activation,
+#    now support 'abs_max', 'range_abs_max' and 'moving_average_abs_max'.
+#    If use 'abs_max' mode, the quantization scale will be calculated
+#    dynamically each step in both training and testing period. If use
+#    'range_abs_max', a static quantization scale will be calculated
+#    during training and used in inference.
+#
+#save_in_nodes(list<str>): A list of variable names used to prune graph
+#                          for saving inference model.
+#
+#save_out_nodes(list<str>): A list of variable names used to prune graph
+#                                      for saving inference model.
+version: 1.0
+strategies:
+    quantization_strategy:
+        class: 'QuantizationStrategy'
+        start_epoch: 0
+        end_epoch: 0
+        float_model_save_path: './output/float'
+        weight_bits: 8
+        activation_bits: 8
+        weight_quantize_type: 'abs_max'
+        activation_quantize_type: 'abs_max'
+        save_in_nodes: ['image']
+        save_out_nodes: ['quan.tmp_2']
+compressor:
+    epoch: 1
+    checkpoint_path: './checkpoints_quan/'
+    strategies:
+        - quantization_strategy

python/paddle/fluid/contrib/slim/tests/test_distillation_strategy.py

+#   copyright (c) 2019 paddlepaddle authors. all rights reserved.
+#
+# licensed under the apache license, version 2.0 (the "license");
+# you may not use this file except in compliance with the license.
+# you may obtain a copy of the license at
+#
+#     http://www.apache.org/licenses/license-2.0
+#
+# unless required by applicable law or agreed to in writing, software
+# distributed under the license is distributed on an "as is" basis,
+# without warranties or conditions of any kind, either express or implied.
+# see the license for the specific language governing permissions and
+# limitations under the license.
+
+import paddle
+import unittest
+import paddle.fluid as fluid
+from mobilenet import MobileNet
+from paddle.fluid.contrib.slim.core import Compressor
+from paddle.fluid.contrib.slim.graph import GraphWrapper
+
+
+class TestDistillationStrategy(unittest.TestCase):
+    """
+    Test API of distillation strategy.
+    """
+
+    def test_compression(self):
+        if not fluid.core.is_compiled_with_cuda():
+            return
+        class_dim = 10
+        image_shape = [1, 28, 28]
+        image = fluid.layers.data(
+            name='image', shape=image_shape, dtype='float32')
+        image.stop_gradient = False
+        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        out = MobileNet(name="student").net(input=image, class_dim=class_dim)
+        acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
+        acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
+        val_program = fluid.default_main_program().clone(for_test=False)
+
+        cost = fluid.layers.cross_entropy(input=out, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+        optimizer = fluid.optimizer.Momentum(
+            momentum=0.9,
+            learning_rate=0.01,
+            regularization=fluid.regularizer.L2Decay(4e-5))
+
+        place = fluid.CUDAPlace(0)
+        exe = fluid.Executor(place)
+        exe.run(fluid.default_startup_program())
+
+        val_reader = paddle.batch(paddle.dataset.mnist.test(), batch_size=128)
+
+        val_feed_list = [('img', image.name), ('label', label.name)]
+        val_fetch_list = [('acc_top1', acc_top1.name), ('acc_top5',
+                                                        acc_top5.name)]
+
+        train_reader = paddle.batch(
+            paddle.dataset.mnist.train(), batch_size=128)
+        train_feed_list = [('img', image.name), ('label', label.name)]
+        train_fetch_list = [('loss', avg_cost.name)]
+
+        # define teacher program
+        teacher_program = fluid.Program()
+        startup_program = fluid.Program()
+        with fluid.program_guard(teacher_program, startup_program):
+            img = teacher_program.global_block()._clone_variable(
+                image, force_persistable=False)
+            predict = MobileNet(name="teacher").net(input=img,
+                                                    class_dim=class_dim)
+
+        exe.run(startup_program)
+
+        com_pass = Compressor(
+            place,
+            fluid.global_scope(),
+            fluid.default_main_program(),
+            train_reader=train_reader,
+            train_feed_list=train_feed_list,
+            train_fetch_list=train_fetch_list,
+            eval_program=val_program,
+            eval_reader=val_reader,
+            eval_feed_list=val_feed_list,
+            eval_fetch_list=val_fetch_list,
+            teacher_programs=[teacher_program.clone(for_test=True)],
+            train_optimizer=optimizer,
+            distiller_optimizer=optimizer)
+        com_pass.config('./distillation/compress.yaml')
+        eval_graph = com_pass.run()
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/contrib/slim/tests/test_factory.py

@@ -12,29 +12,25 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from paddle.fluid.contrib.slim import ConfigFactory
+from paddle.fluid.contrib.slim.core import ConfigFactory
 import unittest
 
 
 class TestFactory(unittest.TestCase):
-    def test_parse(self):
-        factory = ConfigFactory('./configs/config.yaml')
+    def test_parse_pruning(self):
+        factory = ConfigFactory('./configs/filter_pruning.yaml')
 
-        pruner = factory.instance('pruner_1')
-        self.assertEquals(pruner.ratios['conv1_1.w'], 0.3)
+        pruner_1 = factory.instance('pruner_1')
+        self.assertEquals(pruner_1.pruning_axis['*'], 0)
+        self.assertEquals(pruner_1.criterions['*'], 'l1_norm')
 
-        pruner = factory.instance('pruner_2')
-        self.assertEquals(pruner.ratios['*'], 0.7)
+        strategy = factory.instance('sensitive_pruning_strategy')
+        pruner_1 = strategy.pruner
+        self.assertEquals(pruner_1.criterions['*'], 'l1_norm')
 
-        strategy = factory.instance('strategy_1')
-        pruner = strategy.pruner
-        self.assertEquals(pruner.ratios['*'], 0.7)
-
-        compress_pass = factory.get_compress_pass()
-        self.assertEquals(compress_pass.epoch, 100)
-
-        strategy = compress_pass.strategies[0]
-        self.assertEquals(strategy.delta_rate, 0.2)
+        self.assertEquals(strategy.start_epoch, 0)
+        self.assertEquals(strategy.sensitivities_file,
+                          'mobilenet_acc_top1_sensitive.data')
 
 
 if __name__ == '__main__':

python/paddle/fluid/contrib/slim/tests/test_filter_pruning.py

+#   copyright (c) 2019 paddlepaddle authors. all rights reserved.
+#
+# licensed under the apache license, version 2.0 (the "license");
+# you may not use this file except in compliance with the license.
+# you may obtain a copy of the license at
+#
+#     http://www.apache.org/licenses/license-2.0
+#
+# unless required by applicable law or agreed to in writing, software
+# distributed under the license is distributed on an "as is" basis,
+# without warranties or conditions of any kind, either express or implied.
+# see the license for the specific language governing permissions and
+# limitations under the license.
+
+import paddle
+import unittest
+import paddle.fluid as fluid
+from mobilenet import MobileNet
+from paddle.fluid.contrib.slim.core import Compressor
+from paddle.fluid.contrib.slim.graph import GraphWrapper
+
+
+class TestFilterPruning(unittest.TestCase):
+    def test_compression(self):
+        """
+        Model: mobilenet_v1
+        data: mnist
+        step1: Training one epoch
+        step2: pruning flops
+        step3: fine-tune one epoch
+        step4: check top1_acc.
+        """
+        if not fluid.core.is_compiled_with_cuda():
+            return
+        class_dim = 10
+        image_shape = [1, 28, 28]
+        image = fluid.layers.data(
+            name='image', shape=image_shape, dtype='float32')
+        image.stop_gradient = False
+        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        out = MobileNet().net(input=image, class_dim=class_dim)
+        acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
+        acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
+        val_program = fluid.default_main_program().clone(for_test=False)
+
+        cost = fluid.layers.cross_entropy(input=out, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+
+        optimizer = fluid.optimizer.Momentum(
+            momentum=0.9,
+            learning_rate=0.01,
+            regularization=fluid.regularizer.L2Decay(4e-5))
+
+        place = fluid.CUDAPlace(0)
+        exe = fluid.Executor(place)
+        exe.run(fluid.default_startup_program())
+
+        val_reader = paddle.batch(paddle.dataset.mnist.test(), batch_size=128)
+
+        val_feed_list = [('img', image.name), ('label', label.name)]
+        val_fetch_list = [('acc_top1', acc_top1.name), ('acc_top5',
+                                                        acc_top5.name)]
+
+        train_reader = paddle.batch(
+            paddle.dataset.mnist.train(), batch_size=128)
+        train_feed_list = [('img', image.name), ('label', label.name)]
+        train_fetch_list = [('loss', avg_cost.name)]
+
+        com_pass = Compressor(
+            place,
+            fluid.global_scope(),
+            fluid.default_main_program(),
+            train_reader=train_reader,
+            train_feed_list=train_feed_list,
+            train_fetch_list=train_fetch_list,
+            eval_program=val_program,
+            eval_reader=val_reader,
+            eval_feed_list=val_feed_list,
+            eval_fetch_list=val_fetch_list,
+            train_optimizer=optimizer)
+        com_pass.config('./filter_pruning/compress.yaml')
+        eval_graph = com_pass.run()
+        self.assertTrue(
+            abs((com_pass.context.eval_results['acc_top1'][-1] - 0.969) / 0.969)
+            < 0.02)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/contrib/slim/tests/test_graph_wrapper.py

+#   copyright (c) 2019 paddlepaddle authors. all rights reserved.
+#
+# licensed under the apache license, version 2.0 (the "license");
+# you may not use this file except in compliance with the license.
+# you may obtain a copy of the license at
+#
+#     http://www.apache.org/licenses/license-2.0
+#
+# unless required by applicable law or agreed to in writing, software
+# distributed under the license is distributed on an "as is" basis,
+# without warranties or conditions of any kind, either express or implied.
+# see the license for the specific language governing permissions and
+# limitations under the license.
+
+from __future__ import print_function
+import unittest
+import paddle.fluid as fluid
+import six
+import numpy as np
+from paddle.fluid.contrib.slim.graph import GraphWrapper
+from paddle.fluid import core
+
+
+def residual_block(num):
+    def conv_bn_layer(input,
+                      ch_out,
+                      filter_size,
+                      stride,
+                      padding,
+                      act='relu',
+                      bias_attr=False):
+        tmp = fluid.layers.conv2d(
+            input=input,
+            filter_size=filter_size,
+            num_filters=ch_out,
+            stride=stride,
+            padding=padding,
+            act=None,
+            bias_attr=bias_attr)
+        return fluid.layers.batch_norm(input=tmp, act=act)
+
+    data = fluid.layers.data(name='image', shape=[1, 8, 8], dtype='float32')
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+    data.stop_gradinet = False
+    hidden = data
+    for _ in six.moves.xrange(num):
+        conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True)
+        short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None)
+        hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu')
+    fc = fluid.layers.fc(input=hidden, size=10)
+
+    loss = fluid.layers.cross_entropy(input=fc, label=label)
+    loss = fluid.layers.mean(loss)
+    return data, label, loss
+
+
+class TestGraphWrapper(unittest.TestCase):
+    def build_program(self):
+        place = fluid.CPUPlace()
+        if fluid.core.is_compiled_with_cuda():
+            place = fluid.CUDAPlace(0)
+        main = fluid.Program()
+        startup = fluid.Program()
+        with fluid.program_guard(main, startup):
+            image, label, self.loss = residual_block(2)
+            eval_program = main.clone()
+            opt = fluid.optimizer.SGD(learning_rate=0.001)
+            opt.minimize(self.loss)
+        self.scope = core.Scope()
+        exe = fluid.Executor(place)
+        exe.run(startup, scope=self.scope)
+        self.eval_graph = GraphWrapper(
+            program=eval_program,
+            in_nodes={'image': image.name,
+                      'label': label.name},
+            out_nodes={'loss': self.loss.name})
+        self.train_graph = GraphWrapper(
+            program=main,
+            in_nodes={'image': image.name,
+                      'label': label.name},
+            out_nodes={'loss': self.loss.name})
+
+    def test_all_parameters(self):
+        self.build_program()
+        self.assertEquals(len(self.train_graph.all_parameters()), 24)
+
+    def test_all_vars(self):
+        self.build_program()
+        self.assertEquals(len(self.train_graph.vars()), 90)
+
+    def test_numel_params(self):
+        self.build_program()
+        self.assertEquals(self.train_graph.numel_params(), 13258)
+
+    def test_compile(self):
+        self.build_program()
+        place = fluid.CPUPlace()
+        if fluid.core.is_compiled_with_cuda():
+            place = fluid.CUDAPlace(0)
+        exe = fluid.Executor(place)
+        self.train_graph.compile()
+        exe.run(self.train_graph.compiled_graph,
+                scope=self.scope,
+                feed={
+                    'image':
+                    np.random.randint(0, 40, [16, 1, 8, 8]).astype('float32'),
+                    'label': np.random.randint(0, 10, [16, 1]).astype('int64')
+                })
+
+    def test_pre_and_next_ops(self):
+        self.build_program()
+        for op in self.train_graph.ops():
+            for next_op in self.train_graph.next_ops(op):
+                self.assertTrue(op in self.train_graph.pre_ops(next_op))
+
+    def test_get_optimize_graph(self):
+        self.build_program()
+        place = fluid.CPUPlace()
+        if fluid.core.is_compiled_with_cuda():
+            place = fluid.CUDAPlace(0)
+        opt = fluid.optimizer.SGD(learning_rate=0.001)
+        train_graph = self.eval_graph.get_optimize_graph(
+            opt, place, self.scope, no_grad_var_names=['image'])
+        self.assertEquals(len(self.train_graph.ops()), len(train_graph.ops()))
+        exe = fluid.Executor(place)
+        train_graph.compile()
+        image = np.random.randint(0, 225, [16, 1, 8, 8]).astype('float32')
+        label = np.random.randint(0, 10, [16, 1]).astype('int64')
+        exe.run(train_graph.compiled_graph,
+                scope=self.scope,
+                feed={'image': image,
+                      'label': label})
+
+    def test_flops(self):
+        self.build_program()
+        self.assertEquals(self.train_graph.flops(), 354624)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/contrib/slim/tests/test_quantization_strategy.py

+#   copyright (c) 2019 paddlepaddle authors. all rights reserved.
+#
+# licensed under the apache license, version 2.0 (the "license");
+# you may not use this file except in compliance with the license.
+# you may obtain a copy of the license at
+#
+#     http://www.apache.org/licenses/license-2.0
+#
+# unless required by applicable law or agreed to in writing, software
+# distributed under the license is distributed on an "as is" basis,
+# without warranties or conditions of any kind, either express or implied.
+# see the license for the specific language governing permissions and
+# limitations under the license.
+
+import paddle
+import unittest
+import paddle.fluid as fluid
+from mobilenet import MobileNet
+from paddle.fluid.contrib.slim.core import Compressor
+from paddle.fluid.contrib.slim.graph import GraphWrapper
+
+
+class TestQuantizationStrategy(unittest.TestCase):
+    """
+    Test API of quantization strategy.
+    """
+
+    def test_compression(self):
+        if not fluid.core.is_compiled_with_cuda():
+            return
+        class_dim = 10
+        image_shape = [1, 28, 28]
+        image = fluid.layers.data(
+            name='image', shape=image_shape, dtype='float32')
+        image.stop_gradient = False
+        label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+        out = MobileNet(name='quan').net(input=image, class_dim=class_dim)
+        acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
+        acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
+        val_program = fluid.default_main_program().clone(for_test=False)
+
+        cost = fluid.layers.cross_entropy(input=out, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+
+        optimizer = fluid.optimizer.Momentum(
+            momentum=0.9,
+            learning_rate=0.01,
+            regularization=fluid.regularizer.L2Decay(4e-5))
+
+        place = fluid.CUDAPlace(0)
+        exe = fluid.Executor(place)
+        exe.run(fluid.default_startup_program())
+
+        val_reader = paddle.batch(paddle.dataset.mnist.test(), batch_size=128)
+
+        val_feed_list = [('img', image.name), ('label', label.name)]
+        val_fetch_list = [('acc_top1', acc_top1.name), ('acc_top5',
+                                                        acc_top5.name)]
+
+        train_reader = paddle.batch(
+            paddle.dataset.mnist.train(), batch_size=128)
+        train_feed_list = [('img', image.name), ('label', label.name)]
+        train_fetch_list = [('loss', avg_cost.name)]
+
+        com_pass = Compressor(
+            place,
+            fluid.global_scope(),
+            fluid.default_main_program(),
+            train_reader=train_reader,
+            train_feed_list=train_feed_list,
+            train_fetch_list=train_fetch_list,
+            eval_program=val_program,
+            eval_reader=val_reader,
+            eval_feed_list=val_feed_list,
+            eval_fetch_list=val_fetch_list,
+            train_optimizer=optimizer)
+        com_pass.config('./quantization/compress.yaml')
+        eval_graph = com_pass.run()
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/contrib/tests/test_calibration.py

@@ -136,7 +136,7 @@ class TestCalibrationForResnet50(unittest.TestCase):
                                                         "full_data", False)
         else:
             data_urls.append(
-                'http://paddle-inference-dist.cdn.bcebos.com/int8/calibration_test_data.tar.gz'
+                'http://paddle-inference-dist.bj.bcebos.com/int8/calibration_test_data.tar.gz'
             )
             data_md5s.append('1b6c1c434172cca1bf9ba1e4d7a3157d')
             self.data_cache_folder = self.download_data(data_urls, data_md5s,
@@ -189,7 +189,7 @@ class TestCalibrationForResnet50(unittest.TestCase):
     def download_model(self):
         # resnet50 fp32 data
         data_urls = [
-            'http://paddle-inference-dist.cdn.bcebos.com/int8/resnet50_int8_model.tar.gz'
+            'http://paddle-inference-dist.bj.bcebos.com/int8/resnet50_int8_model.tar.gz'
         ]
         data_md5s = ['4a5194524823d9b76da6e738e1367881']
         self.model_cache_folder = self.download_data(data_urls, data_md5s,
@@ -290,7 +290,7 @@ class TestCalibrationForResnet50(unittest.TestCase):
             self.model, self.infer_iterations)
         (int8_throughput, int8_latency,
          int8_acc1) = self.run_program("calibration_out")
-        delta_value = np.abs(fp32_acc1 - int8_acc1)
+        delta_value = fp32_acc1 - int8_acc1
         self.assertLess(delta_value, 0.01)
         print(
             "FP32 {0}: batch_size {1}, throughput {2} images/second, latency {3} second, accuracy {4}".
@@ -307,7 +307,7 @@ class TestCalibrationForMobilenetv1(TestCalibrationForResnet50):
     def download_model(self):
         # mobilenetv1 fp32 data
         data_urls = [
-            'http://paddle-inference-dist.cdn.bcebos.com/int8/mobilenetv1_int8_model.tar.gz'
+            'http://paddle-inference-dist.bj.bcebos.com/int8/mobilenetv1_int8_model.tar.gz'
         ]
         data_md5s = ['13892b0716d26443a8cdea15b3c6438b']
         self.model_cache_folder = self.download_data(data_urls, data_md5s,

python/paddle/fluid/data_feeder.py

@@ -26,6 +26,24 @@ from .framework import Variable, default_main_program
 __all__ = ['DataFeeder']
 
 
+def convert_dtype(dtype):
+    if dtype == core.VarDesc.VarType.FP32:
+        return 'float32'
+    elif dtype == core.VarDesc.VarType.INT64:
+        return 'int64'
+    elif dtype == core.VarDesc.VarType.FP64:
+        return 'float64'
+    elif dtype == core.VarDesc.VarType.FP16:
+        return 'float16'
+    elif dtype == core.VarDesc.VarType.INT32:
+        return 'int32'
+    elif dtype == core.VarDesc.VarType.UINT8:
+        return 'uint8'
+    else:
+        raise ValueError("dtype must be any of [int32, float32, int64, "
+                         "float64, uint8]")
+
+
 class DataToLoDTensorConverter(object):
     def __init__(self, place, lod_level, shape, dtype):
         self.place = place
@@ -38,27 +56,12 @@ class DataToLoDTensorConverter(object):
             if negtive_count > 1:
                 self.shape = None
                 break
-        if dtype == core.VarDesc.VarType.FP32:
-            self.dtype = 'float32'
-        elif dtype == core.VarDesc.VarType.INT64:
-            self.dtype = 'int64'
-        elif dtype == core.VarDesc.VarType.FP64:
-            self.dtype = 'float64'
-        elif dtype == core.VarDesc.VarType.FP16:
-            self.dtype = 'float16'
-        elif dtype == core.VarDesc.VarType.INT32:
-            self.dtype = 'int32'
-        elif dtype == core.VarDesc.VarType.UINT8:
-            self.dtype = 'uint8'
-        else:
-            raise ValueError("dtype must be any of [int32, float32, int64, "
-                             "float64, uint8]")
+        self.dtype = convert_dtype(dtype)
+        self._reset()
 
+    def _reset(self):
         self.data = []
-        self.lod = []
-
-        for i in six.moves.range(lod_level):
-            self.lod.append([])
+        self.lod = [[] for _ in six.moves.range(self.lod_level)]
 
     def feed(self, data):
         self._feed_impl_(data, self.lod, self.lod_level)
@@ -88,15 +91,52 @@ class DataToLoDTensorConverter(object):
                     raise ValueError(
                         "Reshape error. What is defined in data layer is {}, but receive {}"
                         .format(self.shape, arr.shape))
-            #else:
-            #    self._check_shape(arr.shape)
         t = core.LoDTensor()
         t.set(arr, self.place)
         if self.lod_level > 0:
             t.set_recursive_sequence_lengths(self.lod)
+        self._reset()
         return t
 
 
+class BatchedTensorProvider(object):
+    def __init__(self, feed_list, place, batch_size, generator, drop_last):
+        self.place = place
+        self.batch_size = batch_size
+        self.generator = generator
+        self.converters = []
+        self.drop_last = drop_last
+
+        for var in feed_list:
+            assert var.lod_level == 0, "lod_level must be 0"
+            self.converters.append(
+                DataToLoDTensorConverter(
+                    place=self.place,
+                    lod_level=0,
+                    shape=var.shape,
+                    dtype=var.dtype))
+
+    def _done(self):
+        return [c.done() for c in self.converters]
+
+    def __call__(self):
+        idx = 0
+        for each_sample in self.generator():
+            for each_slot, each_converter in six.moves.zip(each_sample,
+                                                           self.converters):
+                each_converter.data.append(each_slot)
+
+            idx += 1
+            if idx == self.batch_size:
+                idx = 0
+                yield self._done()
+
+        if not self.drop_last and idx > 0:
+            yield self._done()
+        else:
+            [c._reset() for c in self.converters]
+
+
 class DataFeeder(object):
     """
     DataFeeder converts the data that returned by a reader into a data

python/paddle/fluid/executor.py

@@ -564,6 +564,10 @@ class Executor(object):
 
         if feed is None:
             feed = {}
+        elif isinstance(feed, (list, tuple)):
+            assert len(feed) == 1, "Not compiled with data parallel"
+            feed = feed[0]
+
         if not isinstance(feed, dict):
             raise TypeError(
                 "feed requires dict as its Parameter. But you passed in %s" %

python/paddle/fluid/framework.py

@@ -26,6 +26,7 @@ import six
 
 import numpy as np
 import subprocess
+import multiprocessing
 
 from .. import compat as cpt
 from .proto import framework_pb2
@@ -63,6 +64,9 @@ __all__ = [
     'default_main_program',
     'program_guard',
     'name_scope',
+    'cuda_places',
+    'cpu_places',
+    'cuda_pinned_places',
 ]
 
 EMPTY_VAR_NAME = core.kEmptyVarName()
@@ -87,6 +91,87 @@ def _current_expected_place():
     return _imperative_current_expected_place_
 
 
+def _cpu_num():
+    return int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+
+
+def cuda_places(device_ids=None):
+    '''
+    Create a list of :code:`fluid.CUDAPlace` objects.
+
+    If :code:`device_ids` is None, environment variable of
+    :code:`FLAGS_selected_gpus` would be checked first. If
+    :code:`FLAGS_selected_gpus=0,1,2`, the returned list would
+    be [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
+    If :code:`FLAGS_selected_gpus` is not set, all visible
+    gpu places would be returned.  
+
+    If :code:`device_ids` is not None, it should be the device
+    ids of gpus. For example, if :code:`device_ids=[0,1,2]`, 
+    the returned list would be 
+    [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
+    
+    Args: 
+        device_ids (None|list(int)|tuple(int)): gpu device id list.
+
+    Returns:
+        out (list(fluid.CUDAPlace)): gpu place list.
+    '''
+    assert core.is_compiled_with_cuda(), \
+        "Not compiled with CUDA"
+    if device_ids is None:
+        gpus_env = os.getenv("FLAGS_selected_gpus")
+        if gpus_env:
+            device_ids = [int(s) for s in gpus_env.split(",")]
+        else:
+            device_ids = six.moves.range(core.get_cuda_device_count())
+    elif not isinstance(device_ids, (list, tuple)):
+        device_ids = [device_ids]
+    return [core.CUDAPlace(dev_id) for dev_id in device_ids]
+
+
+def cpu_places(device_count=None):
+    '''
+    Create a list of :code:`fluid.CPUPlace` objects.
+    
+    If :code:`device_count` is None, the device count would
+    be determined by environment variable :code:`CPU_NUM`. 
+    If :code:`CPU_NUM` is not set, the device count would
+    be determined by :code:`multiprocessing.cpu_count()`. 
+
+    Args:
+        device_count (None|int): device number.
+
+    Returns:
+        out (list(fluid.CPUPlace)): cpu place list.
+    '''
+    if device_count is None:
+        device_count = _cpu_num()
+    return [core.CPUPlace()] * device_count
+
+
+def cuda_pinned_places(device_count=None):
+    '''
+    Create a list of :code:`fluid.CUDAPinnedPlace` objects.
+
+    If :code:`device_count` is None, the device count would
+    be determined by environment variable :code:`CPU_NUM`. 
+    If :code:`CPU_NUM` is not set, the device count would
+    be determined by :code:`multiprocessing.cpu_count()`. 
+
+    Args:
+        device_count (None|int): device number.
+
+    Returns:
+        out (list(fluid.CUDAPinnedPlace)): cuda pinned place list.
+    '''
+    assert core.is_compiled_with_cuda(), \
+        "Not compiled with CUDA"
+    if device_count is None:
+        device_count = _cpu_num()
+    return [core.cuda_pinned_places()] * device_count
+
+
 class NameScope(object):
     def __init__(self, name="", parent=None):
         self._children = dict()
@@ -318,8 +403,8 @@ class Variable(object):
                 self._ivar = core.VarBase(
                     name, dtype if dtype else core.VarDesc.VarType.FP32,
                     list(shape) if shape else [],
-                    _current_expected_place(), True
-                    if persistable else False, stop_gradient)
+                    _current_expected_place(), stop_gradient, True
+                    if persistable else False)
             if persistable:
                 _imperative_tracer().trace_var(name, self)
         else:
@@ -1559,12 +1644,15 @@ class Block(object):
                 name=v.name)
             self.vars[new_p.name] = new_p
 
-    def _clone_variable(self, var):
+    def _clone_variable(self, var, force_persistable=True):
         """
         Clone a variable into current block.
 
         Args:
             var: the variable to be cloned.
+            force_persistable(bool): True means setting the result variable to being persistable.
+                                     False means setting the persistable the same with that of input var.
+                                     default: True.
 
         Returns:
             Variable: the new  variable cloned from 'var' in current block.
@@ -1584,7 +1672,7 @@ class Block(object):
                 shape=var.shape,
                 dtype=var.dtype,
                 type=var.type,
-                persistable=True,
+                persistable=True if force_persistable else var.persistable,
                 is_data=var.is_data)
         else:
             ret_var = self.create_var(
@@ -1593,7 +1681,7 @@ class Block(object):
                 dtype=var.dtype,
                 type=var.type,
                 lod_level=var.lod_level,
-                persistable=True,
+                persistable=True if force_persistable else var.persistable,
                 is_data=var.is_data)
         return ret_var
 

python/paddle/fluid/imperative/base.py

@@ -44,7 +44,7 @@ def guard(place=None):
                     yield
 
 
-def to_variable(value, block=None):
+def to_variable(value, block=None, name=None):
     if isinstance(value, np.ndarray):
         assert enabled(), "to_variable could only be called in imperative mode"
 
@@ -53,7 +53,7 @@ def to_variable(value, block=None):
         py_var = framework.Variable(
             block,
             type=core.VarDesc.VarType.LOD_TENSOR,
-            name=None,
+            name=name,
             shape=value.shape,
             dtype=value.dtype)
         var = py_var._ivar.value()

python/paddle/fluid/imperative/layer_object_helper.py

@@ -105,6 +105,7 @@ class LayerObjectHelper(LayerHelperBase):
 
         Returns dtype of the input
         """
+        inputs_in = inputs_in if (inputs_in is not None) else []
         inputs = self._multiple_input(inputs_in)
         dtype = None
         for each in inputs:

python/paddle/fluid/imperative/layers.py

@@ -17,10 +17,12 @@ import contextlib
 import sys
 import numpy as np
 import collections
+import six
 from .. import unique_name
 from paddle.fluid import core
 from .layer_object_helper import LayerObjectHelper
 from paddle.fluid import framework
+from ..param_attr import ParamAttr
 
 __all__ = ['Layer', 'PyLayer']
 
@@ -72,6 +74,10 @@ class Layer(core.Layer):
 
         Returns created parameter Variable.
         """
+        if isinstance(attr, ParamAttr) and (attr.name is not None):
+            attr.name = ".".join([self._full_name, attr.name])
+        elif isinstance(attr, six.string_types):
+            attr = ".".join([self._full_name, attr])
         return self._helper.create_parameter(attr, shape, dtype, is_bias,
                                              default_initializer)
 
@@ -164,6 +170,7 @@ class Layer(core.Layer):
             the sublayer passed in.
         """
         assert isinstance(sublayer, core.Layer)
+
         self._sub_layers[name] = sublayer
         return sublayer
 

python/paddle/fluid/imperative/nn.py

@@ -20,10 +20,12 @@ from .. import core
 from ..layers import utils
 from . import layers
 from ..framework import Variable, OpProtoHolder
+from ..layers import layer_function_generator
 from ..param_attr import ParamAttr
 from ..initializer import Normal, Constant
-
-__all__ = ['Conv2D', 'Pool2D', 'FC', 'BatchNorm', 'Embedding', 'GRUUnit']
+__all__ = [
+    'Conv2D', 'Pool2D', 'FC', 'BatchNorm', 'Embedding', 'GRUUnit', 'LayerNorm'
+]
 
 
 class Conv2D(layers.Layer):
@@ -438,7 +440,6 @@ class Embedding(layers.Layer):
         self._size = size
         self._is_sparse = is_sparse
         self._is_distributed = is_distributed
-
         self._padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (
             size[0] + padding_idx)
 
@@ -471,6 +472,131 @@ class Embedding(layers.Layer):
         return out
 
 
+class LayerNorm(layers.Layer):
+    def __init__(self,
+                 name_scope,
+                 scale=True,
+                 shift=True,
+                 begin_norm_axis=1,
+                 epsilon=1e-05,
+                 param_attr=None,
+                 bias_attr=None,
+                 act=None):
+        """
+        ${comment}
+
+        The formula is as follows:
+
+        ..  math::
+
+            \\mu & = \\frac{1}{H}\\sum_{i=1}^{H} a_i
+
+            \\sigma & = \\sqrt{\\frac{1}{H}\sum_{i=1}^{H}(a_i - \\mu)^2}
+
+            h & = f(\\frac{g}{\\sigma}(a - \\mu) + b)
+
+        * :math:`a`: the vector representation of the summed inputs to the neurons
+        in that layer.
+
+        * :math:`H`: the number of hidden units in a layers
+
+        * :math:`g`: the trainable scale parameter.
+
+        * :math:`b`: the trainable bias parameter.
+
+        Args:
+            input(Variable): The input tensor variable.
+            scale(bool): Whether to learn the adaptive gain :math:`g` after
+                normalization. Default True.
+            shift(bool): Whether to learn the adaptive bias :math:`b` after
+                normalization. Default True.
+            begin_norm_axis(int): The normalization will be performed along
+                dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`.
+                Default 1.
+            epsilon(float): The small value added to the variance to prevent
+                division by zero. Default 1e-05.
+            param_attr(ParamAttr|None): The parameter attribute for the learnable
+                gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
+                omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
+                a default :code:`ParamAttr` would be added as scale. The
+                :attr:`param_attr` is initialized as 1 if it is added. Default None.
+            bias_attr(ParamAttr|None): The parameter attribute for the learnable
+                bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
+                omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
+                a default :code:`ParamAttr` would be added as bias. The
+                :attr:`bias_attr` is initialized as 0 if it is added. Default None.
+            act(str): Activation to be applied to the output of layer normalizaiton.
+                      Default None.
+        Returns:
+            ${y_comment}
+
+        Examples:
+
+            >>> data = fluid.layers.data(name='data', shape=[3, 32, 32],
+            >>>                          dtype='float32')
+            >>> x = fluid.layers.layer_norm(input=data, begin_norm_axis=1)
+        """
+
+        super(LayerNorm, self).__init__(name_scope)
+        self._scale = scale
+        self._shift = shift
+        self._begin_norm_axis = begin_norm_axis
+        self._epsilon = epsilon
+        self._param_attr = param_attr
+        self._bias_attr = bias_attr
+        self._act = act
+
+    def _build_once(self, input):
+        self._dtype = self._helper.input_dtype(input)
+        input_shape = input.shape
+        param_shape = [
+            reduce(lambda x, y: x * y, input_shape[self._begin_norm_axis:])
+        ]
+        if self._scale:
+            self._scale_w = self.create_parameter(
+                attr=self._param_attr,
+                shape=param_shape,
+                dtype=self._dtype,
+                default_initializer=Constant(1.0))
+        if self._shift:
+            assert self._bias_attr is not False
+            self._bias_w = self.create_parameter(
+                attr=self._bias_attr,
+                shape=param_shape,
+                dtype=self._dtype,
+                is_bias=True)
+
+    def forward(self, input):
+        inputs = dict()
+        inputs['X'] = input
+        if self._scale:
+            inputs['Scale'] = self._scale_w
+        if self._shift:
+            inputs['Bias'] = self._bias_w
+        # create output
+        mean_out = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype, stop_gradient=True)
+        variance_out = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype, stop_gradient=True)
+        layer_norm_out = self._helper.create_variable_for_type_inference(
+            self._dtype)
+
+        self._helper.append_op(
+            type="layer_norm",
+            inputs=inputs,
+            outputs={
+                "Y": layer_norm_out,
+                "Mean": mean_out,
+                "Variance": variance_out,
+            },
+            attrs={
+                "epsilon": self._epsilon,
+                "begin_norm_axis": self._begin_norm_axis
+            })
+
+        return self._helper.append_activation(layer_norm_out)
+
+
 class GRUUnit(layers.Layer):
     """
     **GRU unit layer**

python/paddle/fluid/io.py

@@ -26,12 +26,14 @@ from paddle.fluid import layers
 from paddle.fluid.executor import Executor
 from paddle.fluid.evaluator import Evaluator
 from paddle.fluid.framework import Program, Parameter, default_main_program, default_startup_program, Variable, program_guard
+from . import reader
+from .reader import *
 from . import core
 
 __all__ = [
     'save_vars', 'save_params', 'save_persistables', 'load_vars', 'load_params',
     'load_persistables', 'save_inference_model', 'load_inference_model'
-]
+] + reader.__all__
 
 
 def is_parameter(var):

python/paddle/fluid/layer_helper_base.py

@@ -268,11 +268,9 @@ class LayerHelperBase(object):
         """
         # Deepcopy the attr so that parameters can be shared in program
         attr = copy.deepcopy(attr)
-        if attr is None:
-            attr = ParamAttr._to_attr(attr)
+        attr = ParamAttr._to_attr(attr)
         if not attr:
             return None
-
         assert isinstance(attr, ParamAttr)
         suffix = 'b' if is_bias else 'w'
         if attr.name is None:

python/paddle/fluid/layers/io.py

@@ -563,22 +563,26 @@ def _py_reader(capacity,
 
     def start_provide_thread(func):
         def __provider_thread__():
-            for tensors in func():
-                array = core.LoDTensorArray()
-                for item in tensors:
-                    if not isinstance(item, core.LoDTensor):
-                        tmp = core.LoDTensor()
-                        tmp.set(item, core.CPUPlace())
-                        item = tmp
-
-                    array.append(item)
-
-                if reader.exited:
-                    break
-                feed_queue.push(array)
-                if reader.exited:
-                    break
-            feed_queue.close()
+            try:
+                for tensors in func():
+                    array = core.LoDTensorArray()
+                    for item in tensors:
+                        if not isinstance(item, core.LoDTensor):
+                            tmp = core.LoDTensor()
+                            tmp.set(item, core.CPUPlace())
+                            item = tmp
+
+                        array.append(item)
+
+                    if reader.exited:
+                        break
+                    feed_queue.push(array)
+                    if reader.exited:
+                        break
+                feed_queue.close()
+            except Exception as ex:
+                feed_queue.close()
+                raise ex
 
         reader.thread = threading.Thread(target=__provider_thread__)
         reader.thread.daemon = True
@@ -628,6 +632,9 @@ def _py_reader(capacity,
     reader.reset = __reset__
     reader.decorate_tensor_provider = __set_tensor_provider__
     reader.decorate_paddle_reader = __set_paddle_reader__
+
+    reader.decorate_batch_generator = __set_tensor_provider__
+    reader.decorate_sample_list_generator = __set_paddle_reader__
     reader.start = __start__
 
     return reader
@@ -692,6 +699,11 @@ def py_reader(capacity,
         >>>             exe.run(fetch_list=[loss.name])
         >>>     except fluid.core.EOFException:
         >>>         reader.reset()
+        >>>
+        >>> ...
+        >>>
+        >>> fluid.io.save_inference_model(dirname='./model', feeded_var_names=[img, label],
+        >>>     target_vars=[loss], executor=fluid.Executor(fluid.CUDAPlace(0)))  
 
         2. When training and testing are both performed, two different
         :code:`py_reader` should be created with different names, e.g.:

python/paddle/fluid/layers/nn.py

@@ -189,6 +189,7 @@ __all__ = [
     'huber_loss',
     'tree_conv',
     'npair_loss',
+    'fsp_matrix',
 ]
 
 kIgnoreIndex = -100
@@ -6205,7 +6206,8 @@ def one_hot(input, depth):
         type="one_hot",
         inputs={'X': input},
         attrs={'depth': depth},
-        outputs={'Out': one_hot_out})
+        outputs={'Out': one_hot_out},
+        stop_gradient=True)
     return one_hot_out
 
 
@@ -9768,7 +9770,7 @@ def affine_channel(x,
                 'Bias': bias},
         attrs={"data_layout": data_layout},
         outputs={"Out": out})
-    return helper.append_activation(pre_activation)
+    return helper.append_activation(out)
 
 
 def similarity_focus(input, axis, indexes, name=None):
@@ -10790,3 +10792,46 @@ def npair_loss(anchor, positive, labels, l2_reg=0.002):
     celoss = reduce_mean(cross_entropy)
 
     return l2loss + celoss
+
+
+def fsp_matrix(x, y):
+    """
+
+    **FSP matrix op**
+
+    This op is used to calculate the flow of solution procedure (FSP) matrix of two feature maps.
+    Given feature map x with shape [x_channel, h, w] and feature map y with shape
+    [y_channel, h, w], we can get the fsp matrix of x and y in two steps:
+
+    1. reshape x into matrix with shape [x_channel, h * w] and reshape and
+       transpose y into matrix with shape [h * w, y_channel].
+    2. multiply x and y to get fsp matrix with shape [x_channel, y_channel].
+
+    The output is a batch of fsp matrices.
+
+    Args:
+
+        x (Variable): A feature map with shape [batch_size, x_channel, height, width].
+        y (Variable): A feature map with shape [batch_size, y_channel, height, width].
+                      The y_channel can be different with the x_channel of Input(X)
+                      while the other dimensions must be the same with Input(X)'s.
+
+    Returns:
+
+        fsp matrix (Variable): The output of FSP op with shape [batch_size, x_channel, y_channel].
+        The x_channel is the channel of x and the y_channel is the channel of y.
+
+    Examples:
+
+        .. code-block:: python
+
+            feature_map_0 = fluid.layers.conv2d(x)
+            feature_map_1 = fluid.layers.conv2d(feature_map_0)
+            loss = fluid.layers.fsp_matrix(feature_map_0, feature_map_1)
+
+    """
+    helper = LayerHelper('fsp_matrix', **locals())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype(
+        input_param_name='x'))
+    helper.append_op(type='fsp', inputs={'X': x, 'Y': y}, outputs={'Out': out})
+    return out

python/paddle/fluid/layers/tensor.py

@@ -25,10 +25,26 @@ from .layer_function_generator import templatedoc
 import numpy
 
 __all__ = [
-    'create_tensor', 'create_parameter', 'create_global_var', 'cast',
-    'tensor_array_to_tensor', 'concat', 'sums', 'assign',
-    'fill_constant_batch_size_like', 'fill_constant', 'argmin', 'argmax',
-    'argsort', 'ones', 'zeros', 'reverse', 'has_inf', 'has_nan', 'isfinite'
+    'create_tensor',
+    'create_parameter',
+    'create_global_var',
+    'cast',
+    'tensor_array_to_tensor',
+    'concat',
+    'sums',
+    'assign',
+    'fill_constant_batch_size_like',
+    'fill_constant',
+    'argmin',
+    'argmax',
+    'argsort',
+    'ones',
+    'zeros',
+    'reverse',
+    'has_inf',
+    'has_nan',
+    'isfinite',
+    'range',
 ]
 
 
@@ -764,3 +780,50 @@ def isfinite(x):
     out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(type="isfinite", inputs={"X": x}, outputs={"Out": out})
     return out
+
+
+def range(start, end, step, dtype):
+    """
+    Return evenly spaced values within a given interval.
+
+    Values are generated within the half-open interval [start, stop) (in other words,
+    the interval including start but excluding stop).
+
+    args:
+        start(int|float|Variable): Start of interval. The interval includes this value.
+        end(int|float|Variable): End of interval. The interval does not include this
+                                 value, except in some cases where step is not an integer
+                                 and floating point round-off affects the length of out. 
+        step(int|float|Variable): Spacing between values. For any output out, this is the
+                                  distance between two adjacent values, out[i+1] - out[i].
+                                  The default step size is 1.
+        dtype(string): 'float32'|'int32'|..., the data type of the output tensor.
+
+    returns:
+        Evenly spaced values within a given interval.
+
+    examples:
+
+        .. code-block:: python
+
+             data = fluid.layers.range(0, 10, 2, 'int32')
+
+    """
+    helper = LayerHelper("range", **locals())
+
+    if not isinstance(start, Variable):
+        start = fill_constant([1], dtype, start)
+    if not isinstance(end, Variable):
+        end = fill_constant([1], dtype, end)
+    if not isinstance(step, Variable):
+        step = fill_constant([1], dtype, step)
+
+    out = helper.create_variable_for_type_inference(dtype=start.dtype)
+
+    helper.append_op(
+        type='range',
+        inputs={'Start': start,
+                'End': end,
+                'Step': step},
+        outputs={'Out': [out]})
+    return out

python/paddle/fluid/optimizer.py

@@ -70,6 +70,10 @@ class Optimizer(object):
         # {accum_name : { paramter_name : accumulator_for_parameter, ...}, ...}
         self._accumulators = defaultdict(lambda: dict())
         self.helper = None
+        self._opti_name_list = []
+
+    def get_opti_var_name_list(self):
+        return self._opti_name_list
 
     def _create_global_learning_rate(self):
         lr = self._global_learning_rate()
@@ -166,8 +170,13 @@ class Optimizer(object):
         if shape == None:
             shape = param.shape
         assert isinstance(self.helper, LayerHelper)
+
+        var_name = param.name + "_" + name
+        var_name = unique_name.generate(var_name)
+        self._opti_name_list.append(var_name)
+
         var = self.helper.create_global_variable(
-            name=unique_name.generate(name),
+            name=var_name,
             persistable=True,
             dtype=dtype or param.dtype,
             type=param.type,
@@ -388,13 +397,14 @@ class Optimizer(object):
             for param in parameters:
                 if not param.trainable:
                     continue
-                # create gradient variable
-                grad_var = Variable(
-                    block=loss.block,
-                    name=param._ivar._grad_name(),
-                    stop_gradient=True,
-                    ivar=param._ivar._grad_ivar())
-                params_grads.append((param, grad_var))
+                if param._ivar._grad_ivar() is not None:
+                    # create gradient variable
+                    grad_var = Variable(
+                        block=loss.block,
+                        name=param._ivar._grad_name(),
+                        stop_gradient=True,
+                        ivar=param._ivar._grad_ivar())
+                    params_grads.append((param, grad_var))
             with program_guard(framework.default_main_program(),
                                framework.default_startup_program()):
                 optimize_ops = self._create_optimization_pass(params_grads)

python/paddle/fluid/parallel_executor.py

@@ -99,7 +99,8 @@ class ParallelExecutor(object):
         build_strategy.num_trainers = num_trainers
         build_strategy.trainer_id = trainer_id
 
-        self._places = compiler.get_available_places(use_cuda)
+        self._places = framework.cuda_places(
+        ) if use_cuda else framework.cpu_places()
         self._scope = scope if scope is not None else executor.global_scope()
 
         main_program = main_program if main_program is not None \

python/paddle/fluid/reader.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from . import core
+import six
+import threading
+from .framework import Program, Variable, program_guard, default_main_program, default_startup_program
+from .executor import global_scope
+from .data_feeder import DataFeeder, BatchedTensorProvider
+from .layers.io import monkey_patch_reader_methods, _copy_reader_var_, double_buffer
+from .unique_name import UniqueNameGenerator
+
+__all__ = ['PyReader']
+
+
+def _convert_places(places):
+    if not isinstance(places, (list, tuple)):
+        places = [places]
+
+    ret = []
+    for p in places:
+        if not isinstance(p, core.Place):
+            tmp = core.Place()
+            tmp.set_place(p)
+            p = tmp
+
+        ret.append(p)
+    return ret
+
+
+class PyReader(object):
+    """
+    Create a reader object for data feeding in Python. 
+    Data would be prefetched using Python thread and be pushed
+    into a queue asynchronously. Data in the queue would be extracted 
+    automatically when `Executor.run(...)` is called.
+
+    Args:  
+        feed_list (list(Variable)|tuple(Variable)): feed variable list.
+            The variables should be created by :code:`fluid.layers.data()`. 
+        capacity (int): capacity of the queue maintained in PyReader object. 
+        use_double_buffer (bool): whether to use double_buffer_reader to 
+            speed up data feeding. 
+        iterable (bool): whether the created reader object is iterable.   
+
+    Returns:
+        reader (Reader): the created reader object.
+
+    Examples:
+        1. If iterable = False, the created PyReader object is almost the
+           same as :code:`fluid.layers.py_reader()`. Operators would be 
+           inserted into the program. User should call :code:`start()` 
+           before each epoch and catch :code:`fluid.core.EOFException`
+           thrown by :code:`Executor.run()` when epoch ends. Once the 
+           exception is caught, user should call :code:`reset()` to reset 
+           the reader manually.
+
+        .. code-block:: python
+            
+            image = fluid.layers.data(
+                        name='image', shape=[784], dtype='float32')
+            label = fluid.layers.data(
+                        name='label', shape=[1], dtype='int64')
+            
+            reader = fluid.io.PyReader(feed_list=[image, label], 
+                        capacity=4, iterable=False)
+            reader.decorate_sample_list_generator(user_defined_reader)
+            ... # definition of network is omitted
+            executor.run(fluid.default_main_program())
+            for _ in range(EPOCH_NUM):
+                reader.start()
+                while True:
+                    try:
+                        executor.run(feed=None, ...)
+                    except fluid.core.EOFException:
+                        reader.reset()
+                        break
+                    
+        2. If iterable=True, the created PyReader object is decoupled with
+           the program. No operator would be inserted into the program. 
+           In this case, the created reader is a Python generator, which 
+           is iterable. User should feed the data yielded from PyReader 
+           object into :code:`Executor.run(feed=...)`.  
+
+        .. code-block:: python
+
+            image = fluid.layers.data(
+                        name='image', shape=[784], dtype='float32')
+            label = fluid.layers.data(
+                        name='label', shape=[1], dtype='int64')
+
+            reader = fluid.io.PyReader(feed_list=[image, label], 
+                        capacity=4, iterable=True)
+            reader.decorate_sample_list_generator(user_defined_reader, 
+                        places=fluid.cuda_places())
+            ... # definition of network is omitted
+            executor.run(fluid.default_main_program())
+            for _ in range(EPOCH_NUM):
+                for data in reader():
+                    executor.run(feed=data, ...)
+    """
+
+    unique_name_generator = UniqueNameGenerator()
+
+    def __init__(self,
+                 feed_list,
+                 capacity,
+                 use_double_buffer=True,
+                 iterable=False):
+        self._tensor_reader = None
+        self._thread = None
+        self._iterable = iterable
+        self._use_double_buffer = use_double_buffer
+        self._capacity = capacity
+        self._feed_list = feed_list
+        if not self._iterable:
+            self._init_non_iterable()
+
+    def _init_iterable(self, places):
+        self._var_names = [v.name for v in self._feed_list]
+        self._places = _convert_places(places)
+        self._queue = core.init_lod_tensor_blocking_queue(core.Variable(),
+                                                          self._capacity)
+        self._reader = core.create_py_reader(
+            self.queue, self._var_names, self._places, self._use_double_buffer)
+
+    def _init_non_iterable(self):
+        lod_levels = []
+        dtypes = []
+        shape_concat = []
+        ranks = []
+        shapes = []
+
+        for feed_data in self._feed_list:
+            dtypes.append(feed_data.dtype)
+            shape_concat.extend(feed_data.shape)
+            ranks.append(len(feed_data.shape))
+            shapes.append(feed_data.shape)
+            lod_levels.append(feed_data.lod_level)
+
+        queue_name = PyReader.unique_name_generator('lod_tensor_blocking_queue')
+        reader_name = PyReader.unique_name_generator('create_py_reader')
+        double_buffer_name = PyReader.unique_name_generator('double_buffer')
+
+        var = global_scope().var(queue_name)
+        self._queue = core.init_lod_tensor_blocking_queue(var, self._capacity)
+
+        startup_blk = default_startup_program().current_block()
+        startup_var = startup_blk.create_var(name=reader_name)
+
+        startup_blk.append_op(
+            type='create_py_reader',
+            inputs={'blocking_queue': [queue_name]},
+            outputs={'Out': [startup_var]},
+            attrs={
+                'shape_concat': shape_concat,
+                'lod_levels': lod_levels,
+                'ranks': ranks
+            })
+
+        startup_var.desc.set_dtypes(dtypes)
+        startup_var.persistable = True
+
+        main_prog_var = _copy_reader_var_(
+            default_main_program().current_block(), startup_var)
+
+        main_prog_var.stop_gradient = True
+        main_prog_var.persistable = True
+
+        reader = monkey_patch_reader_methods(main_prog_var)
+        if self._use_double_buffer:
+            double_buffer_reader = double_buffer(
+                reader, name=double_buffer_name)
+            # we return a double buffer reader. However, the reset method comes from
+            # py_reader.
+            double_buffer_reader.reset = reader.reset
+            reader = double_buffer_reader
+
+        self._reader = reader
+
+        default_main_program().current_block().append_op(
+            type='read',
+            inputs={'Reader': [self._reader]},
+            outputs={'Out': self._feed_list})
+
+    @property
+    def queue(self):
+        return self._queue
+
+    @property
+    def iterable(self):
+        return self._iterable
+
+    def __call__(self):
+        assert self.iterable, "PyReader is not iterable"
+        assert self._tensor_reader is not None, \
+            "Data source of PyReader has not set yet"
+
+        class Iterator(object):
+            def __init__(self, reader):
+                self._reader = reader._reader
+                self._reset = reader._reset
+
+            def __iter__(self):
+                return self
+
+            def __next__(self):
+                return self.next()
+
+            def next(self):
+                ret = self._reader.read_next()
+                if ret:
+                    return ret
+                else:
+                    self._reset()
+                    raise StopIteration
+
+        self._start()
+        return Iterator(self)
+
+    def _reset(self):
+        self._reader.reset()
+        self._thread.join()
+
+    def start(self):
+        '''
+        Start the data feeding thread. 
+        Can only call when the reader object is not iterable.  
+        '''
+        assert not self._iterable, "start() cannot be called when PyReader is iterable"
+        self._start()
+
+    def reset(self):
+        '''
+        Reset the reader object when :code:`fluid.core.EOFException` raises. 
+        Can only call when the reader object is not iterable.
+        '''
+        assert not self._iterable, "reset() cannot be called when PyReader is iterable"
+        self._reset()
+
+    def _start(self):
+        def __thread_main__():
+            try:
+                for tensors in self._tensor_reader():
+                    array = core.LoDTensorArray()
+                    for item in tensors:
+                        if not isinstance(item, core.LoDTensor):
+                            tmp = core.LoDTensor()
+                            tmp.set(item, core.CPUPlace())
+                            item = tmp
+
+                        array.append(item)
+
+                    if not self._queue.push(array):
+                        break
+
+                self._queue.close()
+            except Exception as ex:
+                self._queue.close()
+                raise ex
+
+        self._thread = threading.Thread(target=__thread_main__)
+        self._thread.daemon = True
+        self._thread.start()
+
+    def decorate_sample_generator(self,
+                                  sample_generator,
+                                  batch_size,
+                                  drop_last=True,
+                                  places=None):
+        '''
+        Set the data source of the PyReader object.
+        
+        The provided :code:`sample_generator` should be a Python generator,
+        which yields numpy.ndarray typed data of each sample.
+
+        :code:`places` must be set when the PyReader object is iterable.
+
+        If all inputs have no lods, this method is faster than 
+        :code:`decorate_sample_list_generator(paddle.batch(sample_generator, ...))` .
+
+        Args:
+            sample_generator (generator): Python generator that yields
+                numpy.ndarray-typed sample data.
+            batch_size (int): batch size. Must be larger than 0.
+            drop_last (bool): Whether to drop the last batch when sample number
+                is less than batch_size. 
+            places (None|list(CUDAPlace)|list(CPUPlace)): place list. Must
+                be provided when PyReader is iterable.
+        '''
+        assert batch_size > 0, "batch_size must be larger than 0"
+        has_lod = False
+        for f in self._feed_list:
+            if f.lod_level != 0:
+                has_lod = True
+                break
+
+        if has_lod:
+            self.decorate_sample_list_generator(
+                paddle.batch(
+                    sample_generator,
+                    batch_size=batch_size,
+                    drop_last=drop_last),
+                places=places)
+        else:
+            reader = BatchedTensorProvider(
+                feed_list=self._feed_list,
+                place=core.CPUPlace(),
+                batch_size=batch_size,
+                generator=sample_generator,
+                drop_last=drop_last)
+            self.decorate_batch_generator(reader, places=places)
+
+    def decorate_sample_list_generator(self, reader, places=None):
+        '''
+        Set the data source of the PyReader object. 
+
+        The provided :code:`reader` should be a Python generator,
+        which yields list(numpy.ndarray) typed batched data. 
+        
+        :code:`places` must be set when the PyReader object is iterable.
+
+        Args:
+            reader (generator): Python generator that yields 
+                list(numpy.ndarray)-typed batched data. 
+            places (None|list(CUDAPlace)|list(CPUPlace)): place list. Must
+                be provided when PyReader is iterable.
+        '''
+        assert self._tensor_reader is None, \
+            "Cannot reset the data source of PyReader"
+        with program_guard(Program(), Program()):
+            feeder = DataFeeder(
+                feed_list=self._feed_list, place=core.CPUPlace())
+            paddle_reader = feeder.decorate_reader(reader, multi_devices=False)
+
+        def __tensor_reader_impl__():
+            for slots in paddle_reader():
+                yield [slots[var.name] for var in self._feed_list]
+
+        self.decorate_batch_generator(__tensor_reader_impl__, places)
+
+    def decorate_batch_generator(self, reader, places=None):
+        '''
+        Set the data source of the PyReader object.
+
+        The provided :code:`reader` should be a Python generator,
+        which yields numpy.ndarray-typed or LoDTensor-typed batched data.
+
+        :code:`places` must be set when the PyReader object is iterable.
+
+        Args:
+            reader (generator): Python generator that yields LoDTensor-typed
+                batched data.
+            places (None|list(CUDAPlace)|list(CPUPlace)): place list. Must
+                be provided when PyReader is iterable.
+        '''
+        assert self._tensor_reader is None, \
+            "Cannot reset the data source of PyReader"
+        self._tensor_reader = reader
+        if self._iterable:
+            assert places is not None, "Places cannot be None when py_reader is iterable"
+            self._init_iterable(places)

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -105,7 +105,7 @@ if(WITH_DISTRIBUTE)
         # set_tests_properties(test_dist_transformer PROPERTIES TIMEOUT 1000)
         set_tests_properties(test_dist_ctr test_dist_mnist test_dist_mnist_batch_merge test_dist_save_load test_dist_se_resnext test_dist_simnet_bow test_dist_text_classification test_dist_train test_dist_word2vec PROPERTIES RUN_SERIAL TRUE)
     endif(NOT APPLE)
-    py_test_modules(test_dist_transpiler MODULES test_dist_transpiler)
+    # py_test_modules(test_dist_transpiler MODULES test_dist_transpiler)
 endif()
 py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SERIAL)
 py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL)
@@ -118,8 +118,8 @@ if(NOT APPLE)
     py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)
 endif()
 if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-    # change the timeout from 600 to 1200, because in debug mode, this test need more time.
-    set_tests_properties(test_parallel_executor_seresnext PROPERTIES TIMEOUT 1200)
+    # change the timeout from 600 to 2200, because in debug mode, this test need more time.
+    set_tests_properties(test_parallel_executor_seresnext PROPERTIES TIMEOUT 2200)
 endif()
 
 if (WITH_NGRAPH)

python/paddle/fluid/tests/unittests/dist_ctr_reader.py

@@ -20,7 +20,7 @@ logging.basicConfig()
 logger = logging.getLogger("paddle")
 logger.setLevel(logging.INFO)
 
-DATA_URL = "http://paddle-ctr-data.cdn.bcebos.com/avazu_ctr_data.tgz"
+DATA_URL = "http://paddle-ctr-data.bj.bcebos.com/avazu_ctr_data.tgz"
 DATA_MD5 = "c11df99fbd14e53cd4bfa6567344b26e"
 """
 avazu_ctr_data/train.txt

python/paddle/fluid/tests/unittests/test_decoupled_py_reader.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import paddle.fluid as fluid
+import numpy as np
+import time
+import six
+import unittest
+
+EPOCH_NUM = 60
+BATCH_SIZE = 32
+CLASS_NUM = 10
+
+
+def random_reader():
+    np.random.seed(1)
+    for i in range(BATCH_SIZE * 40):
+        image = np.random.random([784])
+        label = np.random.random_integers(low=0, high=CLASS_NUM - 1)
+        yield image, label
+
+
+def simple_fc_net(places, use_legacy_py_reader, use_double_buffer):
+    startup_prog = fluid.Program()
+    main_prog = fluid.Program()
+    startup_prog.random_seed = 1
+    main_prog.random_seed = 1
+
+    with fluid.unique_name.guard():
+        with fluid.program_guard(main_prog, startup_prog):
+            image = fluid.layers.data(
+                name='image', shape=[784], dtype='float32')
+            label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+            py_reader = fluid.io.PyReader(
+                feed_list=[image, label],
+                capacity=4,
+                iterable=not use_legacy_py_reader,
+                use_double_buffer=use_double_buffer)
+            hidden = image
+            for hidden_size in [10, 20, 30]:
+                hidden = fluid.layers.fc(
+                    hidden,
+                    size=hidden_size,
+                    act='tanh',
+                    bias_attr=fluid.ParamAttr(
+                        initializer=fluid.initializer.Constant(value=1.0)))
+
+            predict_label = fluid.layers.fc(hidden,
+                                            size=CLASS_NUM,
+                                            act='softmax')
+            loss = fluid.layers.mean(
+                fluid.layers.cross_entropy(
+                    input=predict_label, label=label))
+
+            optimizer = fluid.optimizer.Adam()
+            optimizer.minimize(loss)
+    return startup_prog, main_prog, py_reader, loss
+
+
+class TestBase(unittest.TestCase):
+    def run_main(self, use_legacy_py_reader, with_data_parallel, places,
+                 use_double_buffer):
+        scope = fluid.Scope()
+        with fluid.scope_guard(scope):
+            startup_prog, main_prog, py_reader, loss = simple_fc_net(
+                places, use_legacy_py_reader, use_double_buffer)
+
+            reader = paddle.batch(random_reader, batch_size=BATCH_SIZE)
+
+            ps = places if use_double_buffer else fluid.cpu_places(len(places))
+
+            py_reader.decorate_sample_list_generator(
+                reader, places=ps if py_reader.iterable else None)
+
+            exe = fluid.Executor(place=places[0])
+            exe.run(startup_prog)
+
+            prog = fluid.CompiledProgram(main_prog)
+            if with_data_parallel:
+                prog = prog.with_data_parallel(
+                    loss_name=loss.name, places=places)
+
+            step = 0
+            step_list = []
+            loss_list = []
+            start_t = time.time()
+            if not py_reader.iterable:
+                for _ in six.moves.range(EPOCH_NUM):
+                    step = 0
+                    py_reader.start()
+                    while True:
+                        try:
+                            L, = exe.run(program=prog,
+                                         fetch_list=[loss],
+                                         use_program_cache=True)
+                            loss_list.append(np.mean(L))
+                            step += 1
+                        except fluid.core.EOFException:
+                            py_reader.reset()
+                            break
+                    step_list.append(step)
+            else:
+                for _ in six.moves.range(EPOCH_NUM):
+                    step = 0
+                    for d in py_reader():
+                        assert len(d) == len(places)
+                        for i, item in enumerate(d):
+                            image = item['image']
+                            label = item['label']
+                            assert image.shape() == [BATCH_SIZE, 784]
+                            assert label.shape() == [BATCH_SIZE, 1]
+                            assert image._place()._equals(ps[i])
+                            assert label._place()._equals(ps[i])
+                        L, = exe.run(program=prog,
+                                     feed=d,
+                                     fetch_list=[loss],
+                                     use_program_cache=True)
+                        loss_list.append(np.mean(L))
+                        step += 1
+                    step_list.append(step)
+            end_t = time.time()
+            ret = {
+                "time": end_t - start_t,
+                "step": step_list,
+                "loss": np.array(loss_list)
+            }
+            return ret
+
+    def prepare_places(self, with_data_parallel, with_cpu=True, with_gpu=True):
+        places = []
+        if with_cpu:
+            places.append([fluid.CPUPlace()])
+            if with_data_parallel:
+                places.append([fluid.CPUPlace()] * 2)
+
+        if with_gpu and fluid.core.is_compiled_with_cuda():
+            tmp = fluid.cuda_places()
+            assert len(tmp) > 0, "no gpu detected"
+            if with_data_parallel:
+                places.append(tmp)
+            places.append([tmp[0]])
+        return places
+
+    def test_main(self):
+        for with_data_parallel in [True, False]:
+            for p in self.prepare_places(with_data_parallel):
+                for use_double_buffer in [False, True]:
+                    results = []
+                    for use_legacy_py_reader in [False, True]:
+                        ret = self.run_main(
+                            use_legacy_py_reader=use_legacy_py_reader,
+                            with_data_parallel=with_data_parallel,
+                            places=p,
+                            use_double_buffer=use_double_buffer)
+                        results.append(ret)
+                    if not use_double_buffer:
+                        diff = np.max(
+                            np.abs(results[0]['loss'] - results[1]['loss']))
+                        self.assertLess(diff, 1e-3)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_dist_transformer.py

@@ -21,7 +21,7 @@ from test_dist_base import TestDistBase
 
 
 def download_files():
-    url_prefix = 'http://paddle-unittest-data.cdn.bcebos.com/dist_transformer/'
+    url_prefix = 'http://paddle-unittest-data.bj.bcebos.com/dist_transformer/'
     vocab_url = url_prefix + 'vocab.bpe.32000'
     vocab_md5 = 'a86d345ca6e27f6591d0dccb1b9be853'
     paddle.dataset.common.download(vocab_url, 'test_dist_transformer',

python/paddle/fluid/tests/unittests/test_fsp_op.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+def fsp_matrix(a, b):
+    batch = a.shape[0]
+    a_channel = a.shape[1]
+    b_channel = b.shape[1]
+    h = a.shape[2]
+    w = a.shape[3]
+    a_t = a.transpose([0, 2, 3, 1])
+    a_t = a_t.reshape([batch, h * w, a_channel])
+    b_t = b.transpose([0, 2, 3, 1]).reshape([batch, h * w, b_channel])
+    a_r = a_t.repeat(
+        b_channel, axis=1).reshape(
+            [batch, h * w, b_channel, a_channel]).transpose([0, 1, 3, 2])
+    b_r = b_t.repeat(
+        a_channel, axis=1).reshape([batch, h * w, a_channel, b_channel])
+    return np.mean(a_r * b_r, axis=1)
+
+
+class TestFSPOp(OpTest):
+    def setUp(self):
+        self.op_type = "fsp"
+        self.initTestCase()
+
+        feature_map_0 = np.random.uniform(0, 10, self.a_shape).astype('float32')
+        feature_map_1 = np.random.uniform(0, 10, self.b_shape).astype('float32')
+
+        self.inputs = {'X': feature_map_0, 'Y': feature_map_1}
+        self.outputs = {'Out': fsp_matrix(feature_map_0, feature_map_1)}
+
+    def initTestCase(self):
+        self.a_shape = (2, 16, 32, 31)
+        self.b_shape = (2, 28, 32, 31)
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.05)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_imperative_deepcf.py

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import numpy as np
+import random
+import sys
+
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+from test_imperative_base import new_program_scope
+from paddle.fluid.imperative.base import to_variable
+
+NUM_USERS = 100
+NUM_ITEMS = 1000
+
+BATCH_SIZE = 32
+NUM_BATCHES = 2
+
+
+class MLP(fluid.imperative.Layer):
+    def __init__(self, name_scope):
+        super(MLP, self).__init__(name_scope)
+        self._user_latent = fluid.imperative.FC(self.full_name(), 256)
+        self._item_latent = fluid.imperative.FC(self.full_name(), 256)
+
+        self._user_layers = []
+        self._item_layers = []
+        self._hid_sizes = [128, 64]
+        for i in range(len(self._hid_sizes)):
+            self._user_layers.append(
+                self.add_sublayer(
+                    'user_layer_%d' % i,
+                    fluid.imperative.FC(
+                        self.full_name(), self._hid_sizes[i], act='relu')))
+            self._item_layers.append(
+                self.add_sublayer(
+                    'item_layer_%d' % i,
+                    fluid.imperative.FC(
+                        self.full_name(), self._hid_sizes[i], act='relu')))
+
+    def forward(self, users, items):
+        users = self._user_latent(users)
+        items = self._item_latent(items)
+
+        for ul, il in zip(self._user_layers, self._item_layers):
+            users = ul(users)
+            items = il(items)
+        return fluid.layers.elementwise_mul(users, items)
+
+
+class DMF(fluid.imperative.Layer):
+    def __init__(self, name_scope):
+        super(DMF, self).__init__(name_scope)
+        self._user_latent = fluid.imperative.FC(self.full_name(), 256)
+        self._item_latent = fluid.imperative.FC(self.full_name(), 256)
+        self._match_layers = []
+        self._hid_sizes = [128, 64]
+        for i in range(len(self._hid_sizes)):
+            self._match_layers.append(
+                self.add_sublayer(
+                    'match_layer_%d' % i,
+                    fluid.imperative.FC(
+                        self.full_name(), self._hid_sizes[i], act='relu')))
+        self._mat
+
+    def forward(self, users, items):
+        users = self._user_latent(users)
+        items = self._item_latent(items)
+        match_vec = fluid.layers.concat(
+            [users, items], axis=len(users.shape) - 1)
+        for l in self._match_layers:
+            match_vec = l(match_vec)
+        return match_vec
+
+
+class DeepCF(fluid.imperative.Layer):
+    def __init__(self, name_scope):
+        super(DeepCF, self).__init__(name_scope)
+
+        self._user_emb = fluid.imperative.Embedding(self.full_name(),
+                                                    [NUM_USERS, 256])
+        self._item_emb = fluid.imperative.Embedding(self.full_name(),
+                                                    [NUM_ITEMS, 256])
+
+        self._mlp = MLP(self.full_name())
+        self._dmf = DMF(self.full_name())
+        self._match_fc = fluid.imperative.FC(self.full_name(), 1, act='sigmoid')
+
+    def forward(self, users, items):
+        users_emb = self._user_emb(users)
+        items_emb = self._item_emb(items)
+
+        mlp_predictive = self._mlp(users_emb, items_emb)
+        dmf_predictive = self._dmf(users_emb, items_emb)
+        predictive = fluid.layers.concat(
+            [mlp_predictive, dmf_predictive],
+            axis=len(mlp_predictive.shape) - 1)
+        prediction = self._match_fc(predictive)
+        return prediction
+
+
+def get_data():
+    user_ids = []
+    item_ids = []
+    labels = []
+    for uid in range(NUM_USERS):
+        for iid in range(NUM_ITEMS):
+            # 10% positive
+            label = float(random.randint(1, 10) == 1)
+            user_ids.append(uid)
+            item_ids.append(iid)
+            labels.append(label)
+    indices = np.arange(NUM_USERS * NUM_ITEMS)
+    np.random.shuffle(indices)
+    users_np = np.array(user_ids, dtype=np.int64)[indices]
+    items_np = np.array(item_ids, dtype=np.int64)[indices]
+    labels_np = np.array(labels, dtype=np.float32)[indices]
+    return np.expand_dims(users_np, -1), \
+           np.expand_dims(items_np, -1), \
+           np.expand_dims(labels_np, -1)
+
+
+class TestImperativeDeepCF(unittest.TestCase):
+    def test_gan_float32(self):
+        seed = 90
+        users_np, items_np, labels_np = get_data()
+
+        startup = fluid.Program()
+        startup.random_seed = seed
+        main = fluid.Program()
+        main.random_seed = seed
+
+        scope = fluid.core.Scope()
+        with new_program_scope(main=main, startup=startup, scope=scope):
+            users = fluid.layers.data('users', [1], dtype='int64')
+            items = fluid.layers.data('items', [1], dtype='int64')
+            labels = fluid.layers.data('labels', [1], dtype='float32')
+
+            deepcf = DeepCF('deepcf')
+            prediction = deepcf(users, items)
+            loss = fluid.layers.reduce_sum(
+                fluid.layers.log_loss(prediction, labels))
+            adam = fluid.optimizer.AdamOptimizer(0.01)
+            adam.minimize(loss)
+
+            exe = fluid.Executor(fluid.CPUPlace(
+            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
+            exe.run(startup)
+            for slice in range(0, BATCH_SIZE * NUM_BATCHES, BATCH_SIZE):
+                static_loss = exe.run(
+                    main,
+                    feed={
+                        users.name: users_np[slice:slice + BATCH_SIZE],
+                        items.name: items_np[slice:slice + BATCH_SIZE],
+                        labels.name: labels_np[slice:slice + BATCH_SIZE]
+                    },
+                    fetch_list=[loss])[0]
+                sys.stderr.write('static loss %s\n' % static_loss)
+
+        with fluid.imperative.guard():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+
+            deepcf = DeepCF('deepcf')
+            for slice in range(0, BATCH_SIZE * NUM_BATCHES, BATCH_SIZE):
+                prediction = deepcf(
+                    to_variable(users_np[slice:slice + BATCH_SIZE]),
+                    to_variable(items_np[slice:slice + BATCH_SIZE]))
+                loss = fluid.layers.reduce_sum(
+                    fluid.layers.log_loss(prediction,
+                                          to_variable(labels_np[slice:slice +
+                                                                BATCH_SIZE])))
+                loss._backward()
+                adam = fluid.optimizer.AdamOptimizer(0.01)
+                adam.minimize(loss)
+                deepcf.clear_gradients()
+                dy_loss = loss._numpy()
+
+        self.assertEqual(static_loss, dy_loss)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_imperative_gan.py

@@ -51,7 +51,7 @@ class Generator(fluid.imperative.Layer):
         return self._fc3(x)
 
 
-class TestImperativeMnist(unittest.TestCase):
+class TestImperativeGAN(unittest.TestCase):
     def test_gan_float32(self):
         seed = 90
 

python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py

@@ -59,7 +59,7 @@ class SimpleLSTMRNN(fluid.imperative.Layer):
                 dtype="float32",
                 default_initializer=fluid.initializer.UniformInitializer(
                     low=-self._init_scale, high=self._init_scale))
-            self.weight_1_arr.append(weight_1)
+            self.weight_1_arr.append(self.add_parameter('w_%d' % i, weight_1))
             bias_1 = self.create_parameter(
                 attr=fluid.ParamAttr(
                     initializer=fluid.initializer.UniformInitializer(
@@ -67,7 +67,7 @@ class SimpleLSTMRNN(fluid.imperative.Layer):
                 shape=[self._hidden_size * 4],
                 dtype="float32",
                 default_initializer=fluid.initializer.Constant(0.0))
-            self.bias_arr.append(bias_1)
+            self.bias_arr.append(self.add_parameter('b_%d' % i, bias_1))
 
     def forward(self, input_embedding, init_hidden=None, init_cell=None):
         self.cell_array = []
@@ -242,7 +242,7 @@ class TestImperativePtbRnn(unittest.TestCase):
             dy_loss = None
             last_hidden = None
             last_cell = None
-            batch_num = 50
+            batch_num = 200
 
             for i in range(batch_num):
                 x_data = np.arange(12).reshape(4, 3).astype('int64')
@@ -264,8 +264,10 @@ class TestImperativePtbRnn(unittest.TestCase):
                         dy_param_init[param.name] = param._numpy()
                 dy_loss._backward()
                 sgd.minimize(dy_loss)
-                for param in ptb_model.parameters():
-                    dy_param_updated[param.name] = param._numpy()
+                ptb_model.clear_gradients()
+                if i == batch_num - 1:
+                    for param in ptb_model.parameters():
+                        dy_param_updated[param.name] = param._numpy()
 
         with new_program_scope():
             fluid.default_startup_program().random_seed = seed
@@ -323,25 +325,28 @@ class TestImperativePtbRnn(unittest.TestCase):
                               },
                               fetch_list=fetch_list)
                 static_loss_value = out[0]
-                static_last_cell_value = out[1]
-                static_last_hidden_value = out[2]
-                for k in range(3, len(out)):
-                    static_param_updated[static_param_name_list[k - 3]] = out[k]
+                static_last_hidden_value = out[1]
+                static_last_cell_value = out[2]
 
+                if i == batch_num - 1:
+                    for k in range(3, len(out)):
+                        static_param_updated[static_param_name_list[k -
+                                                                    3]] = out[k]
+
+        self.assertTrue(np.allclose(static_loss_value, dy_loss._numpy()))
+        self.assertTrue(np.allclose(static_last_cell_value, last_cell._numpy()))
+        self.assertTrue(
+            np.allclose(static_last_hidden_value, last_hidden._numpy()))
+        for key, value in six.iteritems(static_param_init):
+            # print("static_init name: {}, value {}".format(key, value))
+            # print("dy_init name: {}, value {}".format(key, dy_param_init[key]))
+            self.assertTrue(np.allclose(value, dy_param_init[key], atol=1e-5))
+        for key, value in six.iteritems(static_param_updated):
+            # print("static name: {}, value {}".format(key, value))
+            # print("dy name: {}, value {}".format(key, dy_param_updated[key]))
             self.assertTrue(
-                np.allclose(static_loss_value.all(), dy_loss._numpy().all()))
-            self.assertTrue(
-                np.allclose(static_last_cell_value.all(),
-                            last_cell._numpy().all()))
-            self.assertTrue(
-                np.allclose(static_last_hidden_value.all(),
-                            last_hidden._numpy().all()))
-            for key, value in six.iteritems(static_param_init):
-                self.assertTrue(
-                    np.allclose(value.all(), dy_param_init[key].all()))
-            for key, value in six.iteritems(static_param_updated):
-                self.assertTrue(
-                    np.allclose(value.all(), dy_param_updated[key].all()))
+                np.allclose(
+                    value, dy_param_updated[key], atol=1e-5))
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/unittests/test_imperative_transformer.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import paddle.fluid as fluid
+from paddle.fluid.imperative import Embedding, LayerNorm, FC, to_variable, Layer, guard
+from test_imperative_base import new_program_scope
+from paddle.fluid import core
+import numpy as np
+import six
+np.set_printoptions(suppress=True)
+
+
+# Copy from models
+class TrainTaskConfig(object):
+    # support both CPU and GPU now.
+    use_gpu = True
+    # the epoch number to train.
+    pass_num = 30
+    # the number of sequences contained in a mini-batch.
+    # deprecated, set batch_size in args.
+    batch_size = 32
+    # the hyper parameters for Adam optimizer.
+    # This static learning_rate will be multiplied to the LearningRateScheduler
+    # derived learning rate the to get the final learning rate.
+    learning_rate = 2.0
+    beta1 = 0.9
+    beta2 = 0.997
+    eps = 1e-9
+    # the parameters for learning rate scheduling.
+    warmup_steps = 8000
+    # the weight used to mix up the ground-truth distribution and the fixed
+    # uniform distribution in label smoothing when training.
+    # Set this as zero if label smoothing is not wanted.
+    label_smooth_eps = 0.1
+    # the directory for saving trained models.
+    model_dir = "trained_models"
+    # the directory for saving checkpoints.
+    ckpt_dir = "trained_ckpts"
+    # the directory for loading checkpoint.
+    # If provided, continue training from the checkpoint.
+    ckpt_path = None
+    # the parameter to initialize the learning rate scheduler.
+    # It should be provided if use checkpoints, since the checkpoint doesn't
+    # include the training step counter currently.
+    start_step = 0
+    # the frequency to save trained models.
+    save_freq = 10000
+
+
+class InferTaskConfig(object):
+    use_gpu = True
+    # the number of examples in one run for sequence generation.
+    batch_size = 10
+    # the parameters for beam search.
+    beam_size = 5
+    max_out_len = 256
+    # the number of decoded sentences to output.
+    n_best = 1
+    # the flags indicating whether to output the special tokens.
+    output_bos = False
+    output_eos = False
+    output_unk = True
+    # the directory for loading the trained model.
+    model_path = "trained_models/pass_1.infer.model"
+
+
+class ModelHyperParams(object):
+    # These following five vocabularies related configurations will be set
+    # automatically according to the passed vocabulary path and special tokens.
+    # size of source word dictionary.
+    src_vocab_size = 10000
+    # size of target word dictionay
+    trg_vocab_size = 10000
+    # index for <bos> token
+    bos_idx = 0
+    # index for <eos> token
+    eos_idx = 1
+    # index for <unk> token
+    unk_idx = 2
+    # max length of sequences deciding the size of position encoding table.
+    max_length = 4
+    # the dimension for word embeddings, which is also the last dimension of
+    # the input and output of multi-head attention, position-wise feed-forward
+    # networks, encoder and decoder.
+    d_model = 512
+    # size of the hidden layer in position-wise feed-forward networks.
+    d_inner_hid = 2048
+    # the dimension that keys are projected to for dot-product attention.
+    d_key = 64
+    # the dimension that values are projected to for dot-product attention.
+    d_value = 64
+    # number of head used in multi-head attention.
+    n_head = 8
+    # number of sub-layers to be stacked in the encoder and decoder.
+    n_layer = 6
+    # dropout rates of different modules.
+    prepostprocess_dropout = 0.1
+    attention_dropout = 0.1
+    relu_dropout = 0.1
+    # to process before each sub-layer
+    preprocess_cmd = "n"  # layer normalization
+    # to process after each sub-layer
+    postprocess_cmd = "da"  # dropout + residual connection
+    # random seed used in dropout for CE.
+    dropout_seed = 1
+    # the flag indicating whether to share embedding and softmax weights.
+    # vocabularies in source and target should be same for weight sharing.
+    weight_sharing = True
+
+
+def merge_cfg_from_list(cfg_list, g_cfgs):
+    """
+    Set the above global configurations using the cfg_list.
+    """
+    assert len(cfg_list) % 2 == 0
+    for key, value in zip(cfg_list[0::2], cfg_list[1::2]):
+        for g_cfg in g_cfgs:
+            if hasattr(g_cfg, key):
+                try:
+                    value = eval(value)
+                except Exception:  # for file path
+                    pass
+                setattr(g_cfg, key, value)
+                break
+
+
+def position_encoding_init(n_position, d_pos_vec):
+    """
+    Generate the initial values for the sinusoid position encoding table.
+    """
+    channels = d_pos_vec
+    position = np.arange(n_position)
+    num_timescales = channels // 2
+    log_timescale_increment = (np.log(float(1e4) / float(1)) /
+                               (num_timescales - 1))
+    inv_timescales = np.exp(np.arange(
+        num_timescales)) * -log_timescale_increment
+    scaled_time = np.expand_dims(position, 1) * np.expand_dims(inv_timescales,
+                                                               0)
+    signal = np.concatenate([np.sin(scaled_time), np.cos(scaled_time)], axis=1)
+    signal = np.pad(signal, [[0, 0], [0, np.mod(channels, 2)]], 'constant')
+    position_enc = signal
+    return position_enc.astype("float32")
+
+
+def create_data(is_static=False):
+    if is_static:
+        return [
+            src_word_np, src_pos_np, src_slf_attn_bias_np, trg_word_np,
+            trg_pos_np, trg_slf_attn_bias_np, trg_src_attn_bias_np, lbl_word_np,
+            lbl_weight_np
+        ]
+    else:
+        enc_inputs = [
+            to_variable(src_word_np), to_variable(src_pos_np),
+            to_variable(src_slf_attn_bias_np)
+        ]
+        dec_inputs = [
+            to_variable(trg_word_np), to_variable(trg_pos_np),
+            to_variable(trg_slf_attn_bias_np), to_variable(trg_src_attn_bias_np)
+        ]
+        label = to_variable(lbl_word_np)
+        weight = to_variable(lbl_weight_np)
+        return enc_inputs, dec_inputs, label, weight
+
+
+def create_feed_dict_list(data, init=False):
+    if init:
+        data_input_names = encoder_data_input_fields + \
+                           decoder_data_input_fields[:-1] + label_data_input_fields + pos_enc_param_names
+    else:
+        data_input_names = encoder_data_input_fields + \
+                           decoder_data_input_fields[:-1] + label_data_input_fields
+    feed_dict_list = dict()
+    for i in range(len(data_input_names)):
+        feed_dict_list[data_input_names[i]] = data[i]
+    return feed_dict_list
+
+
+def make_all_inputs(input_fields):
+    """
+    Define the input data layers for the transformer model.
+    """
+    inputs = []
+    for input_field in input_fields:
+        input_var = fluid.layers.data(
+            name=input_field,
+            shape=input_descs[input_field][0],
+            dtype=input_descs[input_field][1],
+            lod_level=input_descs[input_field][2]
+            if len(input_descs[input_field]) == 3 else 0,
+            append_batch_size=False)
+        inputs.append(input_var)
+    return inputs
+
+
+# The placeholder for batch_size in compile time. Must be -1 currently to be
+# consistent with some ops' infer-shape output in compile time, such as the
+# sequence_expand op used in beamsearch decoder.
+batch_size = 32
+# The placeholder for squence length in compile time.
+seq_len = ModelHyperParams.max_length
+# Here list the data shapes and data types of all inputs.
+# The shapes here act as placeholder and are set to pass the infer-shape in
+# compile time.
+input_descs = {
+    # The actual data shape of src_word is:
+    # [batch_size, max_src_len_in_batch, 1]
+    "src_word": [(batch_size, seq_len, 1), "int64", 2],
+    # The actual data shape of src_pos is:
+    # [batch_size, max_src_len_in_batch, 1]
+    "src_pos": [(batch_size, seq_len, 1), "int64"],
+    # This input is used to remove attention weights on paddings in the
+    # encoder.
+    # The actual data shape of src_slf_attn_bias is:
+    # [batch_size, n_head, max_src_len_in_batch, max_src_len_in_batch]
+    "src_slf_attn_bias": [(batch_size, ModelHyperParams.n_head, seq_len,
+                           seq_len), "float32"],
+    # The actual data shape of trg_word is:
+    # [batch_size, max_trg_len_in_batch, 1]
+    "trg_word": [(batch_size, seq_len, 1), "int64",
+                 2],  # lod_level is only used in fast decoder.
+    # The actual data shape of trg_pos is:
+    # [batch_size, max_trg_len_in_batch, 1]
+    "trg_pos": [(batch_size, seq_len, 1), "int64"],
+    # This input is used to remove attention weights on paddings and
+    # subsequent words in the decoder.
+    # The actual data shape of trg_slf_attn_bias is:
+    # [batch_size, n_head, max_trg_len_in_batch, max_trg_len_in_batch]
+    "trg_slf_attn_bias": [(batch_size, ModelHyperParams.n_head, seq_len,
+                           seq_len), "float32"],
+    # This input is used to remove attention weights on paddings of the source
+    # input in the encoder-decoder attention.
+    # The actual data shape of trg_src_attn_bias is:
+    # [batch_size, n_head, max_trg_len_in_batch, max_src_len_in_batch]
+    "trg_src_attn_bias": [(batch_size, ModelHyperParams.n_head, seq_len,
+                           seq_len), "float32"],
+    # This input is used in independent decoder program for inference.
+    # The actual data shape of enc_output is:
+    # [batch_size, max_src_len_in_batch, d_model]
+    "enc_output": [(batch_size, seq_len, ModelHyperParams.d_model), "float32"],
+    # The actual data shape of label_word is:
+    # [batch_size * max_trg_len_in_batch, 1]
+    "lbl_word": [(batch_size * seq_len, 1), "int64"],
+    # This input is used to mask out the loss of paddding tokens.
+    # The actual data shape of label_weight is:
+    # [batch_size * max_trg_len_in_batch, 1]
+    "lbl_weight": [(batch_size * seq_len, 1), "float32"],
+    # This input is used in beam-search decoder.
+    "init_score": [(batch_size, 1), "float32", 2],
+    # This input is used in beam-search decoder for the first gather
+    # (cell states updation)
+    "init_idx": [(batch_size, ), "int32"],
+}
+
+# Names of word embedding table which might be reused for weight sharing.
+word_emb_param_names = (
+    "src_word_emb_table",
+    "trg_word_emb_table", )
+# Names of position encoding table which will be initialized externally.
+pos_enc_param_names = (
+    "src_pos_enc_table",
+    "trg_pos_enc_table", )
+# separated inputs for different usages.
+encoder_data_input_fields = (
+    "src_word",
+    "src_pos",
+    "src_slf_attn_bias", )
+decoder_data_input_fields = (
+    "trg_word",
+    "trg_pos",
+    "trg_slf_attn_bias",
+    "trg_src_attn_bias",
+    "enc_output", )
+label_data_input_fields = (
+    "lbl_word",
+    "lbl_weight", )
+# In fast decoder, trg_pos (only containing the current time step) is generated
+# by ops and trg_slf_attn_bias is not needed.
+fast_decoder_data_input_fields = (
+    "trg_word",
+    "init_score",
+    "init_idx",
+    "trg_src_attn_bias", )
+# if we use py_reader
+use_py_reader = False
+
+# if we run sync mode
+sync = False
+
+# how many batches we use
+batch_num = 2
+
+np.random.seed = 1
+src_word_np = np.random.randint(
+    1,
+    ModelHyperParams.src_vocab_size - 1,
+    size=(batch_size, seq_len, 1),
+    dtype='int64')
+src_pos_np = np.random.randint(
+    1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
+src_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
+                                       seq_len, seq_len).astype('float32')
+
+trg_word_np = np.random.randint(
+    1,
+    ModelHyperParams.src_vocab_size - 1,
+    size=(batch_size, seq_len, 1),
+    dtype='int64')
+trg_pos_np = np.random.randint(
+    1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
+trg_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
+                                       seq_len, seq_len).astype('float32')
+trg_src_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
+                                       seq_len, seq_len).astype('float32')
+
+lbl_word_np = np.random.randint(
+    1,
+    ModelHyperParams.src_vocab_size - 1,
+    size=(batch_size * seq_len, 1),
+    dtype='int64')
+lbl_weight_np = np.random.randn(batch_size * seq_len, 1).astype('float32')
+
+# np.random.seed = 1
+# src_word_np = np.arange(0, 10).reshape([batch_size, seq_len, 1]).astype('int64')
+# src_pos_np = np.random.randint(
+#     1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
+# src_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
+#                                        seq_len, seq_len).astype('float32')
+#
+# trg_word_np =  np.arange(0, 10).reshape([batch_size, seq_len, 1]).astype('int64')
+# trg_pos_np = np.random.randint(
+#     1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
+# trg_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
+#                                        seq_len, seq_len).astype('float32')
+# trg_src_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
+#                                        seq_len, seq_len).astype('float32')
+#
+# lbl_word_np =  np.arange(0, 10).reshape([batch_size * seq_len, 1]).astype('int64')
+# lbl_weight_np = np.random.randn(batch_size * seq_len, 1).astype('float32')
+#
+pos_inp1 = position_encoding_init(ModelHyperParams.max_length,
+                                  ModelHyperParams.d_model)
+pos_inp2 = position_encoding_init(ModelHyperParams.max_length,
+                                  ModelHyperParams.d_model)
+
+
+class PrePostProcessLayer(Layer):
+    def __init__(self, name_scope, process_cmd, shape_len=None):
+        super(PrePostProcessLayer, self).__init__(name_scope)
+        for cmd in process_cmd:
+            if cmd == "n":
+                self._layer_norm = LayerNorm(
+                    name_scope=self.full_name(),
+                    begin_norm_axis=shape_len - 1,
+                    param_attr=fluid.ParamAttr(
+                        initializer=fluid.initializer.Constant(1.)),
+                    bias_attr=fluid.ParamAttr(
+                        initializer=fluid.initializer.Constant(0.)))
+
+    def forward(self, prev_out, out, process_cmd, dropout_rate=0.):
+        for cmd in process_cmd:
+            if cmd == "a":  # add residual connection
+                out = out + prev_out if prev_out else out
+            elif cmd == "n":  # add layer normalization
+                out = self._layer_norm(out)
+            elif cmd == "d":  # add dropout
+                if dropout_rate:
+                    out = fluid.layers.dropout(
+                        out,
+                        dropout_prob=dropout_rate,
+                        seed=ModelHyperParams.dropout_seed,
+                        is_test=False)
+        return out
+
+
+class PositionwiseFeedForwardLayer(Layer):
+    def __init__(self, name_scope, d_inner_hid, d_hid, dropout_rate):
+        super(PositionwiseFeedForwardLayer, self).__init__(name_scope)
+        self._i2h = FC(name_scope=self.full_name(),
+                       size=d_inner_hid,
+                       num_flatten_dims=2,
+                       act="relu")
+        self._h2o = FC(name_scope=self.full_name(),
+                       size=d_hid,
+                       num_flatten_dims=2)
+        self._dropout_rate = dropout_rate
+
+    def forward(self, x):
+        hidden = self._i2h(x)
+        if self._dropout_rate:
+            hidden = fluid.layers.dropout(
+                hidden,
+                dropout_prob=self._dropout_rate,
+                seed=ModelHyperParams.dropout_seed,
+                is_test=False)
+        out = self._h2o(hidden)
+        return out
+
+
+class MultiHeadAttentionLayer(Layer):
+    def __init__(self,
+                 name_scope,
+                 d_key,
+                 d_value,
+                 d_model,
+                 n_head=1,
+                 dropout_rate=0.,
+                 cache=None,
+                 gather_idx=None,
+                 static_kv=False):
+        super(MultiHeadAttentionLayer, self).__init__(name_scope)
+        self._n_head = n_head
+        self._d_key = d_key
+        self._d_value = d_value
+        self._d_model = d_model
+        self._dropout_rate = dropout_rate
+        self._q_fc = FC(name_scope=self.full_name(),
+                        size=d_key * n_head,
+                        bias_attr=False,
+                        num_flatten_dims=2)
+        self._k_fc = FC(name_scope=self.full_name(),
+                        size=d_key * n_head,
+                        bias_attr=False,
+                        num_flatten_dims=2)
+        self._v_fc = FC(name_scope=self.full_name(),
+                        size=d_value * n_head,
+                        bias_attr=False,
+                        num_flatten_dims=2)
+        self._proj_fc = FC(name_scope=self.full_name(),
+                           size=self._d_model,
+                           bias_attr=False,
+                           num_flatten_dims=2)
+
+    def forward(self, queries, keys, values, attn_bias):
+        # compute q ,k ,v
+        keys = queries if keys is None else keys
+        values = keys if values is None else values
+
+        q = self._q_fc(queries)
+        k = self._k_fc(keys)
+        v = self._v_fc(values)
+
+        # split head
+        reshaped_q = fluid.layers.reshape(
+            x=q, shape=[0, 0, self._n_head, self._d_key], inplace=False)
+        transpose_q = fluid.layers.transpose(x=reshaped_q, perm=[0, 2, 1, 3])
+        reshaped_k = fluid.layers.reshape(
+            x=k, shape=[0, 0, self._n_head, self._d_key], inplace=False)
+        transpose_k = fluid.layers.transpose(x=reshaped_k, perm=[0, 2, 1, 3])
+        reshaped_v = fluid.layers.reshape(
+            x=v, shape=[0, 0, self._n_head, self._d_value], inplace=False)
+        transpose_v = fluid.layers.transpose(x=reshaped_v, perm=[0, 2, 1, 3])
+
+        #scale dot product attention
+        product = fluid.layers.matmul(
+            x=transpose_q,
+            y=transpose_k,
+            transpose_y=True,
+            alpha=self._d_model**-0.5)
+        if attn_bias:
+            product += attn_bias
+        weights = fluid.layers.softmax(product)
+        if self._dropout_rate:
+            weights_droped = fluid.layers.dropout(
+                weights,
+                dropout_prob=self._dropout_rate,
+                seed=ModelHyperParams.dropout_seed,
+                is_test=False)
+            out = fluid.layers.matmul(weights_droped, transpose_v)
+        else:
+            out = fluid.layers.matmul(weights, transpose_v)
+
+        # combine heads
+        if len(out.shape) != 4:
+            raise ValueError("Input(x) should be a 4-D Tensor.")
+        trans_x = fluid.layers.transpose(out, perm=[0, 2, 1, 3])
+        final_out = fluid.layers.reshape(
+            x=trans_x,
+            shape=[0, 0, trans_x.shape[2] * trans_x.shape[3]],
+            inplace=False)
+
+        # fc to output
+        proj_out = self._proj_fc(final_out)
+        return proj_out
+
+
+class EncoderSubLayer(Layer):
+    def __init__(self,
+                 name_scope,
+                 n_head,
+                 d_key,
+                 d_value,
+                 d_model,
+                 d_inner_hid,
+                 prepostprocess_dropout,
+                 attention_dropout,
+                 relu_dropout,
+                 preprocess_cmd="n",
+                 postprocess_cmd="da"):
+
+        super(EncoderSubLayer, self).__init__(name_scope)
+        self._preprocess_cmd = preprocess_cmd
+        self._postprocess_cmd = postprocess_cmd
+        self._prepostprocess_dropout = prepostprocess_dropout
+
+        self._preprocess_layer = PrePostProcessLayer(self.full_name(),
+                                                     self._preprocess_cmd, 3)
+        self._multihead_attention_layer = MultiHeadAttentionLayer(
+            self.full_name(), d_key, d_value, d_model, n_head,
+            attention_dropout)
+        self._postprocess_layer = PrePostProcessLayer(
+            self.full_name(), self._postprocess_cmd, None)
+        self._preprocess_layer2 = PrePostProcessLayer(self.full_name(),
+                                                      self._preprocess_cmd, 3)
+        self._positionwise_feed_forward = PositionwiseFeedForwardLayer(
+            self.full_name(), d_inner_hid, d_model, relu_dropout)
+        self._postprocess_layer2 = PrePostProcessLayer(
+            self.full_name(), self._postprocess_cmd, None)
+
+    def forward(self, enc_input, attn_bias):
+        pre_process_multihead = self._preprocess_layer(
+            None, enc_input, self._preprocess_cmd, self._prepostprocess_dropout)
+        attn_output = self._multihead_attention_layer(pre_process_multihead,
+                                                      None, None, attn_bias)
+        attn_output = self._postprocess_layer(enc_input, attn_output,
+                                              self._postprocess_cmd,
+                                              self._prepostprocess_dropout)
+        pre_process2_output = self._preprocess_layer2(
+            None, attn_output, self._preprocess_cmd,
+            self._prepostprocess_dropout)
+        ffd_output = self._positionwise_feed_forward(pre_process2_output)
+        return self._postprocess_layer2(attn_output, ffd_output,
+                                        self._postprocess_cmd,
+                                        self._prepostprocess_dropout)
+
+
+class EncoderLayer(Layer):
+    def __init__(self,
+                 name_scope,
+                 n_layer,
+                 n_head,
+                 d_key,
+                 d_value,
+                 d_model,
+                 d_inner_hid,
+                 prepostprocess_dropout,
+                 attention_dropout,
+                 relu_dropout,
+                 preprocess_cmd="n",
+                 postprocess_cmd="da"):
+
+        super(EncoderLayer, self).__init__(name_scope)
+        self._preprocess_cmd = preprocess_cmd
+        self._encoder_sublayers = list()
+        self._prepostprocess_dropout = prepostprocess_dropout
+        self._n_layer = n_layer
+        self._preprocess_layer = PrePostProcessLayer(self.full_name(),
+                                                     self._preprocess_cmd, 3)
+        for i in range(n_layer):
+            self._encoder_sublayers.append(
+                self.add_sublayer(
+                    'esl_%d' % i,
+                    EncoderSubLayer(
+                        self.full_name(), n_head, d_key, d_value, d_model,
+                        d_inner_hid, prepostprocess_dropout, attention_dropout,
+                        relu_dropout, preprocess_cmd, postprocess_cmd)))
+
+    def forward(self, enc_input, attn_bias):
+        for i in range(self._n_layer):
+            enc_output = self._encoder_sublayers[i](enc_input, attn_bias)
+            enc_input = enc_output
+
+        return self._preprocess_layer(None, enc_output, self._preprocess_cmd,
+                                      self._prepostprocess_dropout)
+
+
+class PrepareEncoderDecoderLayer(Layer):
+    def __init__(self,
+                 name_scope,
+                 src_vocab_size,
+                 src_emb_dim,
+                 src_max_len,
+                 dropout_rate,
+                 word_emb_param_name=None,
+                 pos_enc_param_name=None):
+        super(PrepareEncoderDecoderLayer, self).__init__(name_scope)
+        self._src_max_len = src_max_len
+        self._src_emb_dim = src_emb_dim
+        self._src_vocab_size = src_vocab_size
+        self._dropout_rate = dropout_rate
+        self._input_emb = Embedding(
+            name_scope=self.full_name(),
+            size=[src_vocab_size, src_emb_dim],
+            padding_idx=0,
+            param_attr=fluid.ParamAttr(
+                name=word_emb_param_name,
+                initializer=fluid.initializer.Normal(0., src_emb_dim**-0.5)))
+
+        if pos_enc_param_name is pos_enc_param_names[0]:
+            pos_inp = pos_inp1
+        else:
+            pos_inp = pos_inp2
+        self._pos_emb = Embedding(
+            name_scope=self.full_name(),
+            size=[self._src_max_len, src_emb_dim],
+            param_attr=fluid.ParamAttr(
+                name=pos_enc_param_name,
+                initializer=fluid.initializer.NumpyArrayInitializer(pos_inp),
+                trainable=False))
+
+        # use in imperative_mode to fit different length batch
+        # self._pos_emb._w = to_variable(
+        #     position_encoding_init(self._src_max_len, self._src_emb_dim))
+
+    def forward(self, src_word, src_pos):
+        src_word_emb = self._input_emb(src_word)
+        src_word_emb = fluid.layers.scale(
+            x=src_word_emb, scale=self._src_emb_dim**0.5)
+        # # TODO change this to fit dynamic length input
+        src_pos_emb = self._pos_emb(src_pos)
+        src_pos_emb.stop_gradient = True
+        enc_input = src_word_emb + src_pos_emb
+        return fluid.layers.dropout(
+            enc_input,
+            dropout_prob=self._dropout_rate,
+            seed=ModelHyperParams.dropout_seed,
+            is_test=False) if self._dropout_rate else enc_input
+
+
+class WrapEncoderLayer(Layer):
+    def __init__(self, name_cope, src_vocab_size, max_length, n_layer, n_head,
+                 d_key, d_value, d_model, d_inner_hid, prepostprocess_dropout,
+                 attention_dropout, relu_dropout, preprocess_cmd,
+                 postprocess_cmd, weight_sharing):
+        """
+        The wrapper assembles together all needed layers for the encoder.
+        """
+        super(WrapEncoderLayer, self).__init__(name_cope)
+
+        self._prepare_encoder_layer = PrepareEncoderDecoderLayer(
+            self.full_name(),
+            src_vocab_size,
+            d_model,
+            max_length,
+            prepostprocess_dropout,
+            word_emb_param_name=word_emb_param_names[0],
+            pos_enc_param_name=pos_enc_param_names[0])
+        self._encoder = EncoderLayer(
+            self.full_name(), n_layer, n_head, d_key, d_value, d_model,
+            d_inner_hid, prepostprocess_dropout, attention_dropout,
+            relu_dropout, preprocess_cmd, postprocess_cmd)
+
+    def forward(self, enc_inputs):
+        src_word, src_pos, src_slf_attn_bias = enc_inputs
+        enc_input = self._prepare_encoder_layer(src_word, src_pos)
+        enc_output = self._encoder(enc_input, src_slf_attn_bias)
+        return enc_output
+
+
+class DecoderSubLayer(Layer):
+    def __init__(self,
+                 name_scope,
+                 n_head,
+                 d_key,
+                 d_value,
+                 d_model,
+                 d_inner_hid,
+                 prepostprocess_dropout,
+                 attention_dropout,
+                 relu_dropout,
+                 preprocess_cmd,
+                 postprocess_cmd,
+                 cache=None,
+                 gather_idx=None):
+        super(DecoderSubLayer, self).__init__(name_scope)
+        self._postprocess_cmd = postprocess_cmd
+        self._preprocess_cmd = preprocess_cmd
+        self._prepostprcess_dropout = prepostprocess_dropout
+        self._pre_process_layer = PrePostProcessLayer(self.full_name(),
+                                                      preprocess_cmd, 3)
+        self._multihead_attention_layer = MultiHeadAttentionLayer(
+            self.full_name(),
+            d_key,
+            d_value,
+            d_model,
+            n_head,
+            attention_dropout,
+            cache=cache,
+            gather_idx=gather_idx)
+        self._post_process_layer = PrePostProcessLayer(self.full_name(),
+                                                       postprocess_cmd, None)
+        self._pre_process_layer2 = PrePostProcessLayer(self.full_name(),
+                                                       preprocess_cmd, 3)
+        self._multihead_attention_layer2 = MultiHeadAttentionLayer(
+            self.full_name(),
+            d_key,
+            d_value,
+            d_model,
+            n_head,
+            attention_dropout,
+            cache=cache,
+            gather_idx=gather_idx,
+            static_kv=True)
+        self._post_process_layer2 = PrePostProcessLayer(self.full_name(),
+                                                        postprocess_cmd, None)
+        self._pre_process_layer3 = PrePostProcessLayer(self.full_name(),
+                                                       preprocess_cmd, 3)
+        self._positionwise_feed_forward_layer = PositionwiseFeedForwardLayer(
+            self.full_name(), d_inner_hid, d_model, relu_dropout)
+        self._post_process_layer3 = PrePostProcessLayer(self.full_name(),
+                                                        postprocess_cmd, None)
+
+    def forward(self, dec_input, enc_output, slf_attn_bias, dec_enc_attn_bias):
+        pre_process_rlt = self._pre_process_layer(
+            None, dec_input, self._preprocess_cmd, self._prepostprcess_dropout)
+        slf_attn_output = self._multihead_attention_layer(pre_process_rlt, None,
+                                                          None, slf_attn_bias)
+        slf_attn_output_pp = self._post_process_layer(
+            dec_input, slf_attn_output, self._postprocess_cmd,
+            self._prepostprcess_dropout)
+        pre_process_rlt2 = self._pre_process_layer2(None, slf_attn_output_pp,
+                                                    self._preprocess_cmd,
+                                                    self._prepostprcess_dropout)
+        enc_attn_output_pp = self._multihead_attention_layer2(
+            pre_process_rlt2, enc_output, enc_output, dec_enc_attn_bias)
+        enc_attn_output = self._post_process_layer2(
+            slf_attn_output, enc_attn_output_pp, self._postprocess_cmd,
+            self._prepostprcess_dropout)
+        pre_process_rlt3 = self._pre_process_layer3(None, enc_attn_output,
+                                                    self._preprocess_cmd,
+                                                    self._prepostprcess_dropout)
+        ffd_output = self._positionwise_feed_forward_layer(pre_process_rlt3)
+        dec_output = self._post_process_layer3(enc_attn_output, ffd_output,
+                                               self._postprocess_cmd,
+                                               self._prepostprcess_dropout)
+        return dec_output
+
+
+class DecoderLayer(Layer):
+    def __init__(self,
+                 name_scope,
+                 n_layer,
+                 n_head,
+                 d_key,
+                 d_value,
+                 d_model,
+                 d_inner_hid,
+                 prepostprocess_dropout,
+                 attention_dropout,
+                 relu_dropout,
+                 preprocess_cmd,
+                 postprocess_cmd,
+                 caches=None,
+                 gather_idx=None):
+        super(DecoderLayer, self).__init__(name_scope)
+        self._pre_process_layer = PrePostProcessLayer(self.full_name(),
+                                                      preprocess_cmd, 3)
+        self._decoder_sub_layers = list()
+        self._n_layer = n_layer
+        self._preprocess_cmd = preprocess_cmd
+        self._prepostprocess_dropout = prepostprocess_dropout
+        for i in range(n_layer):
+            self._decoder_sub_layers.append(
+                self.add_sublayer(
+                    'dsl_%d' % i,
+                    DecoderSubLayer(
+                        self.full_name(),
+                        n_head,
+                        d_key,
+                        d_value,
+                        d_model,
+                        d_inner_hid,
+                        prepostprocess_dropout,
+                        attention_dropout,
+                        relu_dropout,
+                        preprocess_cmd,
+                        postprocess_cmd,
+                        cache=None if caches is None else caches[i],
+                        gather_idx=gather_idx)))
+
+    def forward(self, dec_input, enc_output, dec_slf_attn_bias,
+                dec_enc_attn_bias):
+        for i in range(self._n_layer):
+            tmp_dec_output = self._decoder_sub_layers[i](
+                dec_input, enc_output, dec_slf_attn_bias, dec_enc_attn_bias)
+            dec_input = tmp_dec_output
+
+        dec_output = self._pre_process_layer(None, tmp_dec_output,
+                                             self._preprocess_cmd,
+                                             self._prepostprocess_dropout)
+        return dec_output
+
+
+class WrapDecoderLayer(Layer):
+    def __init__(self,
+                 name_scope,
+                 trg_vocab_size,
+                 max_length,
+                 n_layer,
+                 n_head,
+                 d_key,
+                 d_value,
+                 d_model,
+                 d_inner_hid,
+                 prepostprocess_dropout,
+                 attention_dropout,
+                 relu_dropout,
+                 preprocess_cmd,
+                 postprocess_cmd,
+                 weight_sharing,
+                 caches=None,
+                 gather_idx=None):
+        """
+        The wrapper assembles together all needed layers for the encoder.
+        """
+        super(WrapDecoderLayer, self).__init__(name_scope)
+
+        self._prepare_decoder_layer = PrepareEncoderDecoderLayer(
+            self.full_name(),
+            trg_vocab_size,
+            d_model,
+            max_length,
+            prepostprocess_dropout,
+            word_emb_param_name=word_emb_param_names[1],
+            pos_enc_param_name=pos_enc_param_names[1])
+        self._decoder_layer = DecoderLayer(
+            self.full_name(),
+            n_layer,
+            n_head,
+            d_key,
+            d_value,
+            d_model,
+            d_inner_hid,
+            prepostprocess_dropout,
+            attention_dropout,
+            relu_dropout,
+            preprocess_cmd,
+            postprocess_cmd,
+            caches=caches,
+            gather_idx=gather_idx)
+        self._weight_sharing = weight_sharing
+        if not weight_sharing:
+            self._fc = FC(self.full_name(),
+                          size=trg_vocab_size,
+                          bias_attr=False)
+
+    def forward(self, dec_inputs=None, enc_output=None):
+        trg_word, trg_pos, trg_slf_attn_bias, trg_src_attn_bias = dec_inputs
+        dec_input = self._prepare_decoder_layer(trg_word, trg_pos)
+        dec_output = self._decoder_layer(dec_input, enc_output,
+                                         trg_slf_attn_bias, trg_src_attn_bias)
+
+        dec_output_reshape = fluid.layers.reshape(
+            dec_output, shape=[-1, dec_output.shape[-1]], inplace=False)
+
+        if self._weight_sharing:
+            predict = fluid.layers.matmul(
+                x=dec_output_reshape,
+                y=self._prepare_decoder_layer._input_emb._w,
+                transpose_y=True)
+        else:
+            predict = self._fc(dec_output_reshape)
+
+        if dec_inputs is None:
+            # Return probs for independent decoder program.
+            predict_out = fluid.layers.softmax(predict)
+            return predict_out
+        return predict
+
+
+class TransFormer(Layer):
+    def __init__(self,
+                 name_scope,
+                 src_vocab_size,
+                 trg_vocab_size,
+                 max_length,
+                 n_layer,
+                 n_head,
+                 d_key,
+                 d_value,
+                 d_model,
+                 d_inner_hid,
+                 prepostprocess_dropout,
+                 attention_dropout,
+                 relu_dropout,
+                 preprocess_cmd,
+                 postprocess_cmd,
+                 weight_sharing,
+                 label_smooth_eps,
+                 use_py_reader=False,
+                 is_test=False):
+        super(TransFormer, self).__init__(name_scope)
+        self._label_smooth_eps = label_smooth_eps
+        self._trg_vocab_size = trg_vocab_size
+        if weight_sharing:
+            assert src_vocab_size == trg_vocab_size, (
+                "Vocabularies in source and target should be same for weight sharing."
+            )
+        self._wrap_encoder_layer = WrapEncoderLayer(
+            self.full_name(), src_vocab_size, max_length, n_layer, n_head,
+            d_key, d_value, d_model, d_inner_hid, prepostprocess_dropout,
+            attention_dropout, relu_dropout, preprocess_cmd, postprocess_cmd,
+            weight_sharing)
+        self._wrap_decoder_layer = WrapDecoderLayer(
+            self.full_name(), trg_vocab_size, max_length, n_layer, n_head,
+            d_key, d_value, d_model, d_inner_hid, prepostprocess_dropout,
+            attention_dropout, relu_dropout, preprocess_cmd, postprocess_cmd,
+            weight_sharing)
+
+        if weight_sharing:
+            self._wrap_decoder_layer._prepare_decoder_layer._input_emb._w = self._wrap_encoder_layer._prepare_encoder_layer._input_emb._w
+
+    def forward(self, enc_inputs, dec_inputs, label, weights):
+        enc_output = self._wrap_encoder_layer(enc_inputs)
+        predict = self._wrap_decoder_layer(dec_inputs, enc_output)
+        if self._label_smooth_eps:
+            label_out = fluid.layers.label_smooth(
+                label=fluid.layers.one_hot(
+                    input=label, depth=self._trg_vocab_size),
+                epsilon=self._label_smooth_eps)
+
+        cost = fluid.layers.softmax_with_cross_entropy(
+            logits=predict,
+            label=label_out,
+            soft_label=True if self._label_smooth_eps else False)
+        weighted_cost = cost * weights
+        sum_cost = fluid.layers.reduce_sum(weighted_cost)
+        token_num = fluid.layers.reduce_sum(weights)
+        token_num.stop_gradient = True
+        avg_cost = sum_cost / token_num
+        return sum_cost, avg_cost, predict, token_num
+
+
+class TestImperativeTransformer(unittest.TestCase):
+    def test_transformer_float32(self):
+        seed = 90
+        with guard():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+            transformer = TransFormer(
+                'transformer',
+                ModelHyperParams.src_vocab_size,
+                ModelHyperParams.trg_vocab_size,
+                ModelHyperParams.max_length + 1,
+                ModelHyperParams.n_layer,
+                ModelHyperParams.n_head,
+                ModelHyperParams.d_key,
+                ModelHyperParams.d_value,
+                ModelHyperParams.d_model,
+                ModelHyperParams.d_inner_hid,
+                ModelHyperParams.prepostprocess_dropout,
+                ModelHyperParams.attention_dropout,
+                ModelHyperParams.relu_dropout,
+                ModelHyperParams.preprocess_cmd,
+                ModelHyperParams.postprocess_cmd,
+                ModelHyperParams.weight_sharing,
+                TrainTaskConfig.label_smooth_eps,
+                use_py_reader=use_py_reader,
+                is_test=False)
+            if sync:
+                lr_decay = fluid.layers.learning_rate_scheduler.noam_decay(
+                    ModelHyperParams.d_model, TrainTaskConfig.warmup_steps)
+                with fluid.default_main_program()._lr_schedule_guard():
+                    learning_rate = lr_decay * TrainTaskConfig.learning_rate
+                optimizer = fluid.optimizer.Adam(
+                    learning_rate=learning_rate,
+                    beta1=TrainTaskConfig.beta1,
+                    beta2=TrainTaskConfig.beta2,
+                    epsilon=TrainTaskConfig.eps)
+            else:
+                optimizer = fluid.optimizer.SGD(learning_rate=0.003)
+            dy_param_init = dict()
+            dy_param_updated = dict()
+            for i in range(batch_num):
+                enc_inputs, dec_inputs, label, weights = create_data()
+                dy_sum_cost, dy_avg_cost, dy_predict, dy_token_num = transformer(
+                    enc_inputs, dec_inputs, label, weights)
+                if i == 0:
+                    for param in transformer.parameters():
+                        dy_param_init[param.name] = param._numpy()
+
+                dy_avg_cost._backward()
+                optimizer.minimize(dy_avg_cost)
+                transformer.clear_gradients()
+                if i == batch_num - 1:
+                    for param in transformer.parameters():
+                        dy_param_updated[param.name] = param._numpy()
+
+        with new_program_scope():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+            transformer = TransFormer(
+                'transformer',
+                ModelHyperParams.src_vocab_size,
+                ModelHyperParams.trg_vocab_size,
+                ModelHyperParams.max_length + 1,
+                ModelHyperParams.n_layer,
+                ModelHyperParams.n_head,
+                ModelHyperParams.d_key,
+                ModelHyperParams.d_value,
+                ModelHyperParams.d_model,
+                ModelHyperParams.d_inner_hid,
+                ModelHyperParams.prepostprocess_dropout,
+                ModelHyperParams.attention_dropout,
+                ModelHyperParams.relu_dropout,
+                ModelHyperParams.preprocess_cmd,
+                ModelHyperParams.postprocess_cmd,
+                ModelHyperParams.weight_sharing,
+                TrainTaskConfig.label_smooth_eps,
+                use_py_reader=use_py_reader,
+                is_test=False)
+            exe = fluid.Executor(fluid.CPUPlace(
+            ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
+            optimizer = fluid.optimizer.SGD(learning_rate=0.003)
+
+            data_input_names = encoder_data_input_fields + decoder_data_input_fields[:
+                                                                                     -1] + label_data_input_fields
+            all_inputs = make_all_inputs(data_input_names)
+            enc_inputs_len = len(encoder_data_input_fields)
+            dec_inputs_len = len(decoder_data_input_fields[:-1])
+            enc_inputs = all_inputs[0:enc_inputs_len]
+            dec_inputs = all_inputs[enc_inputs_len:enc_inputs_len +
+                                    dec_inputs_len]
+            label = all_inputs[-2]
+            weights = all_inputs[-1]
+            static_param_updated = dict()
+            static_param_init = dict()
+            static_param_name_list = list()
+            static_sum_cost, static_avg_cost, static_predict, static_token_num = transformer(
+                enc_inputs, dec_inputs, label, weights)
+
+            optimizer.minimize(static_avg_cost)
+            for param in transformer.parameters():
+                static_param_name_list.append(param.name)
+            out = exe.run(fluid.default_startup_program(),
+                          fetch_list=static_param_name_list)
+            for i in range(len(static_param_name_list)):
+                static_param_init[static_param_name_list[i]] = out[i]
+            static_sum_cost_value = None
+            static_avg_cost_value = None
+            static_predict_value = None
+            static_token_num_value = None
+            for i in range(batch_num):
+                feed_dict = create_feed_dict_list(create_data(True))
+                fetch_list = [
+                    static_sum_cost, static_avg_cost, static_predict,
+                    static_token_num
+                ]
+                fetch_list.extend(static_param_name_list)
+
+                out = exe.run(fluid.default_main_program(),
+                              feed=feed_dict,
+                              fetch_list=fetch_list)
+                static_sum_cost_value = out[0]
+                static_avg_cost_value = out[1]
+                static_predict_value = out[2]
+                static_token_num_value = out[3]
+                if i == batch_num - 1:
+                    for k in range(4, len(out)):
+                        static_param_updated[static_param_name_list[k -
+                                                                    4]] = out[k]
+
+        self.assertTrue(
+            np.allclose(static_avg_cost_value, dy_avg_cost._numpy()))
+        self.assertTrue(
+            np.allclose(static_sum_cost_value, dy_sum_cost._numpy()))
+        self.assertTrue(
+            np.allclose(
+                static_predict_value, dy_predict._numpy(), atol=1e-5))
+        self.assertTrue(
+            np.allclose(static_token_num_value, dy_token_num._numpy()))
+        for key, value in six.iteritems(static_param_init):
+            self.assertTrue(np.allclose(value, dy_param_init[key]))
+        for key, value in six.iteritems(static_param_updated):
+            self.assertTrue(
+                np.allclose(
+                    value, dy_param_updated[key], atol=1e-4))
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_layers.py

@@ -70,6 +70,34 @@ class LayerTest(unittest.TestCase):
 
 
 class TestLayer(LayerTest):
+    def test_layer_norm(self):
+        inp = np.ones([3, 32, 32], dtype='float32')
+        with self.static_graph():
+            t = layers.data(
+                name='data',
+                shape=[3, 32, 32],
+                dtype='float32',
+                append_batch_size=False)
+            ret = layers.layer_norm(t)
+            static_ret = self.get_static_graph_result(
+                feed={'data': inp}, fetch_list=[ret])[0]
+        with self.static_graph():
+            t = layers.data(
+                name='data',
+                shape=[3, 32, 32],
+                dtype='float32',
+                append_batch_size=False)
+            lm = nn.LayerNorm('layer_norm')
+            ret = lm(t)
+            static_ret2 = self.get_static_graph_result(
+                feed={'data': inp}, fetch_list=[ret])[0]
+        with self.dynamic_graph():
+            lm = nn.LayerNorm('layer_norm')
+            dy_ret = lm(base.to_variable(inp))
+
+        self.assertTrue(np.allclose(static_ret, static_ret2))
+        self.assertTrue(np.allclose(dy_ret._numpy(), static_ret2))
+
     def test_relu(self):
         with self.static_graph():
             t = layers.data(name='t', shape=[3, 3], dtype='float32')
@@ -1240,6 +1268,14 @@ class TestBook(unittest.TestCase):
 
         print(str(program))
 
+    def test_range(self):
+        program = Program()
+        with program_guard(program):
+            layers.range(0, 10, 2, 'int32')
+            layers.range(0.1, 10.0, 0.2, 'float32')
+
+        print(str(program))
+
     def test_spectral_norm(self):
         program = Program()
         with program_guard(program):
@@ -1261,6 +1297,15 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(out)
         print(str(program))
 
+    def test_fsp(self):
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name="X", shape=[16, 4, 4], dtype="float32")
+            y = layers.data(name="Y", shape=[8, 4, 4], dtype="float32")
+            out = layers.fsp_matrix(x, y)
+            self.assertIsNotNone(out)
+        print(str(program))
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_py_reader_sample_generator.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import paddle.fluid as fluid
+import math
+import unittest
+import numpy as np
+import os
+
+os.environ['CPU_NUM'] = '1'
+
+
+def random_reader(sample_num):
+    def __impl__():
+        for _ in range(sample_num):
+            yield np.random.random(
+                size=[784]).astype('float32'), np.random.random_integers(
+                    low=0, high=9, size=[1]).astype('int64')
+
+    return paddle.reader.cache(__impl__)
+
+
+class TestCaseBase(unittest.TestCase):
+    def setUp(self):
+        self.batch_size = 32
+        self.epoch_num = 2
+        self.sample_num = 165
+
+    def generate_all_data(self, reader):
+        ret = []
+        for d in reader():
+            slots = [[], []]
+            for item in d:
+                slots[0].append(item[0])
+                slots[1].append(item[1])
+            slots = [np.array(slot) for slot in slots]
+            ret.append(slots)
+        return ret
+
+    def run_main(self, reader, use_sample_generator, iterable, drop_last):
+        image = fluid.layers.data(name='image', dtype='float32', shape=[784])
+        label = fluid.layers.data(name='label', dtype='int64', shape=[1])
+        py_reader = fluid.io.PyReader(
+            feed_list=[image, label],
+            capacity=16,
+            iterable=iterable,
+            use_double_buffer=False)
+
+        batch_reader = paddle.batch(reader, self.batch_size, drop_last)
+        all_datas = self.generate_all_data(batch_reader)
+
+        if not use_sample_generator:
+            py_reader.decorate_sample_list_generator(
+                batch_reader, places=fluid.cpu_places())
+        else:
+            py_reader.decorate_sample_generator(
+                reader, self.batch_size, drop_last, places=fluid.cpu_places())
+
+        if drop_last:
+            batch_num = int(self.sample_num / self.batch_size)
+        else:
+            batch_num = math.ceil(float(self.sample_num) / self.batch_size)
+
+        exe = fluid.Executor(fluid.CPUPlace())
+        exe.run(fluid.default_startup_program())
+        for _ in range(self.epoch_num):
+            if py_reader.iterable:
+                step = 0
+                for data in py_reader():
+                    img, lbl = exe.run(feed=data, fetch_list=[image, label])
+                    self.assertArrayEqual(img, all_datas[step][0])
+                    self.assertArrayEqual(lbl, all_datas[step][1])
+                    step += 1
+                self.assertEqual(step, len(all_datas))
+            else:
+                step = 0
+                try:
+                    py_reader.start()
+                    while True:
+                        img, lbl = exe.run(fetch_list=[image, label])
+                        self.assertArrayEqual(img, all_datas[step][0])
+                        self.assertArrayEqual(lbl, all_datas[step][1])
+                        step += 1
+                except fluid.core.EOFException:
+                    py_reader.reset()
+                    self.assertEqual(step, len(all_datas))
+                    break
+
+    def assertArrayEqual(self, arr1, arr2):
+        self.assertEqual(arr1.shape, arr2.shape)
+        self.assertTrue((arr1 == arr2).all())
+
+    def test_main(self):
+        reader = random_reader(self.sample_num)
+        for use_sample_generator in [False, True]:
+            for iterable in [False, True]:
+                for drop_last in [False, True]:
+                    with fluid.program_guard(fluid.Program(), fluid.Program()):
+                        self.run_main(reader, use_sample_generator, iterable,
+                                      drop_last)
+
+
+class TestCase1(TestCaseBase):
+    def setUp(self):
+        self.batch_size = 32
+        self.epoch_num = 10
+        self.sample_num = 160
+
+
+class TestCase2(TestCaseBase):
+    def setUp(self):
+        self.batch_size = 32
+        self.epoch_num = 2
+        self.sample_num = 200
+
+
+class TestCase3(TestCaseBase):
+    def setUp(self):
+        self.batch_size = 32
+        self.epoch_num = 2
+        self.sample_num = 159
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/contrib/slim/graph/graph.py → python/paddle/fluid/tests/unittests/test_range.py

-# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -11,39 +11,60 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 from __future__ import print_function
-import os
-import subprocess
-from ....framework import Program
-from ....framework import Block
-from .... import core
 
-__all__ = ['Graph', 'ImitationGraph', 'IRGraph']
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestRangeOp(OpTest):
+    def setUp(self):
+        self.op_type = "range"
+        self.init_config()
+        self.inputs = {
+            'Start': np.array([self.case[0]]).astype(self.dtype),
+            'End': np.array([self.case[1]]).astype(self.dtype),
+            'Step': np.array([self.case[2]]).astype(self.dtype)
+        }
+
+        self.outputs = {
+            'Out': np.arange(self.case[0], self.case[1],
+                             self.case[2]).astype(self.dtype)
+        }
+
+    def init_config(self):
+        self.dtype = np.float32
+        self.case = (0, 1, 0.2)
+
+    def test_check_output(self):
+        self.check_output()
+
 
+class TestFloatRangeOpCase0(TestRangeOp):
+    def init_config(self):
+        self.dtype = np.float32
+        self.case = (0, 5, 1)
 
-class Graph(object):
-    """
-    Base class for all graph.
-    """
 
-    def __init__(self):
-        pass
+class TestInt32RangeOpCase0(TestRangeOp):
+    def init_config(self):
+        self.dtype = np.int32
+        self.case = (0, 5, 2)
 
-    def all_parameters(self):
-        """
-        Return all the parameters in current graph.
-        """
-        pass
 
+class TestInt32RangeOpCase1(TestRangeOp):
+    def init_config(self):
+        self.dtype = np.int32
+        self.case = (10, 1, -2)
 
-class ImitationGraph(Graph):
-    def __init__(self, program=None):
-        super(ImitationGraph, self).__init__()
-        self.program = Program() if program is None else program
 
-    def all_parameters(self):
-        return self.program.global_block().all_parameters()
+class TestInt32RangeOpCase2(TestRangeOp):
+    def init_config(self):
+        self.dtype = np.int32
+        self.case = (-1, -10, -2)
 
 
-class IRGraph(Graph):
-    pass
+if __name__ == "__main__":
+    unittest.main()

python/paddle/reader/decorator.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 __all__ = [
-    'map_readers', 'buffered', 'compose', 'chain', 'shuffle',
+    'cache', 'map_readers', 'buffered', 'compose', 'chain', 'shuffle',
     'ComposeNotAligned', 'firstn', 'xmap_readers', 'PipeReader',
     'multiprocess_reader', 'Fake'
 ]
@@ -33,6 +33,30 @@ import zlib
 import paddle.compat as cpt
 
 
+def cache(reader):
+    """
+    Cache the reader data into memory. 
+
+    Be careful that this method may take long time to process, 
+    and consume lots of memory. :code:`reader()` would only 
+    call once. 
+
+    Args:
+        reader (generator): a reader object which yields 
+            data each time.
+
+    Returns:
+        generator: a decorated reader object which yields data from cached memory.
+    """
+    all_data = tuple(reader())
+
+    def __impl__():
+        for item in all_data:
+            yield item
+
+    return __impl__
+
+
 def map_readers(func, *readers):
     """
     Creates a data reader that outputs return value of function using

python/requirements.txt

@@ -12,3 +12,4 @@ six
 funcsigs
 pyyaml
 decorator
+prettytable

python/setup.py.in

@@ -117,6 +117,7 @@ packages=['paddle',
           'paddle.fluid.contrib.slim.graph',
           'paddle.fluid.contrib.slim.prune',
           'paddle.fluid.contrib.slim.quantization',
+          'paddle.fluid.contrib.slim.distillation',
           'paddle.fluid.contrib.utils',
           'paddle.fluid.transpiler',
           'paddle.fluid.transpiler.details']

tools/manylinux1/Dockerfile.x64

@@ -52,7 +52,7 @@ RUN LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH} /o
     LD_LIBRARY_PATH=/opt/_internal/cpython-3.6.0/lib/:${LD_LIBRARY_PATH} /opt/_internal/cpython-3.6.0/bin/pip3 install pre-commit 'ipython==5.3.0' opencv-python && \
     LD_LIBRARY_PATH=/opt/_internal/cpython-3.7.0/lib/:${LD_LIBRARY_PATH} /opt/_internal/cpython-3.7.0/bin/pip3 install pre-commit 'ipython==5.3.0' opencv-python
 
-RUN wget -O /opt/swig-2.0.12.tar.gz https://cytranet.dl.sourceforge.net/project/swig/swig/swig-2.0.12/swig-2.0.12.tar.gz && \
+RUN wget -O /opt/swig-2.0.12.tar.gz https://sourceforge.net/projects/swig/files/swig/swig-2.0.12/swig-2.0.12.tar.gz/download && \
     cd /opt && tar xzf swig-2.0.12.tar.gz && cd /opt/swig-2.0.12 && ./configure && make && make install && cd /opt && rm swig-2.0.12.tar.gz
 
 CMD ["bash", "/paddle/paddle/scripts/docker/build.sh"]