Merge pull request #1398 from luotao1/reshape

add SequenceReshapeLayer in trainer_config_helpers

doc/api/trainer_config_helpers/layers.rst

@@ -308,6 +308,12 @@ repeat_layer
     :members: repeat_layer
     :noindex:
 
+seq_reshape_layer
+-----------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: seq_reshape_layer
+    :noindex:
+
 Math Layers
 ===========
 

paddle/gserver/layers/SequenceReshapeLayer.cpp

@@ -20,9 +20,12 @@ limitations under the License. */
 namespace paddle {
 
 /**
- * A layer for reshaping the sequence
- * Input: a sequence
- * Output: a sequence
+ *  A layer for reshaping the sequence. Assume the input sequence has
+ *  T instances, the dimension of each instance is M, and the input
+ *  reshape_dim is N, then the output sequence has T*M/N instances,
+ *  the dimension of each instance is N.
+ *
+ *  Note that T*M/N must be an integer.
  */
 
 class SequenceReshapeLayer : public Layer {

python/paddle/trainer_config_helpers/layers.py

@@ -37,6 +37,7 @@ __all__ = [
     "dotmul_projection",
     "dotmul_operator",
     "repeat_layer",
+    "seq_reshape_layer",
     "table_projection",
     "mixed_layer",
     "data_layer",
@@ -125,6 +126,7 @@ class LayerType(object):
     GRUMEMORY = "gated_recurrent"
     SEQUENCE_LAST_INSTANCE = "seqlastins"
     SEQUENCE_FIRST_INSTANCE = "seqfirstins"
+    SEQUENCE_RESHAPE = "seqreshape"
     POOLING_MAX = "max"
     POOLING_AVG = 'average'
     FC_LAYER = "fc"
@@ -1450,6 +1452,61 @@ def repeat_layer(input, num_repeats, name=None, layer_attr=None):
         parents=[input])
 
 
+@wrap_name_default("seqreshape")
+@wrap_act_default(act=IdentityActivation())
+@wrap_bias_attr_default(has_bias=False)
+@layer_support()
+def seq_reshape_layer(input,
+                      reshape_size,
+                      act=None,
+                      name=None,
+                      layer_attr=None,
+                      bias_attr=None):
+    """
+    A layer for reshaping the sequence. Assume the input sequence has T instances,
+    the dimension of each instance is M, and the input reshape_size is N, then the 
+    output sequence has T*M/N instances, the dimension of each instance is N.
+
+    Note that T*M/N must be an integer.
+
+    The example usage is:
+
+    .. code-block:: python
+
+       reshape = seq_reshape_layer(input=layer, reshape_size=4)
+
+    :param input: Input layer.
+    :type input: LayerOutput
+    :param reshape_size: the size of reshaped sequence.
+    :type reshape_size: int
+    :param name: Layer name.
+    :type name: basestring
+    :param act: Activation type.
+    :type act: BaseActivation
+    :param layer_attr: extra layer attributes.
+    :type layer_attr: ExtraLayerAttribute.
+    :param bias_attr: The Bias Attribute. If no bias, then pass False or
+                      something not type of ParameterAttribute. None will get a
+                      default Bias.
+    :type bias_attr: ParameterAttribute or None or bool
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+
+    Layer(
+        inputs=[input.name],
+        name=name,
+        size=reshape_size,
+        type=LayerType.SEQUENCE_RESHAPE,
+        bias=ParamAttr.to_bias(bias_attr),
+        **ExtraAttr.to_kwargs(layer_attr))
+    return LayerOutput(
+        name=name,
+        size=reshape_size,
+        layer_type=LayerType.SEQUENCE_RESHAPE,
+        parents=[input])
+
+
 @wrap_name_default()
 @layer_support()
 def interpolation_layer(input, weight, name=None, layer_attr=None):
@@ -2604,6 +2661,10 @@ def seq_concat_layer(a, b, act=None, name=None, layer_attr=None,
     :type act: BaseActivation
     :param layer_attr: Extra Layer Attribute.
     :type layer_attr: ExtraLayerAttribute
+    :param bias_attr: The Bias Attribute. If no bias, then pass False or
+                      something not type of ParameterAttribute. None will get a
+                      default Bias.
+    :type bias_attr: ParameterAttribute or None or bool
     :return: LayerOutput object.
     :rtype: LayerOutput
     """

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -5,6 +5,6 @@ last_first_seq test_expand_layer test_ntm_layers test_hsigmoid
 img_layers img_trans_layers util_layers simple_rnn_layers unused_layers test_cost_layers
 test_rnn_group shared_fc shared_lstm shared_gru test_cost_layers_with_weight
 test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
-test_seq_concat)
+test_seq_concat_reshape)
 
 export whole_configs=(test_split_datasource)

python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_concat.protostr → python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_concat_reshape.protostr

@@ -23,17 +23,29 @@ layers {
     input_layer_name: "data2"
   }
 }
+layers {
+  name: "__seqreshape_0__"
+  type: "seqreshape"
+  size: 5
+  active_type: "linear"
+  inputs {
+    input_layer_name: "data1"
+  }
+}
 input_layer_names: "data1"
 input_layer_names: "data2"
 output_layer_names: "__seqconcat_0__"
+output_layer_names: "__seqreshape_0__"
 sub_models {
   name: "root"
   layer_names: "data1"
   layer_names: "data2"
   layer_names: "__seqconcat_0__"
+  layer_names: "__seqreshape_0__"
   input_layer_names: "data1"
   input_layer_names: "data2"
   output_layer_names: "__seqconcat_0__"
+  output_layer_names: "__seqreshape_0__"
   is_recurrent_layer_group: false
 }
 

python/paddle/trainer_config_helpers/tests/configs/test_seq_concat.py → python/paddle/trainer_config_helpers/tests/configs/test_seq_concat_reshape.py

@@ -3,7 +3,10 @@ from paddle.trainer_config_helpers import *
 settings(batch_size=1000, learning_rate=1e-5)
 
 din1 = data_layer(name='data1', size=30)
-
 din2 = data_layer(name='data2', size=30)
 
-outputs(seq_concat_layer(a=din1, b=din2))
+opts = []
+opts.append(seq_concat_layer(a=din1, b=din2))
+opts.append(seq_reshape_layer(input=din1, reshape_size=5))
+
+outputs(opts)