Merge pull request #2480 from emailweixu/repeat_layer

Repeat layer for column vector

paddle/gserver/layers/FeatureMapExpandLayer.cpp

@@ -40,6 +40,7 @@ namespace paddle {
 class FeatureMapExpandLayer : public Layer {
 private:
   int numFilters_;
+  bool asRowVector_;
 
 public:
   explicit FeatureMapExpandLayer(const LayerConfig& config) : Layer(config) {}
@@ -62,6 +63,7 @@ bool FeatureMapExpandLayer::init(const LayerMap& layerMap,
 
   CHECK_EQ(inputLayers_.size(), 1UL);
   numFilters_ = config_.num_filters();
+  asRowVector_ = config_.user_arg() != "as_col_vec";
   return true;
 }
 
@@ -76,6 +78,7 @@ void FeatureMapExpandLayer::forward(PassType passType) {
 
   {
     AsyncGpuBlock asyncGpuBlock;
+    if (asRowVector_) {
       for (size_t i = 0; i < batchSize; i++) {
         MatrixPtr outVTmp =
             Matrix::create(outputV->getData() + i * imgSize * numFilters_,
@@ -87,6 +90,19 @@ void FeatureMapExpandLayer::forward(PassType passType) {
             inputV->getData() + i * imgSize, 1, imgSize, false, useGpu_);
         outVTmp->addRowVector(*inVTmp);
       }
+    } else {
+      for (size_t i = 0; i < batchSize; i++) {
+        MatrixPtr outVTmp =
+            Matrix::create(outputV->getData() + i * imgSize * numFilters_,
+                           imgSize,
+                           numFilters_,
+                           false,
+                           useGpu_);
+        MatrixPtr inVTmp = Matrix::create(
+            inputV->getData() + i * imgSize, imgSize, 1, false, useGpu_);
+        outVTmp->addColVector(*inVTmp);
+      }
+    }
   }
   /* activation */ {
     REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str());
@@ -102,8 +118,13 @@ void FeatureMapExpandLayer::backward(const UpdateCallback& callback) {
   MatrixPtr outGrad = getOutputGrad();
   size_t batchSize = getInput(0).getBatchSize();
   int imgSize = inGrad->getWidth();
+  /* Do activation */ {
+    REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str());
+    backwardActivation();
+  }
   {
     AsyncGpuBlock asyncGpuBlock;
+    if (asRowVector_) {
       for (size_t i = 0; i < batchSize; i++) {
         MatrixPtr outGradTmp =
             Matrix::create(outGrad->getData() + i * imgSize * numFilters_,
@@ -115,10 +136,19 @@ void FeatureMapExpandLayer::backward(const UpdateCallback& callback) {
             inGrad->getData() + i * imgSize, 1, imgSize, false, useGpu_);
         inGradTmp->collectBias(*outGradTmp, 1);
       }
+    } else {
+      for (size_t i = 0; i < batchSize; i++) {
+        MatrixPtr outGradTmp =
+            Matrix::create(outGrad->getData() + i * imgSize * numFilters_,
+                           imgSize,
+                           numFilters_,
+                           false,
+                           useGpu_);
+        MatrixPtr inGradTmp = Matrix::create(
+            inGrad->getData() + i * imgSize, imgSize, 1, false, useGpu_);
+        inGradTmp->sumRows(*outGradTmp, 1, 1);
+      }
     }
-  /* Do derivation */ {
-    REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str());
-    backwardActivation();
   }
 }
 

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1598,6 +1598,8 @@ TEST(Layer, FeatureMapExpandLayer) {
                               /* paraSize= */ 0});
   config.layerConfig.add_inputs();
   for (auto useGpu : {false, true}) {
+    for (auto asRowVec : {false, true}) {
+      config.layerConfig.set_user_arg(asRowVec ? "as_row_vec" : "as_col_vec");
       testLayerGrad(config,
                     "featmap_expand",
                     /*batch_size*/ 100,
@@ -1605,6 +1607,7 @@ TEST(Layer, FeatureMapExpandLayer) {
                     useGpu,
                     /* useWeight */ true);
     }
+  }
 }
 
 TEST(Layer, MultiplexLayer) {

python/paddle/trainer/config_parser.py

@@ -1926,7 +1926,6 @@ class BatchNormLayer(LayerBase):
     def __init__(self,
                  name,
                  inputs,
-                 active_type="linear",
                  bias=True,
                  use_global_stats=True,
                  moving_average_fraction=0.9,
@@ -1964,12 +1963,7 @@ class BatchNormLayer(LayerBase):
             cudnn_version >= 4007
         self.layer_type = "cudnn_batch_norm" if use_cudnn else "batch_norm"
         super(BatchNormLayer, self).__init__(
-            name,
-            self.layer_type,
-            0,
-            active_type=active_type,
-            inputs=inputs,
-            **xargs)
+            name, self.layer_type, 0, inputs=inputs, **xargs)
 
         if use_global_stats is not None:
             self.config.use_global_stats = use_global_stats
@@ -2377,15 +2371,23 @@ class ExpandLayer(LayerBase):
 
 @config_layer('featmap_expand')
 class FeatMapExpandLayer(LayerBase):
-    def __init__(self, name, inputs, device=None, num_filters=None, bias=False):
+    def __init__(self,
+                 name,
+                 inputs,
+                 num_filters=None,
+                 as_row_vector=True,
+                 bias=False,
+                 **xargs):
         super(FeatMapExpandLayer, self).__init__(
-            name, 'featmap_expand', 0, inputs=inputs, device=device)
+            name, 'featmap_expand', 0, inputs=inputs, **xargs)
         config_assert(
             len(self.inputs) == 1, 'ExpandLayer takes 1 and only 1 inputs')
         if num_filters is not None:
             self.config.num_filters = num_filters
         else:
             logger.fatal("FeatMapExpandLayer must specify num_filters.")
+        if not as_row_vector:
+            self.config.user_arg = "as_col_vec"
         self.set_layer_size(self.get_input_layer(0).size * num_filters)
 
 
@@ -2395,14 +2397,12 @@ class MaxLayer(LayerBase):
                  name,
                  inputs,
                  trans_type='non-seq',
-                 active_type='linear',
                  bias=False,
                  output_max_index=None,
                  **xargs):
         super(MaxLayer, self).__init__(name, 'max', 0, inputs=inputs, **xargs)
         config_assert(len(self.inputs) == 1, 'MaxLayer must have 1 input')
         self.config.trans_type = trans_type
-        self.config.active_type = active_type
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
             self.set_layer_size(input_layer.size)
@@ -2444,18 +2444,12 @@ class SequenceLastInstanceLayer(LayerBase):
     def __init__(self,
                  name,
                  inputs,
-                 active_type='linear',
                  trans_type='non-seq',
                  bias=False,
                  stride=-1,
                  **xargs):
         super(SequenceLastInstanceLayer, self).__init__(
-            name,
-            'seqlastins',
-            0,
-            inputs=inputs,
-            active_type=active_type,
-            **xargs)
+            name, 'seqlastins', 0, inputs=inputs, **xargs)
         config_assert(
             len(inputs) == 1, 'SequenceLastInstanceLayer must have 1 input')
         if trans_type == 'seq':
@@ -2471,7 +2465,6 @@ class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
     def __init__(self,
                  name,
                  inputs,
-                 active_type='linear',
                  trans_type='non-seq',
                  bias=False,
                  stride=-1,
@@ -2479,7 +2472,6 @@ class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
         super(SequenceFirstInstanceLayer, self).__init__(
             name,
             inputs=inputs,
-            active_type=active_type,
             trans_type=trans_type,
             bias=bias,
             stride=stride,
@@ -2489,14 +2481,9 @@ class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
 
 @config_layer('seqconcat')
 class SequenceConcatLayer(LayerBase):
-    def __init__(self, name, inputs, active_type='linear', bias=False, **xargs):
+    def __init__(self, name, inputs, bias=False, **xargs):
         super(SequenceConcatLayer, self).__init__(
-            name,
-            'seqconcat',
-            0,
-            inputs=inputs,
-            active_type=active_type,
-            **xargs)
+            name, 'seqconcat', 0, inputs=inputs, **xargs)
         config_assert(
             len(inputs) == 2, 'SequenceConcatLayer must have 2 inputs')
         for input_index in xrange(len(self.inputs)):
@@ -2507,20 +2494,9 @@ class SequenceConcatLayer(LayerBase):
 
 @config_layer('seqreshape')
 class SequenceReshapeLayer(LayerBase):
-    def __init__(self,
-                 name,
-                 size,
-                 inputs,
-                 active_type='linear',
-                 bias=False,
-                 **xargs):
+    def __init__(self, name, size, inputs, bias=False, **xargs):
         super(SequenceReshapeLayer, self).__init__(
-            name,
-            'seqreshape',
-            size,
-            inputs=inputs,
-            active_type=active_type,
-            **xargs)
+            name, 'seqreshape', size, inputs=inputs, **xargs)
         config_assert(
             len(inputs) == 1, 'SequenceReshapeLayer must have 1 inputs')
         self.set_layer_size(size)
@@ -2529,9 +2505,9 @@ class SequenceReshapeLayer(LayerBase):
 
 @config_layer('subseq')
 class SubSequenceLayer(LayerBase):
-    def __init__(self, name, inputs, active_type='linear', bias=False, **xargs):
+    def __init__(self, name, inputs, bias=False, **xargs):
         super(SubSequenceLayer, self).__init__(
-            name, 'subseq', 0, inputs=inputs, active_type=active_type, **xargs)
+            name, 'subseq', 0, inputs=inputs, **xargs)
         config_assert(len(inputs) == 3, 'SubSequenceLayer must have 3 inputs')
         input_layer0 = self.get_input_layer(0)
         size = input_layer0.size
@@ -2687,11 +2663,10 @@ class AverageLayer(LayerBase):
                  inputs,
                  average_strategy='average',
                  trans_type='non-seq',
-                 active_type='linear',
                  bias=False,
                  **xargs):
         super(AverageLayer, self).__init__(
-            name, 'average', 0, inputs=inputs, active_type=active_type, **xargs)
+            name, 'average', 0, inputs=inputs, **xargs)
         self.config.average_strategy = average_strategy
         self.config.trans_type = trans_type
         config_assert(len(inputs) == 1, 'AverageLayer must have 1 input')

python/paddle/trainer_config_helpers/layers.py

@@ -1553,14 +1553,24 @@ def expand_layer(input,
 
 
 @wrap_name_default()
+@wrap_act_default(act=IdentityActivation())
 @layer_support()
-def repeat_layer(input, num_repeats, name=None, layer_attr=None):
+def repeat_layer(input,
+                 num_repeats,
+                 as_row_vector=True,
+                 act=None,
+                 name=None,
+                 layer_attr=None):
     """
-    A layer for repeating the input for num_repeats times. This is equivalent
-    to apply concat_layer() with num_repeats same input.
+    A layer for repeating the input for num_repeats times.
 
+    If as_row_vector:
     .. math::
-       y  = [x, x, \cdots, x]
+       y  = [x_1,\cdots, x_n, \cdots, x_1, \cdots, x_n]
+    If not as_row_vector:
+    .. math::
+       y  = [x_1,\cdots, x_1, \cdots, x_n, \cdots, x_n]
+
 
     The example usage is:
 
@@ -1573,6 +1583,14 @@ def repeat_layer(input, num_repeats, name=None, layer_attr=None):
     :param num_repeats: Repeat the input so many times
     :type num_repeats: int
     :param name: Layer name.
+    :param as_row_vector: True for treating input as row vector and repeating
+                          in the column direction.  This is equivalent to apply
+                          concat_layer() with num_repeats same input.
+                          False for treating input as column vector and repeating
+                          in the row direction.
+    :type as_row_vector: bool
+    :param act: Activation type.
+    :type act: BaseActivation
     :type name: basestring
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
@@ -1583,13 +1601,16 @@ def repeat_layer(input, num_repeats, name=None, layer_attr=None):
     l = Layer(
         inputs=[input.name],
         name=name,
+        active_type=act.name,
         num_filters=num_repeats,
+        as_row_vector=as_row_vector,
         type=LayerType.FEATURE_MAP_EXPAND_LAYER,
         **ExtraAttr.to_kwargs(layer_attr))
     return LayerOutput(
         name=name,
         size=l.config.size,
         layer_type=LayerType.FEATURE_MAP_EXPAND_LAYER,
+        activation=act,
         parents=[input])
 
 
@@ -2834,11 +2855,13 @@ def seq_concat_layer(a, b, act=None, name=None, layer_attr=None,
     Concat sequence a with sequence b.
 
     Inputs:
-      - a = [a1, a2, ..., an]
+      - a = [a1, a2, ..., am]
       - b = [b1, b2, ..., bn]
-      - Note that the length of a and b should be the same.
 
-    Output: [a1, b1, a2, b2, ..., an, bn]
+    Output: [a1, ..., am, b1, ..., bn]
+
+    Note that the above computation is for one sample. Multiple samples are
+    processed in one batch.
 
     The example usage is:
 

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

 #!/bin/bash
-export configs=(test_fc layer_activations projections test_print_layer
+export configs=(test_repeat_layer test_fc layer_activations projections test_print_layer
 test_sequence_pooling test_lstmemory_layer test_grumemory_layer
 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

python/paddle/trainer_config_helpers/tests/configs/protostr/last_first_seq.protostr

@@ -9,7 +9,7 @@ layers {
   name: "__first_seq_0__"
   type: "seqlastins"
   size: 30
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "data"
   }
@@ -21,7 +21,7 @@ layers {
   name: "__first_seq_1__"
   type: "seqlastins"
   size: 30
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "data"
   }
@@ -33,7 +33,7 @@ layers {
   name: "__last_seq_0__"
   type: "seqlastins"
   size: 30
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "data"
   }
@@ -44,7 +44,7 @@ layers {
   name: "__last_seq_1__"
   type: "seqlastins"
   size: 30
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "data"
   }
@@ -55,7 +55,7 @@ layers {
   name: "__first_seq_2__"
   type: "seqlastins"
   size: 30
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "data"
   }
@@ -67,7 +67,7 @@ layers {
   name: "__last_seq_2__"
   type: "seqlastins"
   size: 30
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "data"
   }

python/paddle/trainer_config_helpers/tests/configs/protostr/shared_gru.protostr

@@ -123,7 +123,7 @@ layers {
   name: "__last_seq_0__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__simple_gru_0__"
   }
@@ -134,7 +134,7 @@ layers {
   name: "__last_seq_1__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__simple_gru_1__"
   }

python/paddle/trainer_config_helpers/tests/configs/protostr/shared_lstm.protostr

@@ -205,7 +205,7 @@ layers {
   name: "__last_seq_0__"
   type: "seqlastins"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__lstm_group_0__"
   }
@@ -216,7 +216,7 @@ layers {
   name: "__last_seq_1__"
   type: "seqlastins"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__lstm_group_1__"
   }

python/paddle/trainer_config_helpers/tests/configs/protostr/simple_rnn_layers.protostr

@@ -138,7 +138,7 @@ layers {
   name: "__last_seq_0__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__recurrent_layer_0__"
   }
@@ -149,7 +149,7 @@ layers {
   name: "__first_seq_0__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__recurrent_layer_1__"
   }
@@ -161,7 +161,7 @@ layers {
   name: "__last_seq_1__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__lstmemory_0__"
   }
@@ -172,7 +172,7 @@ layers {
   name: "__first_seq_1__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__lstmemory_1__"
   }
@@ -184,7 +184,7 @@ layers {
   name: "__last_seq_2__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__gru_0__"
   }
@@ -195,7 +195,7 @@ layers {
   name: "__first_seq_2__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__gru_1__"
   }

python/paddle/trainer_config_helpers/tests/configs/protostr/test_repeat_layer.protostr

+type: "nn"
+layers {
+  name: "data"
+  type: "data"
+  size: 30
+  active_type: ""
+}
+layers {
+  name: "__repeat_layer_0__"
+  type: "featmap_expand"
+  size: 300
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+  }
+  num_filters: 10
+}
+layers {
+  name: "__repeat_layer_1__"
+  type: "featmap_expand"
+  size: 300
+  active_type: "tanh"
+  inputs {
+    input_layer_name: "data"
+  }
+  num_filters: 10
+  user_arg: "as_col_vec"
+}
+input_layer_names: "data"
+output_layer_names: "__repeat_layer_0__"
+output_layer_names: "__repeat_layer_1__"
+sub_models {
+  name: "root"
+  layer_names: "data"
+  layer_names: "__repeat_layer_0__"
+  layer_names: "__repeat_layer_1__"
+  input_layer_names: "data"
+  output_layer_names: "__repeat_layer_0__"
+  output_layer_names: "__repeat_layer_1__"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/protostr/test_rnn_group.protostr

@@ -91,7 +91,7 @@ layers {
   name: "__last_seq_0__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "rnn_forward"
   }
@@ -140,7 +140,7 @@ layers {
   name: "__first_seq_0__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "rnn_back"
   }
@@ -190,7 +190,7 @@ layers {
   name: "__last_seq_1__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "rnn_subseq_forward"
   }
@@ -280,7 +280,7 @@ layers {
   name: "__last_seq_2__"
   type: "seqlastins"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__lstm_group_0__"
   }
@@ -329,7 +329,7 @@ layers {
   name: "__last_seq_3__"
   type: "seqlastins"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__gru_group_0__"
   }
@@ -378,7 +378,7 @@ layers {
   name: "__last_seq_4__"
   type: "seqlastins"
   size: 200
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "__fc_layer_0__"
   }

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

@@ -27,7 +27,7 @@ layers {
   name: "__seqreshape_0__"
   type: "seqreshape"
   size: 5
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "data1"
   }

python/paddle/trainer_config_helpers/tests/configs/protostr/test_sequence_pooling.protostr

@@ -9,7 +9,7 @@ layers {
   name: "__seq_pooling_0__"
   type: "max"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "dat_in"
   }
@@ -19,7 +19,7 @@ layers {
   name: "__seq_pooling_1__"
   type: "max"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "dat_in"
   }
@@ -29,7 +29,7 @@ layers {
   name: "__seq_pooling_2__"
   type: "average"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "dat_in"
   }
@@ -40,7 +40,7 @@ layers {
   name: "__seq_pooling_3__"
   type: "average"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "dat_in"
   }
@@ -51,7 +51,7 @@ layers {
   name: "__seq_pooling_4__"
   type: "average"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "dat_in"
   }
@@ -62,7 +62,7 @@ layers {
   name: "__seq_pooling_5__"
   type: "average"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "dat_in"
   }
@@ -73,7 +73,7 @@ layers {
   name: "__seq_pooling_6__"
   type: "max"
   size: 100
-  active_type: "linear"
+  active_type: ""
   inputs {
     input_layer_name: "dat_in"
   }

python/paddle/trainer_config_helpers/tests/configs/test_repeat_layer.py

+from paddle.trainer_config_helpers import *
+
+settings(batch_size=1000, learning_rate=1e-5)
+
+din = data_layer(name='data', size=30)
+
+outputs(
+    repeat_layer(
+        input=din, num_repeats=10, as_row_vector=True),
+    repeat_layer(
+        input=din, num_repeats=10, act=TanhActivation(), as_row_vector=False))