Merge branch 'develop' into layer

python/paddle/trainer_config_helpers/layers.py

@@ -30,28 +30,90 @@ except ImportError:
 import copy
 
 __all__ = [
-    "full_matrix_projection", "AggregateLevel", "ExpandLevel",
-    "identity_projection", "dotmul_projection", "dotmul_operator",
-    "repeat_layer", "seq_reshape_layer", "table_projection", "mixed_layer",
-    "data_layer", "embedding_layer", "fc_layer", "grumemory", "pooling_layer",
-    "lstmemory", "last_seq", "first_seq", "cos_sim", "hsigmoid",
-    "conv_projection", "regression_cost", 'classification_cost', "LayerOutput",
-    'img_conv_layer', 'img_pool_layer', 'batch_norm_layer', 'img_cmrnorm_layer',
-    'addto_layer', 'concat_layer', 'seq_concat_layer', 'lstm_step_layer',
-    'recurrent_group', 'memory', 'StaticInput', 'expand_layer', 'scaling_layer',
-    'scaling_projection', 'power_layer', 'interpolation_layer',
-    'bilinear_interp_layer', 'trans_layer', 'rotate_layer',
-    'sum_to_one_norm_layer', 'get_output_layer', 'LayerType',
-    'context_projection', 'beam_search', 'maxid_layer', 'GeneratedInput',
-    'SubsequenceInput', 'gru_step_layer', 'recurrent_layer',
-    'BaseGeneratedInput', 'conv_operator', 'conv_shift_layer', 'tensor_layer',
-    'selective_fc_layer', 'sampling_id_layer', 'slope_intercept_layer',
-    'trans_full_matrix_projection', 'linear_comb_layer', 'convex_comb_layer',
-    'ctc_layer', 'warp_ctc_layer', 'crf_layer', 'crf_decoding_layer',
-    'nce_layer', 'cross_entropy_with_selfnorm', 'cross_entropy',
-    'multi_binary_label_cross_entropy', 'sum_cost', 'rank_cost', 'lambda_cost',
-    'huber_cost', 'block_expand_layer', 'maxout_layer', 'out_prod_layer',
-    'print_layer', 'priorbox_layer', 'spp_layer', 'pad_layer', 'eos_layer'
+    "full_matrix_projection",
+    "AggregateLevel",
+    "ExpandLevel",
+    "identity_projection",
+    "dotmul_projection",
+    "dotmul_operator",
+    "repeat_layer",
+    "seq_reshape_layer",
+    "table_projection",
+    "mixed_layer",
+    "data_layer",
+    "embedding_layer",
+    "fc_layer",
+    "grumemory",
+    "pooling_layer",
+    "lstmemory",
+    "last_seq",
+    "first_seq",
+    "cos_sim",
+    "hsigmoid",
+    "conv_projection",
+    "regression_cost",
+    'classification_cost',
+    "LayerOutput",
+    'img_conv_layer',
+    'img_pool_layer',
+    'batch_norm_layer',
+    'img_cmrnorm_layer',
+    'addto_layer',
+    'concat_layer',
+    'seq_concat_layer',
+    'lstm_step_layer',
+    'recurrent_group',
+    'memory',
+    'StaticInput',
+    'expand_layer',
+    'scaling_layer',
+    'scaling_projection',
+    'power_layer',
+    'interpolation_layer',
+    'bilinear_interp_layer',
+    'trans_layer',
+    'rotate_layer',
+    'sum_to_one_norm_layer',
+    'get_output_layer',
+    'LayerType',
+    'context_projection',
+    'beam_search',
+    'maxid_layer',
+    'GeneratedInput',
+    'SubsequenceInput',
+    'gru_step_layer',
+    'recurrent_layer',
+    'BaseGeneratedInput',
+    'conv_operator',
+    'conv_shift_layer',
+    'tensor_layer',
+    'selective_fc_layer',
+    'sampling_id_layer',
+    'slope_intercept_layer',
+    'trans_full_matrix_projection',
+    'linear_comb_layer',
+    'convex_comb_layer',
+    'ctc_layer',
+    'warp_ctc_layer',
+    'crf_layer',
+    'crf_decoding_layer',
+    'nce_layer',
+    'cross_entropy_with_selfnorm',
+    'cross_entropy',
+    'multi_binary_label_cross_entropy',
+    'sum_cost',
+    'rank_cost',
+    'lambda_cost',
+    'huber_cost',
+    'block_expand_layer',
+    'maxout_layer',
+    'out_prod_layer',
+    'print_layer',
+    'priorbox_layer',
+    'spp_layer',
+    'pad_layer',
+    'eos_layer',
+    'layer_support',
 ]
 
 
@@ -648,6 +710,7 @@ class MixedLayerType(LayerOutput):
         # update the size which might be computed inside MixedLayer
         # according to the operator's output size
         self.size = ml.config.size
+        self.finalized = True
 
 
 @wrap_name_default("mixed")

python/paddle/v2/layer.py

@@ -71,9 +71,15 @@ import collections
 import paddle.trainer_config_helpers as conf_helps
 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
+from paddle.trainer_config_helpers.default_decorators import wrap_act_default
+from paddle.trainer_config_helpers.default_decorators import wrap_bias_attr_default
+from paddle.trainer_config_helpers.layers import layer_support
 
 import data_type
+import activation
+import attr
 
 __all__ = [
     'parse_network', 'data', 'fc', 'conv_shift', 'img_conv', 'img_pool', 'spp',
@@ -89,6 +95,13 @@ __all__ = [
     'hsigmoid', 'eos'
 ]
 
+__projection_names__ = filter(lambda x: x.endswith('_projection'),
+                              dir(conf_helps))
+__all__ += __projection_names__
+
+__operator_names__ = filter(lambda x: x.endswith('_operator'), dir(conf_helps))
+__all__ += __operator_names__
+
 
 def parse_network(*outputs):
     """
@@ -106,9 +119,8 @@ def parse_network(*outputs):
 
 
 class Layer(object):
-    def __init__(self, name, parent_layers):
+    def __init__(self, name=None, parent_layers=None):
         assert isinstance(parent_layers, dict)
-        assert isinstance(name, basestring)
         self.name = name
         self.__parent_layers__ = parent_layers
 
@@ -127,19 +139,25 @@ class Layer(object):
                                self.__parent_layers__[layer_name])
             kwargs[layer_name] = v1_layer
 
-        if self.name not in context:
+        if self.name is None:
+            return self.to_proto_impl(**kwargs)
+        elif self.name not in context:
             context[self.name] = self.to_proto_impl(**kwargs)
+
         return context[self.name]
 
     def to_proto_impl(self, **kwargs):
         raise NotImplementedError()
 
 
-def __convert_to_v2__(method_name, parent_names):
+def __convert_to_v2__(method_name, parent_names, is_default_name=True):
+    if is_default_name:
         wrapper = wrap_name_default(name_prefix=method_name)
+    else:
+        wrapper = None
 
     class V2LayerImpl(Layer):
-        def __init__(self, name=None, **kwargs):
+        def __init__(self, **kwargs):
             parent_layers = dict()
             other_kwargs = dict()
             for pname in parent_names:
@@ -150,6 +168,7 @@ def __convert_to_v2__(method_name, parent_names):
                 if key not in parent_names:
                     other_kwargs[key] = kwargs[key]
 
+            name = kwargs.get('name', None)
             super(V2LayerImpl, self).__init__(name, parent_layers)
             self.__other_kwargs__ = other_kwargs
 
@@ -162,7 +181,7 @@ def __convert_to_v2__(method_name, parent_names):
                 args[each] = kwargs[each]
             for each in self.__other_kwargs__:
                 args[each] = self.__other_kwargs__[each]
-            return getattr(conf_helps, method_name)(name=self.name, **args)
+            return getattr(conf_helps, method_name)(**args)
 
     return V2LayerImpl
 
@@ -193,6 +212,78 @@ class DataLayerV2(Layer):
         return getattr(conf_helps, self.__method_name__)(name=self.name, **args)
 
 
+class MixedLayerV2(Layer):
+    """
+    This class is use to support `with` grammar. If not, the following code
+    could convert mixed_layer simply.
+
+        mixed = __convert_to_v2__(
+            'mixed_layer', name_prefix='mixed', parent_names=['input'])
+    """
+
+    class AddToSealedMixedLayerExceptionV2(Exception):
+        pass
+
+    def __init__(self,
+                 size=0,
+                 input=None,
+                 name=None,
+                 act=None,
+                 bias_attr=None,
+                 layer_attr=None):
+        self.__method_name__ = 'mixed_layer'
+        self.finalized = False
+        self.__inputs__ = []
+        if input is not None:
+            self.__inputs__ = input
+
+        other_kwargs = dict()
+        other_kwargs['name'] = name
+        other_kwargs['size'] = size
+        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)
+        self.__other_kwargs__ = other_kwargs
+
+    def __iadd__(self, other):
+        if not self.finalized:
+            self.__inputs__.append(other)
+            return self
+        else:
+            raise MixedLayerTypeV2.AddToSealedMixedLayerExceptionV2()
+
+    def __enter__(self):
+        assert len(self.__inputs__) == 0
+        return self
+
+    def __exit__(self, *args, **kwargs):
+        self.finalized = True
+
+    def to_proto_impl(self, **kwargs):
+        args = dict()
+        for each in kwargs:
+            args[each] = kwargs[each]
+        for each in self.__other_kwargs__:
+            args[each] = self.__other_kwargs__[each]
+        return getattr(conf_helps, self.__method_name__)(**args)
+
+
+@wrap_name_default("mixed")
+@wrap_act_default(act=activation.Linear())
+@wrap_bias_attr_default(has_bias=False)
+@layer_support(conf_helps.layers.ERROR_CLIPPING, conf_helps.layers.DROPOUT)
+def mixed(size=0,
+          name=None,
+          input=None,
+          act=None,
+          bias_attr=False,
+          layer_attr=None):
+    return MixedLayerV2(size, input, name, act, bias_attr, layer_attr)
+
+
 data = DataLayerV2
 AggregateLevel = conf_helps.layers.AggregateLevel
 ExpandLevel = conf_helps.layers.ExpandLevel
@@ -274,3 +365,18 @@ layer_list = [
 ]
 for l in layer_list:
     globals()[l[0]] = __convert_to_v2__(l[1], l[2])
+
+# convert projection
+for prj in __projection_names__:
+    globals()[prj] = __convert_to_v2__(
+        prj, parent_names=['input'], is_default_name=False)
+
+# convert operator
+operator_list = [
+    # [V1_method_name, parent_names],
+    ['dotmul_operator', ['a', 'b']],
+    ['conv_operator', ['img', 'filter']]
+]
+for op in operator_list:
+    globals()[op[0]] = __convert_to_v2__(
+        op[0], parent_names=op[1], is_default_name=False)

python/paddle/v2/tests/test_layer.py

@@ -167,5 +167,93 @@ class OtherLayerTest(unittest.TestCase):
         print layer.parse_network(pad)
 
 
+class ProjOpTest(unittest.TestCase):
+    def test_projection(self):
+        input = layer.data(name='data', type=data_type.dense_vector(784))
+        word = layer.data(
+            name='word', type=data_type.integer_value_sequence(10000))
+        fc0 = layer.fc(input=input, size=100, act=activation.Sigmoid())
+        fc1 = layer.fc(input=input, size=200, act=activation.Sigmoid())
+        mixed0 = layer.mixed(
+            size=256,
+            input=[
+                layer.full_matrix_projection(input=fc0),
+                layer.full_matrix_projection(input=fc1)
+            ])
+        with layer.mixed(size=200) as mixed1:
+            mixed1 += layer.full_matrix_projection(input=fc0)
+            mixed1 += layer.identity_projection(input=fc1)
+
+        table = layer.table_projection(input=word)
+        emb0 = layer.mixed(size=512, input=table)
+        with layer.mixed(size=512) as emb1:
+            emb1 += table
+
+        scale = layer.scaling_projection(input=fc0)
+        scale0 = layer.mixed(size=100, input=scale)
+        with layer.mixed(size=100) as scale1:
+            scale1 += scale
+
+        dotmul = layer.dotmul_projection(input=fc0)
+        dotmul0 = layer.mixed(size=100, input=dotmul)
+        with layer.mixed(size=100) as dotmul1:
+            dotmul1 += dotmul
+
+        context = layer.context_projection(input=fc0, context_len=5)
+        context0 = layer.mixed(size=100, input=context)
+        with layer.mixed(size=100) as context1:
+            context1 += context
+
+        conv = layer.conv_projection(
+            input=input,
+            filter_size=1,
+            num_channels=1,
+            num_filters=128,
+            stride=1,
+            padding=0)
+        conv0 = layer.mixed(input=conv, bias_attr=True)
+        with layer.mixed(bias_attr=True) as conv1:
+            conv1 += conv
+
+        print layer.parse_network(mixed0)
+        print layer.parse_network(mixed1)
+        print layer.parse_network(emb0)
+        print layer.parse_network(emb1)
+        print layer.parse_network(scale0)
+        print layer.parse_network(scale1)
+        print layer.parse_network(dotmul0)
+        print layer.parse_network(dotmul1)
+        print layer.parse_network(conv0)
+        print layer.parse_network(conv1)
+
+    def test_operator(self):
+        ipt0 = layer.data(name='data', type=data_type.dense_vector(784))
+        ipt1 = layer.data(name='word', type=data_type.dense_vector(128))
+        fc0 = layer.fc(input=ipt0, size=100, act=activation.Sigmoid())
+        fc1 = layer.fc(input=ipt0, size=100, act=activation.Sigmoid())
+
+        dotmul_op = layer.dotmul_operator(a=fc0, b=fc1)
+        dotmul0 = layer.mixed(input=dotmul_op)
+        with layer.mixed() as dotmul1:
+            dotmul1 += dotmul_op
+
+        conv = layer.conv_operator(
+            img=ipt0,
+            filter=ipt1,
+            filter_size=1,
+            num_channels=1,
+            num_filters=128,
+            stride=1,
+            padding=0)
+        conv0 = layer.mixed(input=conv)
+        with layer.mixed() as conv1:
+            conv1 += conv
+
+        print layer.parse_network(dotmul0)
+        print layer.parse_network(dotmul1)
+        print layer.parse_network(conv0)
+        print layer.parse_network(conv1)
+
+
 if __name__ == '__main__':
     unittest.main()