Merge pull request #2384 from lcy-seso/fix_config_parsing_bug

fix bugs of parsing network topology in text generation task.

doc/api/v2/config/layer.rst

@@ -442,6 +442,14 @@ eos
 ..  autoclass:: paddle.v2.layer.eos
     :noindex:
 
+Miscs
+=====
+
+dropout
+--------------
+..  autoclass:: paddle.v2.layer.dropout
+    :noindex:
+
 Activation with learnable parameter
 ===================================
 
@@ -449,4 +457,3 @@ prelu
 --------
 ..  autoclass:: paddle.v2.layer.prelu
     :noindex:
-

doc/api/v2/config/networks.rst

@@ -125,11 +125,3 @@ simple_attention
     :members: simple_attention
     :noindex:
 
-Miscs
-=====
-
-dropout_layer
---------------
-..  automodule:: paddle.v2.networks
-    :members: dropout_layer
-    :noindex:

python/paddle/trainer/config_parser.py

@@ -3546,11 +3546,7 @@ def update_g_config():
     return g_config
 
 
-def begin_parse(config_arg_str=''):
-    '''
-    @param config_arg_str: a string of the form var1=val1,var2=val2. It will be
-    passed to config script as a dictionary CONFIG_ARGS
-    '''
+def begin_parse():
     init_config_environment()
     for hook in _parse_config_hooks:
         hook()
@@ -3568,8 +3564,12 @@ def begin_parse(config_arg_str=''):
 
 
 def parse_config(trainer_config, config_arg_str):
-    begin_parse(config_arg_str)
+    '''
+    @param config_arg_str: a string of the form var1=val1,var2=val2. It will be
+    passed to config script as a dictionary CONFIG_ARGS
+    '''
 
+    begin_parse()
     config_args = {}
 
     if config_arg_str:

python/paddle/trainer_config_helpers/layers.py

@@ -121,6 +121,7 @@ __all__ = [
     'smooth_l1_cost',
     'layer_support',
     'multiplex_layer',
+    'dropout_layer',
     'prelu_layer',
 ]
 
@@ -3771,7 +3772,6 @@ def beam_search(step,
     assert generated_input_index != -1
 
     gipt = input[generated_input_index]
-    assert isinstance(gipt, BaseGeneratedInput)
 
     gipt.bos_id = bos_id
     gipt.eos_id = eos_id
@@ -3791,7 +3791,6 @@ def beam_search(step,
         predict = gipt.after_real_step(step(*args))
 
         eos_layer(input=predict, eos_id=eos_id, name=eos_name)
-
         return predict
 
     tmp = recurrent_group(
@@ -5567,6 +5566,24 @@ def multiplex_layer(input, name=None, layer_attr=None):
         size=l.config.size)
 
 
+@wrap_name_default("dropout")
+def dropout_layer(input, dropout_rate, name=None):
+    """
+    @TODO(yuyang18): Add comments.
+
+    :param name:
+    :param input:
+    :param dropout_rate:
+    :return:
+    """
+    return addto_layer(
+        name=name,
+        input=input,
+        act=LinearActivation(),
+        bias_attr=False,
+        layer_attr=ExtraAttr(drop_rate=dropout_rate))
+
+
 @wrap_name_default()
 @layer_support()
 @wrap_name_default()

python/paddle/trainer_config_helpers/networks.py

@@ -26,10 +26,10 @@ from paddle.trainer.config_parser import *
 
 __all__ = [
     'sequence_conv_pool', 'simple_lstm', "simple_img_conv_pool",
-    "img_conv_bn_pool", 'dropout_layer', 'lstmemory_group', 'lstmemory_unit',
-    'small_vgg', 'img_conv_group', 'vgg_16_network', 'gru_unit', 'gru_group',
-    'simple_gru', 'simple_attention', 'simple_gru2', 'bidirectional_gru',
-    'text_conv_pool', 'bidirectional_lstm', 'inputs', 'outputs'
+    "img_conv_bn_pool", 'lstmemory_group', 'lstmemory_unit', 'small_vgg',
+    'img_conv_group', 'vgg_16_network', 'gru_unit', 'gru_group', 'simple_gru',
+    'simple_attention', 'simple_gru2', 'bidirectional_gru', 'text_conv_pool',
+    'bidirectional_lstm', 'inputs', 'outputs'
 ]
 
 ######################################################
@@ -1366,29 +1366,6 @@ def simple_attention(encoded_sequence,
         input=scaled, pooling_type=SumPooling(), name="%s_pooling" % name)
 
 
-############################################################################
-#                         Miscs                                            #
-############################################################################
-
-
-@wrap_name_default("dropout")
-def dropout_layer(input, dropout_rate, name=None):
-    """
-    @TODO(yuyang18): Add comments.
-
-    :param name:
-    :param input:
-    :param dropout_rate:
-    :return:
-    """
-    return addto_layer(
-        name=name,
-        input=input,
-        act=LinearActivation(),
-        bias_attr=False,
-        layer_attr=ExtraAttr(drop_rate=dropout_rate))
-
-
 def inputs(layers, *args):
     """
     Declare the inputs of network. The order of input should be as same as

python/paddle/v2/layer.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 """
 `paddle.v2.layer` is a part of model config packages in paddle.v2. In API v2,
-we want to make Paddle a plain Python package. The model config package defined
+we want to make Paddle a plain Python package. The model config package defines
 the way how to configure a neural network topology in Paddle Python code.
 
 The primary usage shows below.
@@ -30,7 +30,6 @@ The primary usage shows below.
     # use prediction instance where needed.
     parameters = paddle.parameters.create(cost)
 """
-
 import collections
 import copy
 import re
@@ -44,9 +43,10 @@ __all__ = ['data', 'parse_network']
 
 
 def __need_to_keep__(name):
-    if name in ['StaticInput', 'LayerType', 'layer_support']:
-        return False
-    return True
+    return name in [
+        'StaticInput', 'SubsequenceInput', 'GeneratedInput', 'LayerType',
+        'layer_support'
+    ]
 
 
 def __need_to_wrap__(name):
@@ -54,6 +54,8 @@ def __need_to_wrap__(name):
 
 
 def __convert_name__(inname):
+    if __need_to_keep__(inname):
+        return inname
     if inname == 'maxid_layer':
         return 'max_id'
     elif inname.endswith('memory') or inname.endswith(
@@ -74,8 +76,6 @@ def __convert_name__(inname):
 
 for name in v1_layers.__all__:
     obj = getattr(v1_layers, name)
-    if not __need_to_keep__(name):
-        continue
     new_name = __convert_name__(name)
     if callable(obj) and __need_to_wrap__(name):
         globals()[new_name] = __convert_to_v2__(obj, new_name, __name__)
@@ -107,7 +107,7 @@ __data_layer__.__doc__ = __map_data_docstr__(v1_layers.data_layer.__doc__)
 data = __convert_to_v2__(__data_layer__, 'name', __name__)
 
 
-def __get_used_layers__(output_layers, extra_layers=None):
+def __get_used_layers__(output_layers):
     layer_names = set()
     parents = {}
 
@@ -132,6 +132,13 @@ def __get_used_layers__(output_layers, extra_layers=None):
                     add_parent(mem.layer_name, mem.boot_layer_name)
                 add_parent(mem.link_name, mem.layer_name)
 
+            if sub_model.HasField('generator'):
+                # according to the implementation of text generation
+                # in recurrent layer group, the generated word must be
+                # the first out link
+                add_parent(sub_model.out_links[0].layer_name,
+                           sub_model.generator.eos_layer_name)
+
     def dfs_travel(layer_name):
         if layer_name in layer_names:
             return
@@ -247,8 +254,8 @@ def __trim_submodel__(old_submodel, layer_names, input_layer_names,
 def parse_network(output_layers, extra_layers=None):
     if not isinstance(output_layers, collections.Sequence):
         output_layers = [output_layers]
-    if extra_layers is not None and not isinstance(extra_layers,
-                                                   collections.Sequence):
+    if extra_layers is not None:
+        if not isinstance(extra_layers, collections.Sequence):
             extra_layers = [extra_layers]
     else:
         extra_layers = []
@@ -262,18 +269,29 @@ def parse_network(output_layers, extra_layers=None):
 
     model_config = ModelConfig()
     model_config.type = cp.g_config.model_config.type
+
+    for layer in output_layers:
+        model_config.output_layer_names.append(layer.full_name)
+        output_layer_names.add(layer.full_name)
+
     for l in cp.g_config.model_config.layers:
         if l.name not in layer_names:
             continue
         model_config.layers.extend([l])
         if l.type == 'data':
+            if l.name in model_config.output_layer_names:
+                """
+                In text generation, the outlink to save the generated word
+                indices is a data_layer defined in recurrent_group. This
+                data_layer is sure to be the output of the network in text
+                generation task, so this statement excludes such a special
+                data_layer from being inputs of the network, otherwise an error
+                will occur during data feeding.
+                """
+                continue
             model_config.input_layer_names.append(l.name)
             input_layer_names.add(l.name)
 
-    for layer in output_layers:
-        model_config.output_layer_names.append(layer.full_name)
-        output_layer_names.add(layer.full_name)
-
     for e in cp.g_config.model_config.evaluators:
         if e.name in evaluator_names:
             model_config.evaluators.extend([e])

python/paddle/v2/topology.py

@@ -31,7 +31,6 @@ class Topology(object):
     def __init__(self, layers, extra_layers=None):
         def __check__(layers):
             if not isinstance(layers, collections.Sequence):
-                __check_layer_type__(layers)
                 layers = [layers]
             for layer in layers:
                 __check_layer_type__(layer)
@@ -91,6 +90,7 @@ class Topology(object):
         [('image', dense_vector(768)), ('label', integer_value(10))]
         """
         data_layers = self.data_layers()
+
         return [(nm, data_layers[nm].data_type)
                 for nm in self.proto().input_layer_names]