diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index 0fd283b868f68095501a9d0236b84267905355ab..149cf3b86d03665252b57bc295e54f33d02bddfa 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -576,7 +576,7 @@ endif() py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf) # Coverage pipeline use cuda 10.1 now, profiler will random hang in cuda 10.1, # see https://github.com/PaddlePaddle/Paddle/issues/29082 for details. -# We guess there are some bugs in cuda 10.1 or 10.2, +# We guess there are some bugs in cuda 10.1 or 10.2, # since this unittest is stable in cuda 11 (py3 pipeline) now. if(NOT WITH_COVERAGE) py_test_modules(test_parallel_executor_profiler MODULES test_parallel_executor_profiler) @@ -601,8 +601,8 @@ py_test_modules(test_fuse_bn_act_pass MODULES test_fuse_bn_act_pass ENVS FLAGS_c py_test_modules(test_fuse_bn_add_act_pass MODULES test_fuse_bn_add_act_pass ENVS FLAGS_cudnn_deterministic=1 FLAGS_cudnn_batchnorm_spatial_persistent=1 FLAGS_conv_workspace_size_limit=1000) # NOTE: These unittests will appear NaN steadily in windows CI. After analysis, -# it is found that windows CI will run all the training unittests with the ON_INFER option turned on, -# which will not appear in other CIs. The calculation behavior of some ops in inference mode is +# it is found that windows CI will run all the training unittests with the ON_INFER option turned on, +# which will not appear in other CIs. The calculation behavior of some ops in inference mode is # inconsistent with that in non-inference mode. if(NOT ON_INFER) py_test_modules(test_parallel_executor_seresnext_base_cpu MODULES test_parallel_executor_seresnext_base_cpu) @@ -645,7 +645,7 @@ if (WITH_XPU) add_subdirectory(xpu) endif() -# dist xpu tests: +# dist xpu tests: if (WITH_XPU_BKCL) py_test(test_collective_reduce_api_xpu SRCS "test_collective_reduce_api.py") py_test(test_collective_allreduce_api_xpu SRCS "test_collective_allreduce_api.py") @@ -713,6 +713,7 @@ if (WITH_DISTRIBUTE) set_tests_properties(test_dist_fleet_ctr2 PROPERTIES TIMEOUT 200) set_tests_properties(test_dist_fleet_sparse_embedding_ctr PROPERTIES TIMEOUT 200) set_tests_properties(test_dist_fleet_infer PROPERTIES TIMEOUT 200) + set_tests_properties(test_dist_fleet_raw_program_optimizer PROPERTIES TIMEOUT 120) endif() if (WITH_DISTRIBUTE AND NOT APPLE) diff --git a/python/paddle/fluid/tests/unittests/dist_fleet_raw_program_optimizer.py b/python/paddle/fluid/tests/unittests/dist_fleet_raw_program_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..575c07390a35bbef00694a1e1c40bc0598e741ab --- /dev/null +++ b/python/paddle/fluid/tests/unittests/dist_fleet_raw_program_optimizer.py @@ -0,0 +1,109 @@ +# Copyright (c) 2021 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 test_dist_base import TestDistRunnerBase, runtime_main +import unittest +import paddle +import os +import paddle.distributed.fleet as fleet +import paddle.distributed.fleet.base.role_maker as role_maker +import numpy as np +from functools import reduce +import paddle.fluid as fluid + +paddle.enable_static() + +DTYPE = "float32" +paddle.dataset.mnist.fetch() + +# Fix seed for test +fluid.default_startup_program().random_seed = 1 +fluid.default_main_program().random_seed = 1 + + +def cnn_model(data): + conv_pool_1 = fluid.nets.simple_img_conv_pool( + input=data, + filter_size=5, + num_filters=20, + pool_size=2, + pool_stride=2, + act="relu", + param_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant( + value=0.01))) + conv_pool_2 = fluid.nets.simple_img_conv_pool( + input=conv_pool_1, + filter_size=5, + num_filters=50, + pool_size=2, + pool_stride=2, + act="relu", + param_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant( + value=0.01))) + + SIZE = 10 + input_shape = conv_pool_2.shape + param_shape = [reduce(lambda a, b: a * b, input_shape[1:], 1)] + [SIZE] + scale = (2.0 / (param_shape[0]**2 * SIZE))**0.5 + + predict = fluid.layers.fc( + input=conv_pool_2, + size=SIZE, + act="softmax", + param_attr=fluid.param_attr.ParamAttr( + initializer=fluid.initializer.Constant(value=0.01))) + return predict + + +class TestFleetMetaOptimizerPrecision(TestDistRunnerBase): + def get_model(self, batch_size=2, single_device=False): + # Input data + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + # Train program + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_size_tensor = fluid.layers.create_tensor(dtype='int64') + batch_acc = fluid.layers.accuracy( + input=predict, label=label, total=batch_size_tensor) + + test_program = fluid.default_main_program().clone(for_test=True) + + # Reader + train_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + + optimizer = paddle.fluid.optimizer.Adam(0.01) + if single_device: + optimizer.minimize(avg_cost) + else: + role = role_maker.PaddleCloudRoleMaker(is_collective=True) + fleet.init(role) + strategy = paddle.distributed.fleet.DistributedStrategy() + strategy.without_graph_optimization = True + optimizer = fleet.distributed_optimizer( + optimizer, strategy=strategy) + optimizer.minimize(avg_cost) + + return test_program, avg_cost, train_reader, test_reader, batch_acc, predict + + +if __name__ == "__main__": + runtime_main(TestFleetMetaOptimizerPrecision) diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py index edc510e4e766d7f1e8898c831204806b0b8f954d..78b06bd5333d79b4aa90d00f1c1f16a399e61929 100755 --- a/python/paddle/fluid/tests/unittests/test_dist_base.py +++ b/python/paddle/fluid/tests/unittests/test_dist_base.py @@ -186,6 +186,76 @@ class TestDistRunnerBase(object): fleet.save_inference_model(exe, infer_save_dir_fleet, feeded_var_names, [avg_cost]) + def run_use_fleet_api_20_trainer(self, args): + """ + 1. remove codes for DistributedStrategy and leave the DistributedStrategy part to get_model() + 2. to run with fleet 2.0 api, set flags _use_fleet_api and _use_fleet_api_20 to True + 3. for now, not support test for model save + """ + assert args.update_method == "nccl2" or "bkcl" + + self.lr = args.lr + print_to_err("use_fleet 2.0", "fleet.node_num:") + + test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \ + self.get_model(batch_size=args.batch_size) + + if fluid.core.is_compiled_with_cuda(): + device_id = int(os.getenv("FLAGS_selected_gpus", "0")) + place = fluid.CUDAPlace(device_id) + elif fluid.core.is_compiled_with_xpu(): + device_id = int(os.getenv("FLAGS_selected_xpus", "0")) + place = fluid.XPUPlace(device_id) + else: + raise ValueError( + "fleet dygraph api must in paddlepaddle-xpu or paddlepaddle-gpu." + ) + + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + eprint(type(self).__name__, "run worker startup program done.") + + feed_var_list = [ + var + for var in fluid.default_main_program().global_block().vars.values() + if var.is_data + ] + + eprint("feed_var_list:", feed_var_list) + + if feed_var_list[0].name == 'label': + feed_var_list = feed_var_list[::-1] + + feeder = fluid.DataFeeder(feed_var_list, place) + reader_generator = train_reader() + + def get_data(): + origin_batch = next(reader_generator) + if args.update_method != "local" and args.use_reader_alloc: + new_batch = [] + for offset, item in enumerate(origin_batch): + if offset % 2 == args.trainer_id: + new_batch.append(item) + return new_batch + else: + return origin_batch + + print_to_err(type(self).__name__, "begin to train on trainer") + out_losses = [] + for i in six.moves.xrange(RUN_STEP): + loss, = exe.run(fluid.default_main_program(), + fetch_list=[avg_cost.name], + feed=feeder.feed(get_data())) + out_losses.append(loss[0]) + print_to_err(type(self).__name__, "run step %d finished" % i) + print_to_err(type(self).__name__, "trainer run finished") + print_to_err(type(self).__name__, "dist losses: {}".format(out_losses)) + + if six.PY2: + print(pickle.dumps(out_losses)) + else: + sys.stdout.buffer.write(pickle.dumps(out_losses)) + def run_use_fleet_api_trainer(self, args): assert args.update_method == "nccl2" or "bkcl" @@ -630,6 +700,7 @@ def runtime_main(test_class): parser.add_argument('--use_hallreduce', action='store_true') parser.add_argument('--use_pipeline', action='store_true') parser.add_argument('--use_fleet_api', action='store_true') + parser.add_argument('--use_fleet_api_20', action='store_true') parser.add_argument('--use_local_sgd', action='store_true') parser.add_argument('--ut4grad_allreduce', action='store_true') parser.add_argument( @@ -671,6 +742,8 @@ def runtime_main(test_class): model.run_pserver(args) elif args.use_fleet_api: model.run_use_fleet_api_trainer(args) + elif args.use_fleet_api_20: + model.run_use_fleet_api_20_trainer(args) elif args.use_pipeline: model.run_pipeline_trainer(args) else: @@ -734,6 +807,7 @@ class TestDistBase(unittest.TestCase): self._nccl_comm_num = 1 self._enable_backward_deps = False self._use_fleet_api = False + self._use_fleet_api_20 = False self._use_local_sgd = False self._ut4grad_allreduce = False self._use_hallreduce = False @@ -1060,7 +1134,7 @@ class TestDistBase(unittest.TestCase): tr_cmd += " --fuse_all_reduce {}".format(self._fuse_all_reduce) if self._use_fleet_api: - tr_cmd += " --use_fleet_api" + tr_cmd += " --use_fleet_api_20" if self._use_fleet_api_20 else " --use_fleet_api" if self._use_local_sgd: tr_cmd += " --use_local_sgd" if self._ut4grad_allreduce: diff --git a/python/paddle/fluid/tests/unittests/test_dist_fleet_raw_program_optimizer.py b/python/paddle/fluid/tests/unittests/test_dist_fleet_raw_program_optimizer.py new file mode 100644 index 0000000000000000000000000000000000000000..e729bfe0537528ed9d225e65823f1eb4f06a0f5d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_fleet_raw_program_optimizer.py @@ -0,0 +1,45 @@ +# Copyright (c) 2021 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 +from test_dist_base import TestDistBase +import paddle +import os + +paddle.enable_static() +flag_name = os.path.splitext(__file__)[0] + + +class TestFleetMetaOptimizerPrecision(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._use_reduce = False + self._use_reader_alloc = False + self._nccl2_mode = True + self._nccl2_reduce_layer = True + self._use_fleet_api = True + self._use_fleet_api_20 = True + + def test_dist_train(self): + import paddle.fluid as fluid + if fluid.core.is_compiled_with_cuda(): + self.check_with_place( + "dist_fleet_raw_program_optimizer.py", + delta=1e-5, + check_error_log=True, + log_name=flag_name) + + +if __name__ == '__main__': + unittest.main()