support sequence_rnn_multi_input

demo/mnist/api_train_v2.py

@@ -3,9 +3,6 @@ import paddle.v2 as paddle
 
 import mnist_util
 
-import pudb
-pudb.set_trace()
-
 
 def train_reader():
     train_file = './data/raw_data/train'

python/paddle/trainer/config_parser.py

@@ -3474,6 +3474,8 @@ def update_g_config():
     for name in g_config.model_config.output_layer_names:
         assert name in g_layer_map, \
             'input name "%s" does not correspond to a layer name' % name
+    for hook in _parse_config_hooks:
+        hook()
     return g_config
 
 
@@ -3485,8 +3487,8 @@ def parse_config(trainer_config, config_arg_str):
     passed to config script as a dictionary CONFIG_ARGS
     '''
     init_config_environment()
-    for hook in _parse_config_hooks:
-        hook()
+    # for hook in _parse_config_hooks:
+    #     hook()
 
     config_args = {}
 

python/paddle/v2/layer.py

@@ -124,11 +124,13 @@ class Layer(object):
             return self.to_proto_impl(**kwargs)
 
         # memory may have the same name with some layer
-        if isinstance(self, MemoryV2) or isinstance(self, LayerOutputV2):
+        if isinstance(self, MemoryV2):
             return self.to_proto_impl(**kwargs)
 
+        # store v1 API's layer_output in context with the key of it's name.
         if self.name not in context:
             context[self.name] = self.to_proto_impl(**kwargs)
+
         return context[self.name]
 
     def to_proto_impl(self, **kwargs):
@@ -200,8 +202,19 @@ class MemoryV2(Layer):
     def __init__(self, name, size, **kwargs):
         self.name = name
         self.size = size
-        self.__kwargs__ = kwargs
-        super(MemoryV2, self).__init__(name=name, parent_layers=dict())
+
+        parent_names = ['boot_layer']
+        parent_layers = dict()
+        other_kwargs = dict()
+        for pname in parent_names:
+            if kwargs.has_key(pname):
+                parent_layers[pname] = kwargs[pname]
+
+        for key in kwargs.keys():
+            if key not in parent_names:
+                other_kwargs[key] = kwargs[key]
+        super(MemoryV2, self).__init__(name=name, parent_layers=parent_layers)
+        self.__kwargs__ = other_kwargs
 
     def to_proto_impl(self, **kwargs):
         args = dict()
@@ -209,10 +222,16 @@ class MemoryV2(Layer):
             args[each] = kwargs[each]
         for each in self.__kwargs__:
             args[each] = self.__kwargs__[each]
+
         return conf_helps.memory(name=self.name, size=self.size, **args)
 
 
 class LayerOutputV2(Layer):
+    """
+    LayerOutputV2 is used to store the result of LayerOutput in v1 api.
+    It will not store it's parents because layer_output has been parsed already.
+    """
+
     def __init__(self, layer_output):
         assert isinstance(layer_output, conf_helps.LayerOutput)
         self.layer_output = layer_output
@@ -239,8 +258,11 @@ class RecurrentGroupV2(Layer):
         super(RecurrentGroupV2, self).__init__(
             name=name, parent_layers=parent_layers)
 
+    wrapper = wrap_name_default(name_prefix='recurrent_group')
+    __init__ = wrapper(__init__)
+
     def to_proto_impl(self, **kwargs):
-        def in_args_converter(in_args):
+        def in_args_converter(*in_args):
             if not isinstance(in_args, collections.Sequence):
                 in_args = [in_args]
             return [LayerOutputV2(input) for input in in_args]

python/paddle/v2/tests/CMakeLists.txt

@@ -1,3 +1,7 @@
 add_test(NAME test_v2_layer
         COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
-        ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/v2/tests/test_layer.py
+        ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/v2/tests/test_layer.py)
+
+add_test(NAME test_v2_rnn_layer
+        COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
+        ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/v2/tests/test_rnn_layer.py)

python/paddle/v2/tests/test_layer.py

@@ -11,16 +11,12 @@
 # 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 difflib
 import unittest
 
-import paddle.trainer_config_helpers as conf_helps
 import paddle.v2.activation as activation
 import paddle.v2.attr as attr
 import paddle.v2.data_type as data_type
 import paddle.v2.layer as layer
-from paddle.trainer_config_helpers.config_parser_utils import \
-    parse_network_config as parse_network
 
 pixel = layer.data(name='pixel', type=data_type.dense_vector(784))
 label = layer.data(name='label', type=data_type.integer_value(10))
@@ -57,51 +53,5 @@ class CostLayerTest(unittest.TestCase):
         print layer.parse_network(cost7, cost8, cost9, cost10, cost11)
 
 
-class RNNTest(unittest.TestCase):
-    def test_simple_rnn(self):
-        dict_dim = 10
-        word_dim = 8
-        hidden_dim = 8
-
-        def parse_old_rnn():
-            def step(y):
-                mem = conf_helps.memory(name="rnn_state", size=hidden_dim)
-                out = conf_helps.fc_layer(
-                    input=[y, mem],
-                    size=hidden_dim,
-                    act=activation.Tanh(),
-                    bias_attr=True,
-                    name="rnn_state")
-                return out
-
-            def test():
-                data1 = conf_helps.data_layer(name="word", size=dict_dim)
-                embd = conf_helps.embedding_layer(input=data1, size=word_dim)
-                conf_helps.recurrent_group(name="rnn", step=step, input=embd)
-
-            return str(parse_network(test))
-
-        def parse_new_rnn():
-            def new_step(y):
-                mem = layer.memory(name="rnn_state", size=hidden_dim)
-                out = layer.fc(input=[y, mem],
-                               size=hidden_dim,
-                               act=activation.Tanh(),
-                               bias_attr=True,
-                               name="rnn_state")
-                return out
-
-            data1 = layer.data(
-                name="word", type=data_type.integer_value(dict_dim))
-            embd = layer.embedding(input=data1, size=word_dim)
-            rnn_layer = layer.recurrent_group(
-                name="rnn", step=new_step, input=embd)
-            return str(layer.parse_network(rnn_layer))
-
-        diff = difflib.unified_diff(parse_old_rnn().splitlines(1),
-                                    parse_new_rnn().splitlines(1))
-        print ''.join(diff)
-
-
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/tests/test_rnn_layer.py

@@ -0,0 +1,143 @@
+# Copyright PaddlePaddle contributors. 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 difflib
+import unittest
+
+import paddle.trainer_config_helpers as conf_helps
+import paddle.v2.activation as activation
+import paddle.v2.data_type as data_type
+import paddle.v2.layer as layer
+from paddle.trainer_config_helpers.config_parser_utils import \
+    parse_network_config as parse_network
+
+
+class RNNTest(unittest.TestCase):
+    def test_simple_rnn(self):
+        dict_dim = 10
+        word_dim = 8
+        hidden_dim = 8
+
+        def parse_old_rnn():
+            def step(y):
+                mem = conf_helps.memory(name="rnn_state", size=hidden_dim)
+                out = conf_helps.fc_layer(
+                    input=[y, mem],
+                    size=hidden_dim,
+                    act=activation.Tanh(),
+                    bias_attr=True,
+                    name="rnn_state")
+                return out
+
+            def test():
+                data = conf_helps.data_layer(name="word", size=dict_dim)
+                embd = conf_helps.embedding_layer(input=data, size=word_dim)
+                conf_helps.recurrent_group(name="rnn", step=step, input=embd)
+
+            return str(parse_network(test))
+
+        def parse_new_rnn():
+            def new_step(y):
+                mem = layer.memory(name="rnn_state", size=hidden_dim)
+                out = layer.fc(input=[y, mem],
+                               size=hidden_dim,
+                               act=activation.Tanh(),
+                               bias_attr=True,
+                               name="rnn_state")
+                return out
+
+            data = layer.data(
+                name="word", type=data_type.integer_value(dict_dim))
+            embd = layer.embedding(input=data, size=word_dim)
+            rnn_layer = layer.recurrent_group(
+                name="rnn", step=new_step, input=embd)
+            return str(layer.parse_network(rnn_layer))
+
+        diff = difflib.unified_diff(parse_old_rnn().splitlines(1),
+                                    parse_new_rnn().splitlines(1))
+        print ''.join(diff)
+
+    def test_sequence_rnn_multi_input(self):
+        dict_dim = 10
+        word_dim = 8
+        hidden_dim = 8
+        label_dim = 3
+
+        def parse_old_rnn():
+            def step(y, wid):
+                z = conf_helps.embedding_layer(input=wid, size=word_dim)
+                mem = conf_helps.memory(name="rnn_state", size=hidden_dim)
+                out = conf_helps.fc_layer(
+                    input=[y, z, mem],
+                    size=hidden_dim,
+                    act=conf_helps.TanhActivation(),
+                    bias_attr=True,
+                    name="rnn_state")
+                return out
+
+            def test():
+                data = conf_helps.data_layer(name="word", size=dict_dim)
+                label = conf_helps.data_layer(name="label", size=label_dim)
+                emb = conf_helps.embedding_layer(input=data, size=word_dim)
+                out = conf_helps.recurrent_group(
+                    name="rnn", step=step, input=[emb, data])
+
+                rep = conf_helps.last_seq(input=out)
+                prob = conf_helps.fc_layer(
+                    size=label_dim,
+                    input=rep,
+                    act=conf_helps.SoftmaxActivation(),
+                    bias_attr=True)
+
+                conf_helps.outputs(
+                    conf_helps.classification_cost(
+                        input=prob, label=label))
+
+            return str(parse_network(test))
+
+        def parse_new_rnn():
+            def step(y, wid):
+                z = layer.embedding(input=wid, size=word_dim)
+                mem = layer.memory(name="rnn_state", size=hidden_dim)
+                out = layer.fc(input=[y, z, mem],
+                               size=hidden_dim,
+                               act=activation.Tanh(),
+                               bias_attr=True,
+                               name="rnn_state")
+                return out
+
+            data = layer.data(
+                name="word", type=data_type.dense_vector(dict_dim))
+            label = layer.data(
+                name="label", type=data_type.dense_vector(label_dim))
+            emb = layer.embedding(input=data, size=word_dim)
+            out = layer.recurrent_group(
+                name="rnn", step=step, input=[emb, data])
+
+            rep = layer.last_seq(input=out)
+            prob = layer.fc(size=label_dim,
+                            input=rep,
+                            act=activation.Softmax(),
+                            bias_attr=True)
+
+            cost = layer.classification_cost(input=prob, label=label)
+
+            return str(layer.parse_network(cost))
+
+        diff = difflib.unified_diff(parse_old_rnn().splitlines(1),
+                                    parse_new_rnn().splitlines(1))
+        print ''.join(diff)
+
+
+if __name__ == '__main__':
+    unittest.main()