Support StaticInput

python/paddle/v2/__init__.py

@@ -20,6 +20,7 @@ import event
 import data_type
 import topology
 import data_feeder
+import networks
 from . import dataset
 from . import reader
 import attr
@@ -29,7 +30,7 @@ import py_paddle.swig_paddle as api
 __all__ = [
     'optimizer', 'layer', 'activation', 'parameters', 'init', 'trainer',
     'event', 'data_type', 'attr', 'pooling', 'data_feeder', 'dataset', 'reader',
-    'topology'
+    'topology', 'networks'
 ]
 
 

python/paddle/v2/layer.py

@@ -109,9 +109,10 @@ def parse_network(*outputs):
 
 
 class Layer(object):
-    def __init__(self, name=None, parent_layers=None):
+    def __init__(self, name=None, size=None, parent_layers=None):
         assert isinstance(parent_layers, dict)
         self.name = name
+        self.size = size
         self.__parent_layers__ = parent_layers
 
     def to_proto(self, context):
@@ -173,7 +174,8 @@ def __convert_to_v2__(method_name, parent_names, is_default_name=True):
                     other_kwargs[key] = kwargs[key]
 
             name = kwargs.get('name', None)
-            super(V2LayerImpl, self).__init__(name, parent_layers)
+            size = kwargs.get('size', None)
+            super(V2LayerImpl, self).__init__(name, size, parent_layers)
             self.__other_kwargs__ = other_kwargs
 
         if wrapper is not None:
@@ -220,9 +222,10 @@ class WithExtraParent(Layer):
     def extra_parent(self):
         return self.__extra_parent__
 
-    def __init__(self, name=None, parent_layers=None):
+    def __init__(self, name=None, size=None, parent_layers=None):
         self.__extra_parent__ = []
-        super(WithExtraParent, self).__init__(name, parent_layers)
+        super(WithExtraParent, self).__init__(
+            name=name, size=size, parent_layers=parent_layers)
 
     def append_extra_parent(self, parent):
         self.__extra_parent__.append(parent)
@@ -261,7 +264,8 @@ class MemoryV2(WithExtraParent):
     def __init__(self, name, size, **kwargs):
         self.name = name
         self.size = size
-        super(MemoryV2, self).__init__(name=name, parent_layers=dict())
+        super(MemoryV2, self).__init__(
+            name=name, size=size, parent_layers=dict())
         self.__kwargs__ = kwargs
         self.__boot_layer_name__ = None
         if 'boot_layer' in kwargs:
@@ -271,7 +275,9 @@ class MemoryV2(WithExtraParent):
             st = inspect.stack()
             for i in xrange(len(st)):
                 locs = inspect.stack()[i][0].f_locals
-                for val in locs.viewvalues():
+                keys = locs.keys()
+                for key in keys:
+                    val = locs[key]
                     if isinstance(val, RecurrentLayerInput):
                         begin_of_current_rnn.append(val)
 
@@ -322,21 +328,15 @@ class LayerOutputV2(Layer):
         return self.layer_output
 
 
-class StaticInputV2(Layer):
-    def __init__(self, input=None, **kwargs):
-        assert input is not None
-        self.__kwargs__ = kwargs
-        super(StaticInputV2, self).__init__(
-            name=input.name, parent_layers={'input': input})
-
-    def context_name(self):
-        return self.name + "#static_input"
-
-    def to_proto_impl(self, **kwargs):
-        args = dict()
-        args.update(kwargs)
-        args.update(self.__kwargs__)
-        return conf_helps.StaticInput(**args)
+class StaticInputV2(object):
+    def __init__(self, input, is_seq=False, size=None):
+        assert isinstance(input, LayerV2)
+        self.name = input.name
+        self.input = input
+        self.is_seq = is_seq
+        self.size = size
+        # TODO(qiaolongfei): add size
+        # assert input.size is not None or size is not None
 
 
 class MixedLayerV2(Layer):
@@ -370,9 +370,8 @@ class MixedLayerV2(Layer):
         other_kwargs['act'] = act
         other_kwargs['bias_attr'] = bias_attr
         other_kwargs['layer_attr'] = layer_attr
-
         parent_layers = {"input": self.__inputs__}
-        super(MixedLayerV2, self).__init__(name, parent_layers)
+        super(MixedLayerV2, self).__init__(name, size, parent_layers)
         self.__other_kwargs__ = other_kwargs
 
     def __iadd__(self, other):
@@ -452,6 +451,12 @@ def recurrent_group(step, input, name=None):
     if not isinstance(input, collections.Sequence):
         input = [input]
 
+    # TODO(qiaolongfei) convert StaticInput to memory according to v2 recurrent_group
+    for i in xrange(len(input)):
+        cur_input = input[i]
+        if isinstance(cur_input, StaticInputV2):
+            input[i] = cur_input.input
+
     actual_input = [
         RecurrentLayerInput(
             recurrent_name=name,
@@ -512,7 +517,7 @@ def __layer_name_mapping_parent_names__(inname):
         lambda x: x in ['input1', 'input2', 'label', 'input', 'a', 'b',
                         'expand_as',
                         'weights', 'vectors', 'weight', 'score', 'left',
-                        'right'],
+                        'right', 'output_mem'],
         all_args)
 
 

python/paddle/v2/networks.py

+# Copyright (c) 2016 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 layer import __convert_to_v2__
+
+simple_gru = __convert_to_v2__('simple_gru', ['input'])
+simple_attention = __convert_to_v2__(
+    'simple_attention', ['encoded_sequence', 'encoded_proj', 'decoder_state'])

python/paddle/v2/tests/test_rnn_layer.py

@@ -74,9 +74,20 @@ class RNNTest(unittest.TestCase):
         label_dim = 3
 
         def parse_old_rnn():
+            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)
+                boot_layer = conf_helps.data_layer(name="boot", size=10)
+                boot_layer = conf_helps.fc_layer(
+                    name='boot_fc', input=boot_layer, size=10)
+
                 def step(y, wid):
                     z = conf_helps.embedding_layer(input=wid, size=word_dim)
-                mem = conf_helps.memory(name="rnn_state", size=hidden_dim)
+                    mem = conf_helps.memory(
+                        name="rnn_state",
+                        size=hidden_dim,
+                        boot_layer=boot_layer)
                     out = conf_helps.fc_layer(
                         input=[y, z, mem],
                         size=hidden_dim,
@@ -85,10 +96,6 @@ class RNNTest(unittest.TestCase):
                         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])
 
@@ -111,11 +118,9 @@ class RNNTest(unittest.TestCase):
             label = layer.data(
                 name="label", type=data_type.dense_vector(label_dim))
             emb = layer.embedding(input=data, size=word_dim)
-
             boot_layer = layer.data(
                 name="boot", type=data_type.dense_vector(10))
-
-            boot_layer = layer.fc(name='wtf', input=boot_layer, size=10)
+            boot_layer = layer.fc(name='boot_fc', input=boot_layer, size=10)
 
             def step(y, wid):
                 z = layer.embedding(input=wid, size=word_dim)
@@ -141,11 +146,9 @@ class RNNTest(unittest.TestCase):
 
             return str(layer.parse_network(cost))
 
-        with open("/Users/baidu/old.out", 'w') as f:
-            print >> f, parse_old_rnn()
-        with open("/Users/baidu/new.out", "w") as f:
-            print >> f, parse_new_rnn()
-        # print ''.join(diff)
+        diff = difflib.unified_diff(parse_old_rnn().splitlines(1),
+                                    parse_new_rnn().splitlines(1))
+        print ''.join(diff)
 
 
 if __name__ == '__main__':