remove return value in dist.merge() (#33)

demo/distillation/distillation_demo.py

@@ -150,7 +150,9 @@ def compress(args):
     #    print(v.name, v.shape)
 
     exe.run(t_startup)
-    _download('http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar', '.')
+    _download(
+        'http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar',
+        '.')
     _decompress('./ResNet50_pretrained.tar')
     assert args.teacher_pretrained_model and os.path.exists(
         args.teacher_pretrained_model
@@ -168,21 +170,17 @@ def compress(args):
         predicate=if_exist)
 
     data_name_map = {'image': 'image'}
-    main = merge(
-        teacher_program,
-        student_program,
-        data_name_map,
-        place)
-
-    with fluid.program_guard(main, s_startup):
-        l2_loss = l2_loss("teacher_fc_0.tmp_0", "fc_0.tmp_0", main)
+    merge(teacher_program, student_program, data_name_map, place)
+
+    with fluid.program_guard(student_program, s_startup):
+        l2_loss = l2_loss("teacher_fc_0.tmp_0", "fc_0.tmp_0", student_program)
         loss = avg_cost + l2_loss
         opt = create_optimizer(args)
         opt.minimize(loss)
     exe.run(s_startup)
     build_strategy = fluid.BuildStrategy()
     build_strategy.fuse_all_reduce_ops = False
-    parallel_main = fluid.CompiledProgram(main).with_data_parallel(
+    parallel_main = fluid.CompiledProgram(student_program).with_data_parallel(
         loss_name=loss.name, build_strategy=build_strategy)
 
     for epoch_id in range(args.num_epochs):
@@ -190,9 +188,7 @@ def compress(args):
             loss_1, loss_2, loss_3 = exe.run(
                 parallel_main,
                 feed=data,
-                fetch_list=[
-                    loss.name, avg_cost.name, l2_loss.name
-                ])
+                fetch_list=[loss.name, avg_cost.name, l2_loss.name])
             if step_id % args.log_period == 0:
                 _logger.info(
                     "train_epoch {} step {} loss {:.6f}, class loss {:.6f}, l2 loss {:.6f}".

docs/docs/api/single_distiller_api.md

 ## merge
 paddleslim.dist.merge(teacher_program, student_program, data_name_map, place, scope=fluid.global_scope(), name_prefix='teacher_') [[源代码]](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/paddleslim/dist/single_distiller.py#L19)
 
-: merge将两个paddle program（teacher_program, student_program）融合为一个program，并将融合得到的program返回。在融合的program中，可以为其中合适的teacher特征图和student特征图添加蒸馏损失函数，从而达到用teacher模型的暗知识（Dark Knowledge）指导student模型学习的目的。
+: merge将teacher_program融合到student_program中。在融合的program中，可以为其中合适的teacher特征图和student特征图添加蒸馏损失函数，从而达到用teacher模型的暗知识（Dark Knowledge）指导student模型学习的目的。
 
 **参数：**
 
@@ -12,7 +12,7 @@ paddleslim.dist.merge(teacher_program, student_program, data_name_map, place, sc
 - **scope**(Scope)-该参数表示程序使用的变量作用域，如果不指定将使用默认的全局作用域。默认值：[*fluid.global_scope()*](https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/fluid_cn/global_scope_cn.html#global-scope)
 - **name_prefix**(str)-merge操作将统一为teacher的[*Variables*](https://www.paddlepaddle.org.cn/documentation/docs/zh/1.3/api_guides/low_level/program.html#variable)添加的名称前缀name_prefix。默认值：'teacher_'
 
-**返回：** 由student_program和teacher_program merge得到的program
+**返回：** 无
 
 !!! note "Note"
     *data_name_map* 是 **teacher_var name到student_var name的映射**，如果写反可能无法正确进行merge
@@ -37,7 +37,7 @@ with fluid.program_guard(teacher_program):
 data_name_map = {'y':'x'}
 USE_GPU = False
 place = fluid.CUDAPlace(0) if USE_GPU else fluid.CPUPlace()
-main_program = dist.merge(teacher_program, student_program,
+dist.merge(teacher_program, student_program,
                           data_name_map, place)
 ```
 
@@ -76,8 +76,8 @@ with fluid.program_guard(teacher_program):
 data_name_map = {'y':'x'}
 USE_GPU = False
 place = fluid.CUDAPlace(0) if USE_GPU else fluid.CPUPlace()
-main_program = merge(teacher_program, student_program, data_name_map, place)
-with fluid.program_guard(main_program):
+merge(teacher_program, student_program, data_name_map, place)
+with fluid.program_guard(student_program):
     distillation_loss = dist.fsp_loss('teacher_t1.tmp_1', 'teacher_t2.tmp_1',
                                       's1.tmp_1', 's2.tmp_1', main_program)
 ```
@@ -116,8 +116,8 @@ with fluid.program_guard(teacher_program):
 data_name_map = {'y':'x'}
 USE_GPU = False
 place = fluid.CUDAPlace(0) if USE_GPU else fluid.CPUPlace()
-main_program = merge(teacher_program, student_program, data_name_map, place)
-with fluid.program_guard(main_program):
+merge(teacher_program, student_program, data_name_map, place)
+with fluid.program_guard(student_program):
     distillation_loss = dist.l2_loss('teacher_t2.tmp_1', 's2.tmp_1',
                                      main_program)
 ```
@@ -158,8 +158,8 @@ with fluid.program_guard(teacher_program):
 data_name_map = {'y':'x'}
 USE_GPU = False
 place = fluid.CUDAPlace(0) if USE_GPU else fluid.CPUPlace()
-main_program = merge(teacher_program, student_program, data_name_map, place)
-with fluid.program_guard(main_program):
+merge(teacher_program, student_program, data_name_map, place)
+with fluid.program_guard(student_program):
     distillation_loss = dist.soft_label_loss('teacher_t2.tmp_1',
                                              's2.tmp_1', main_program, 1., 1.)
 ```
@@ -198,13 +198,13 @@ with fluid.program_guard(teacher_program):
 data_name_map = {'y':'x'}
 USE_GPU = False
 place = fluid.CUDAPlace(0) if USE_GPU else fluid.CPUPlace()
-main_program = merge(teacher_program, student_program, data_name_map, place)
+merge(teacher_program, student_program, data_name_map, place)
 def adaptation_loss(t_var, s_var):
     teacher_channel = t_var.shape[1]
     s_hint = fluid.layers.conv2d(s_var, teacher_channel, 1)
     hint_loss = fluid.layers.reduce_mean(fluid.layers.square(s_hint - t_var))
     return hint_loss
-with fluid.program_guard(main_program):
+with fluid.program_guard(student_program):
     distillation_loss = dist.loss(main_program, adaptation_loss,
             t_var='teacher_t2.tmp_1', s_var='s2.tmp_1')
 ```

docs/docs/tutorials/distillation_demo.md

@@ -86,7 +86,7 @@ merge过程操作较多，具体细节请参考[merge API文档](https://paddlep
 
 ```python
 data_name_map = {'data': 'image'}
-student_program = merge(teacher_program, student_program, data_name_map, place)
+merge(teacher_program, student_program, data_name_map, place)
 ```
 
 ### 5.添加蒸馏loss

paddleslim/dist/single_distiller.py

@@ -34,7 +34,6 @@ def merge(teacher_program,
                                                     paddle run on which device.
         scope(Scope): The input scope
         name_prefix(str): Name prefix added for all vars of the teacher program.
-    Return(Program): Merged program.
     """
     teacher_program = teacher_program.clone(for_test=True)
     for teacher_var in teacher_program.list_vars():
@@ -84,11 +83,13 @@ def merge(teacher_program,
                     attrs[attr_name] = op.attr(attr_name)
                 student_program.global_block().append_op(
                     type=op.type, inputs=inputs, outputs=outputs, attrs=attrs)
-    return student_program
 
 
-def fsp_loss(teacher_var1_name, teacher_var2_name, student_var1_name,
-             student_var2_name, program=fluid.default_main_program()):
+def fsp_loss(teacher_var1_name,
+             teacher_var2_name,
+             student_var1_name,
+             student_var2_name,
+             program=fluid.default_main_program()):
     """
     Combine variables from student model and teacher model by fsp-loss.
     Args:
@@ -115,7 +116,8 @@ def fsp_loss(teacher_var1_name, teacher_var2_name, student_var1_name,
     return fsp_loss
 
 
-def l2_loss(teacher_var_name, student_var_name,
+def l2_loss(teacher_var_name,
+            student_var_name,
             program=fluid.default_main_program()):
     """
     Combine variables from student model and teacher model by l2-loss.