modify api_train_v2

demo/mnist/api_train_v2.py

@@ -26,7 +26,9 @@ def main():
                                 act=paddle.activation.Softmax())
     cost = paddle.layer.classification_cost(input=inference, label=label)
 
-    parameters = paddle.parameters.create(cost)
+    topology = paddle.topology.Topology(cost)
+
+    parameters = paddle.parameters.create(topology)
     for param_name in parameters.keys():
         array = parameters.get(param_name)
         array[:] = numpy.random.uniform(low=-1.0, high=1.0, size=array.shape)
@@ -45,16 +47,12 @@ def main():
 
     trainer = paddle.trainer.SGD(update_equation=adam_optimizer)
 
-    trainer.train(train_data_reader=train_reader,
-                  topology=cost,
+    trainer.train(
+        train_data_reader=train_reader,
+        topology=topology,
         parameters=parameters,
         event_handler=event_handler,
-                  batch_size=32,  # batch size should be refactor in Data reader
-                  data_types={  # data_types will be removed, It should be in
-                      # network topology
-                      'pixel': images.type,
-                      'label': label.type
-                  })
+        batch_size=32)  # batch size should be refactor in Data reader
 
 
 if __name__ == '__main__':

python/paddle/v2/layer.py

@@ -66,12 +66,14 @@ Also, the creation of a protobuf message is hidden in the invocation of
 paddle.v2.parameters.create, no longer exposed to users.
 """
 
+import collections
+
 import paddle.trainer_config_helpers as conf_helps
-from . import data_type as v2_data
 from paddle.trainer_config_helpers.config_parser_utils import \
     parse_network_config as __parse__
 from paddle.trainer_config_helpers.default_decorators import wrap_name_default
-import collections
+
+import data_type as v2_data
 
 __all__ = [
     'parse_network', 'data', 'fc', 'max_id', 'classification_cost',
@@ -184,6 +186,8 @@ class DataLayerV2(Layer):
         return getattr(conf_helps, self.__method_name__)(name=self.name, **args)
 
 
+LayerV2 = Layer
+
 data = DataLayerV2
 fc = __convert_to_v2__('fc_layer', name_prefix='fc', parent_names=['input'])
 max_id = __convert_to_v2__(

python/paddle/v2/parameters.py

 import numpy as np
-from . import layer as v2_layer
 import py_paddle.swig_paddle as api
 from paddle.proto.ParameterConfig_pb2 import ParameterConfig
 
+import topology as v2_topology
+
 __all__ = ['Parameters', 'create']
 
 
-def create(*layers):
+def create(topology):
     """
-    Create parameter pool by layers. In paddle, layer can be represent a
-    model config.
-
-    :param layers:
+    Create parameter pool by topology.
+    :param topology:
     :return:
     """
-    for layer in layers:
-        if not isinstance(layer, v2_layer.Layer):
+    if not isinstance(topology, v2_topology.Topology):
         raise ValueError(
-                'create must pass a topologies which type is paddle.layer.Layer')
-    model_config = v2_layer.parse_network(*layers)
+            'create must pass a topology which type is topology.Topology')
     pool = Parameters()
-    for param in model_config.parameters:
+    for param in topology.proto().parameters:
         pool.__append_config__(param)
     return pool
 

python/paddle/v2/topology.py

@@ -12,7 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from . import layer
+from paddle.proto.ModelConfig_pb2 import ModelConfig
+import paddle.trainer_config_helpers as conf_helps
+import layer as v2_layer
+import data_type
 
 __all__ = ['Topology']
 
@@ -23,22 +26,101 @@ class Topology(object):
     and network configs.
     """
 
-    def __init__(self, cost):
-        self.cost = cost
-        self.__model_config__ = layer.parse_network(cost)
+    def __init__(self, *layers):
+        for layer in layers:
+            if not isinstance(layer, v2_layer.LayerV2):
+                raise ValueError('create must pass a topologies '
+                                 'which type is paddle.layer.Layer')
+        self.layers = layers
+        self.__model_config__ = v2_layer.parse_network(*layers)
+        assert isinstance(self.__model_config__, ModelConfig)
 
-    def __call__(self):
+    def proto(self):
         return self.__model_config__
 
     def get_layer(self, name):
+        """
+        get v2.Layer Class instance by layer name
+        :param name:
+        :return:
+        """
+        result_layer = []
+
+        def find_layer_by_name(layer, layer_name):
+            if layer.name == layer_name and len(result_layer) == 0:
+                result_layer.append(layer)
+            for parent_layer in layer.__parent_layers__.values():
+                find_layer_by_name(parent_layer, layer_name)
+
+        for layer in self.layers:
+            find_layer_by_name(layer, name)
+
+        return result_layer[0]
+
+    def get_data_layer(self):
+        """
+        get all data layer
+        :return:
+        """
+        data_layers = []
+
+        def find_data_layer(layer):
+            assert isinstance(layer, layer.LayerV2)
+            if isinstance(layer, v2_layer.DataLayerV2):
+                if len(
+                        filter(lambda data_layer: data_layer.name == layer.name,
+                               data_layers)) == 0:
+                    data_layers.append(layer)
+            for parent_layer in layer.__parent_layers__.values():
+                find_data_layer(parent_layer)
+
+        for layer in self.layers:
+            find_data_layer(layer)
+
+        return data_layers
+
+    def get_layer_proto(self, name):
         """
         get layer by layer name
         :param name:
         :return:
         """
-        pass
+        layers = filter(lambda layer: layer.name == name,
+                        self.__model_config__.layers)
+        if len(layers) is 1:
+            return layers[0]
+        else:
+            return None
 
     def data_type(self):
         """
+        get data_type from proto, such as:
+        [('image', dense_vector(768)), ('label', integer_value(10))]
+        the order is the same with __model_config__.input_layer_names
         """
-        pass
+        data_types_lists = []
+        for layer_name in self.__model_config__.input_layer_names:
+            data_types_lists.append(
+                (layer_name, self.get_layer(layer_name).type))
+
+        return data_types_lists
+
+
+if __name__ == '__main__':
+    pixel = v2_layer.data(name='pixel', type=data_type.dense_vector(784))
+    label = v2_layer.data(name='label', type=data_type.integer_value(10))
+    hidden = v2_layer.fc(input=pixel,
+                         size=100,
+                         act=conf_helps.SigmoidActivation())
+    inference = v2_layer.fc(input=hidden,
+                            size=10,
+                            act=conf_helps.SoftmaxActivation())
+    maxid = v2_layer.max_id(input=inference)
+    cost1 = v2_layer.classification_cost(input=inference, label=label)
+    cost2 = v2_layer.cross_entropy_cost(input=inference, label=label)
+
+    print Topology(cost1).proto()
+    print Topology(cost2).proto()
+    print Topology(cost1, cost2).proto()
+    print Topology(cost2).proto()
+    print Topology(inference, maxid).proto()

python/paddle/v2/trainer.py

 import collections
 
 import py_paddle.swig_paddle as api
-from paddle.proto.ModelConfig_pb2 import ModelConfig
 from py_paddle import DataProviderConverter
 
 from . import event as v2_event
-from . import layer as v2_layer
 from . import optimizer as v2_optimizer
 from . import parameters as v2_parameters
+from . import topology as v2_topology
 
 __all__ = ['ITrainer', 'SGD']
 
@@ -88,12 +87,11 @@ class SGD(ITrainer):
         if event_handler is None:
             event_handler = default_event_handler
 
-        topology = v2_layer.parse_network(topology)
-
         __check_train_args__(**locals())
 
         gm = api.GradientMachine.createFromConfigProto(
-            topology, api.CREATE_MODE_NORMAL, self.__optimizer__.enable_types())
+            topology.proto(), api.CREATE_MODE_NORMAL,
+            self.__optimizer__.enable_types())
         assert isinstance(gm, api.GradientMachine)
         parameters.append_gradient_machine(gm)
 
@@ -102,13 +100,7 @@ class SGD(ITrainer):
 
         gm.start()
         out_args = api.Arguments.createArguments(0)
-
-        data_types_lists = []
-        for each in topology.input_layer_names:
-            if each not in data_types:
-                raise ValueError()
-            data_types_lists.append(data_types[each])
-
+        data_types_lists = [data_type[1] for data_type in topology.data_type()]
         converter = DataProviderConverter(input_types=data_types_lists)
 
         for pass_id in xrange(num_passes):
@@ -141,7 +133,7 @@ def __data_reader_to_batch__(reader, batch_size, topology):
     def input_reorder(func):
         for item in func():
             retv = []
-            for __layer_name__ in topology.input_layer_names:
+            for __layer_name__ in topology.proto().input_layer_names:
                 retv.append(item[__layer_name__])
             yield retv
 
@@ -178,7 +170,7 @@ def __check_train_args__(train_data_reader, topology, parameters,
             raise ValueError('test_data_reader should be a function, which can '
                              'return a iterator')
 
-    if not isinstance(topology, ModelConfig):
+    if not isinstance(topology, v2_topology.Topology):
         raise ValueError('topology should be a model config')
 
     if not isinstance(parameters, v2_parameters.Parameters):