Merge pull request #10354 from helinwang/scaffold

Improve trainer API

Author: helinwang

python/paddle/fluid/__init__.py

@@ -21,14 +21,15 @@
 from executor import *
 
 import trainer
-from trainer import *
+from trainer import Trainer
+from trainer import BeginEpochEvent
+from trainer import EndEpochEvent
+from trainer import BeginStepEvent
+from trainer import EndStepEvent
 
 import inferencer
 from inferencer import Inferencer
 
-import params
-from params import Params
-
 import io
 import evaluator
 import initializer
@@ -57,7 +58,7 @@
 Tensor = LoDTensor
 
 __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ +\
-          trainer.__all__ + inferencer.__all__ + params.__all__ + [
+          trainer.__all__ + inferencer.__all__ + [
     'io',
     'initializer',
     'layers',

python/paddle/fluid/inferencer.py

@@ -12,18 +12,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import core
+
 __all__ = ['Inferencer', ]
 
 
 class Inferencer(object):
-    def __init__(self, network_func, params, place=None):
+    def __init__(self, network_func, param_path=None, place=None):
         # 1. we need to generate a framework.Program by calling
         # network_func. Reference: fluid.program_guard in test_word2vec.py
 
         # 2. move the default_main_program to self.program.
 
         # 3. run the default_startup program.
-        self.params = params
+
+        # 4. load params from param_path into scope
+        self.scope = core.Scope()
         self.place = place
 
     def infer(self, inputs):

python/paddle/fluid/params.py

@@ -39,7 +39,7 @@ def create_random_lodtensor(lod, place, low, high):
 dict_size = len(word_dict)
 
 
-def inference_network(is_sparse):
+def inference_program(is_sparse):
     first_word = fluid.layers.data(name='firstw', shape=[1], dtype='int64')
     second_word = fluid.layers.data(name='secondw', shape=[1], dtype='int64')
     third_word = fluid.layers.data(name='thirdw', shape=[1], dtype='int64')
@@ -79,9 +79,9 @@ def inference_network(is_sparse):
     return predict_word
 
 
-def train_network(is_sparse):
+def train_program(is_sparse):
     next_word = fluid.layers.data(name='nextw', shape=[1], dtype='int64')
-    predict_word = inference_network(is_sparse)
+    predict_word = inference_program(is_sparse)
     cost = fluid.layers.cross_entropy(input=predict_word, label=next_word)
     avg_cost = fluid.layers.mean(cost)
     return avg_cost
@@ -100,23 +100,25 @@ def event_handler(event):
                 word_dict, N))
 
             if avg_cost < 5.0:
-                trainer.params.save(save_path)
+                trainer.save_params(save_path)
                 return
             if math.isnan(avg_cost):
                 sys.exit("got NaN loss, training failed.")
 
     trainer = fluid.Trainer(
-        partial(train_network, is_sparse),
+        partial(train_program, is_sparse),
         fluid.optimizer.SGD(learning_rate=0.001),
         place=place)
     trainer.train(
         reader=train_reader, num_epochs=100, event_handler=event_handler)
 
 
-def infer(use_cuda, save_path):
-    params = fluid.Params(save_path)
+def infer(use_cuda, is_sparse, save_path):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    inferencer = fluid.Inferencer(inference_network, params, place=place)
+    inferencer = fluid.Inferencer(
+        partial(inference_program, is_sparse),
+        param_path=save_path,
+        place=place)
 
     lod = [0, 1]
     first_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1)
@@ -138,7 +140,7 @@ def main(use_cuda, is_sparse):
 
     save_path = "word2vec.inference.model"
     train(use_cuda, is_sparse, save_path)
-    infer(use_cuda, save_path)
+    infer(use_cuda, is_sparse, save_path)
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/book/word2vec/no_test_word2vec_new_api.py

@@ -56,23 +56,22 @@ class Trainer(object):
     """
 
     Args:
-        network_func(callable): A function which will return loss. The loss must be a scaler.
+        program_func(callable): A function which will return loss. The loss must be a scaler.
         optimizer(optimizer.Optimizer): The optimizer should be an instance of Optimizer
-        params:
         place: The device place of this trainer.
     """
 
-    def __init__(self, network_func, optimizer, params=None, place=None):
+    def __init__(self, program_func, optimizer, param_path=None, place=None):
         # 1. we need to generate a framework.Program by calling
-        # network_func. Reference: fluid.program_guard in
+        # program_func. Reference: fluid.program_guard in
         # test_word2vec.py
-        self.scope = self._get_scope_from_params(params)
+        self.scope = core.Scope()
 
         self.startup_program = framework.Program()
         self.train_program = framework.Program()
 
         with framework.program_guard(self.train_program, self.startup_program):
-            loss = network_func()
+            loss = program_func()
             if not isinstance(optimizer, opt_module.Optimizer):
                 raise TypeError(
                     "The optimizer should be an instance of Optimizer")
@@ -84,14 +83,13 @@ def __init__(self, network_func, optimizer, params=None, place=None):
         # 2. move the default_main_program to self.program and run the
         # default_startup program on an empty core.Scope()
         # Run startup program
-        if params is None:
-            exe = executor.Executor(place)
-            exe.run(self.startup_program, scope=self.scope)
+        exe = executor.Executor(place)
+        exe.run(self.startup_program, scope=self.scope)
 
-        # 3. call self.params.add_vars with the initialized scope, it
-        # will add the new vars of the initialized scope into
-        # self.params.
-        # TODO(yuyang): This depends on parameters implementation.
+        if param_path:
+            # load params from param_path into scope
+            # TODO(yuyang): This depends on parameters implementation.
+            pass
 
         # TODO(helin): support distributed training
 
@@ -124,19 +122,9 @@ def train(self,
     def test(self, reader):
         pass
 
-    def _get_scope_from_params(self, params):
-        """
-        Get Scope from parameter object.
-        Args:
-            params(Parameter|None): The parameter object instance. Could be None.
-
-        Returns: New scope if params is None. Or params.scope()
-        NOTE: This method is WIP. Not fully implemented.
-        """
-        if params is None:
-            return core.Scope()  # new scope when params is None
-        else:
-            raise NotImplementedError("Not implemented right now.")
+    def save_params(self, param_path):
+        # reference: save_persistables in io.py
+        pass
 
     @staticmethod
     def _check_and_get_place(place):