Merge pull request #2701 from luotao1/stride

stride pooling for max and average layer

paddle/gserver/layers/AverageLayer.h

@@ -25,6 +25,10 @@ namespace paddle {
  * If SequenceLevel = kNonSeq:
  *    Output: output size is the number of input sequences (NOT input instances)
  *    output[i] = average_{for each instance in this sequence}{input[i]}
+ *    If stride_ > 0:
+ *      Output: a shorten sequence. Stride is the step size by which we slide a
+ *              window upon the input sequence, and the average pooling
+ *              operation is then applied to each interval independently.
  * If SequenceLevel = kSeq:
  *    Check input sequence must has sub-sequence
  *    Output: output size is the number of input sub-sequences

paddle/gserver/layers/MaxLayer.h

@@ -26,6 +26,10 @@ namespace paddle {
  * If SequenceLevel = kNonSeq:
  *    Output: output size is the number of input sequences (NOT input instances)
  *    output[i] = max_{for each instance in this sequence}{input[i]}
+ *    If stride_ > 0:
+ *      Output: a shorten sequence. Stride is the step size by which we slide a
+ *              window upon the input sequence, and the max pooling operation is
+ *              then applied to each interval independently.
  * If SequenceLevel = kSeq:
  *    Check input sequence must has sub-sequence
  *    Output: output size is the number of input sub-sequences

paddle/gserver/layers/SequenceLastInstanceLayer.cpp

@@ -26,10 +26,9 @@ namespace paddle {
  * If SequenceLevel = kNonseq:
  *   Output: a sequence containing only the last instance of the input sequence
  *   If stride_ > 0:
- *      Output: a shorten sequence. The operation of getting last instance of a
- *              sequence is independently performed on every slice of the input
- *              sequence, which is obtained by sliding a window with the window
- *              size set to stride_.
+ *      Output: a shorten sequence. Stride is the step size by which we slide a
+ *              window upon the input sequence, and getting last instance
+ *              operation is then applied to each interval independently.
  * If SequenceLevel = kSeq:
  *   Check input sequence must has sub-sequence
  *   Output: a sequence containing only the last instance of each sub-sequence
@@ -73,8 +72,7 @@ bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
 void SequenceLastInstanceLayer::forward(PassType passType) {
   SequencePoolLayer::forward(passType);
 
-  auto starts = (stride_ > 0) ? stridePositions_->getData()
-                              : startPositions_->getData(false);
+  auto starts = startPositions_->getData(false);
   MatrixPtr inputValue = getInputValue(0);
   MatrixPtr outputValue = getOutputValue();
 

paddle/gserver/layers/SequencePoolLayer.cpp

@@ -72,9 +72,8 @@ void SequencePoolLayer::forward(PassType passType) {
   if (stride_ > 0) {
     CHECK_EQ(input.hasSubseq(), 0UL)
         << "sequence stride pooling is invalid for hasSubseq now";
-    output_.poolSequenceWithStride(
-        input, stride_, &stridePositions_, reversed_);
-    newBatchSize_ = stridePositions_->getSize() - 1;
+    output_.poolSequenceWithStride(input, stride_, &startPositions_, reversed_);
+    newBatchSize_ = startPositions_->getSize() - 1;
   }
 
   resetOutput(newBatchSize_, dim);

paddle/gserver/layers/SequencePoolLayer.h

@@ -28,8 +28,9 @@ namespace paddle {
  * sequence}{input[i]}
  *    If stride_ > 0:
  *        Check input sequence must not have sub-sequence
- *        Output: a shorten sequence, pooling is performed upon a small local
- *                area
+ *        Output: a shorten sequence. Stride is the step size by which we slide
+ *                a window upon the input sequence, and the pooling operation
+ *                is then applied to each interval independently.
  * If SequenceLevel = kSeq:
  *    Check input sequence must has sub-sequence
  *    Output: output size is the number of input sub-sequences
@@ -47,8 +48,6 @@ protected:
   size_t newBatchSize_;
   ICpuGpuVectorPtr startPositions_;
   int stride_;
-  // Store the start position of each window.
-  IVectorPtr stridePositions_;
   // Whether the input sequence is reversed or not.
   bool reversed_ = false;
 

paddle/gserver/tests/test_LayerGrad.cpp

@@ -845,8 +845,12 @@ void testDegradeLayer(bool hasSubseq,
 
 TEST(Layer, MaxLayer) {
   testDegradeLayer(false, "max", "non-seq", -1);  // seq max to non-seq
-  testDegradeLayer(true, "max", "non-seq", -1);   // hasSubseq max to non-seq
-  testDegradeLayer(true, "max", "seq", -1);       // hasSubseq max to seq
+  testDegradeLayer(false,
+                   "max",
+                   "non-seq",
+                   5);  // seq max to a shorten seq, stride window = 5
+  testDegradeLayer(true, "max", "non-seq", -1);  // hasSubseq max to non-seq
+  testDegradeLayer(true, "max", "seq", -1);      // hasSubseq max to seq
 }
 
 TEST(Layer, SequenceLastInstanceLayer) {
@@ -868,6 +872,10 @@ TEST(Layer, SequenceLastInstanceLayer) {
 
 TEST(Layer, AverageLayer) {
   testDegradeLayer(false, "average", "non-seq", -1);  // seq average to non-seq
+  testDegradeLayer(false,
+                   "average",
+                   "non-seq",
+                   5);  // seq average to a shorten seq, stride window = 5
   testDegradeLayer(
       true, "average", "non-seq", -1);           // hasSubseq average to non-seq
   testDegradeLayer(true, "average", "seq", -1);  // hasSubseq average to seq

paddle/parameter/Argument.cpp

@@ -561,7 +561,7 @@ void Argument::degradeSequence(const Argument& input) {
 
 void Argument::poolSequenceWithStride(const Argument& input,
                                       size_t stride,
-                                      IVectorPtr* stridePostions,
+                                      ICpuGpuVectorPtr* stridePostions,
                                       bool reversed) {
   // If input.sequenceStartPositions = [0, 9, 14, 17, 30] and stride = 5,
   // then sequenceStartPositions = [0, 2, 3, 4, 7].
@@ -598,8 +598,8 @@ void Argument::poolSequenceWithStride(const Argument& input,
   stridePos.emplace_back(starts[numSequences]);
   int size = stridePos.size();
   CHECK_EQ(size - 1, tgtBuf[numSequences]);
-  IVector::resizeOrCreate(*stridePostions, size, false);
-  (*stridePostions)->copyFrom(stridePos.data(), size);
+  ICpuGpuVector::resizeOrCreate(*stridePostions, size, false);
+  (*stridePostions)->getMutableVector(false)->copyFrom(stridePos.data(), size);
 }
 
 void Argument::getValueString(

paddle/parameter/Argument.h

@@ -299,7 +299,7 @@ struct Argument {
    */
   void poolSequenceWithStride(const Argument& input,
                               size_t stride,
-                              IVectorPtr* stridePositions,
+                              ICpuGpuVectorPtr* stridePositions,
                               bool reversed = false);
   /**
    * @brief getValueString will return the argument's output in string. There

paddle/parameter/tests/test_argument.cpp

@@ -31,7 +31,7 @@ TEST(Argument, poolSequenceWithStride) {
   int strideResultReversed[] = {0, 4, 9, 14, 17, 20, 25, 30};
 
   for (auto reversed : {false, true}) {
-    IVectorPtr stridePositions;
+    ICpuGpuVectorPtr stridePositions;
     output.poolSequenceWithStride(
         input, 5 /* stride */, &stridePositions, reversed);
 
@@ -45,7 +45,7 @@ TEST(Argument, poolSequenceWithStride) {
     CHECK_EQ(stridePositions->getSize(), 8UL);
     auto result = reversed ? strideResultReversed : strideResult;
     for (int i = 0; i < 8; i++) {
-      CHECK_EQ(stridePositions->getData()[i], result[i]);
+      CHECK_EQ(stridePositions->getData(false)[i], result[i]);
     }
   }
 }

python/paddle/trainer/config_parser.py

@@ -2466,10 +2466,14 @@ class MaxLayer(LayerBase):
                  trans_type='non-seq',
                  bias=False,
                  output_max_index=None,
+                 stride=-1,
                  **xargs):
         super(MaxLayer, self).__init__(name, 'max', 0, inputs=inputs, **xargs)
         config_assert(len(self.inputs) == 1, 'MaxLayer must have 1 input')
+        if trans_type == 'seq':
+            config_assert(stride == -1, 'subseq does not support stride window')
         self.config.trans_type = trans_type
+        self.config.seq_pool_stride = stride
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
             self.set_layer_size(input_layer.size)
@@ -2731,11 +2735,15 @@ class AverageLayer(LayerBase):
                  average_strategy='average',
                  trans_type='non-seq',
                  bias=False,
+                 stride=-1,
                  **xargs):
         super(AverageLayer, self).__init__(
             name, 'average', 0, inputs=inputs, **xargs)
         self.config.average_strategy = average_strategy
+        if trans_type == 'seq':
+            config_assert(stride == -1, 'subseq does not support stride window')
         self.config.trans_type = trans_type
+        self.config.seq_pool_stride = stride
         config_assert(len(inputs) == 1, 'AverageLayer must have 1 input')
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)

python/paddle/trainer_config_helpers/layers.py

@@ -1246,10 +1246,19 @@ def pooling_layer(input,
                   name=None,
                   bias_attr=None,
                   agg_level=AggregateLevel.TO_NO_SEQUENCE,
+                  stride=-1,
                   layer_attr=None):
     """
     Pooling layer for sequence inputs, not used for Image.
 
+    If stride > 0, this layer slides a window whose size is determined by stride,
+    and return the pooling value of the window as the output. Thus, a long sequence
+    will be shorten. 
+    
+    The parameter stride specifies the intervals at which to apply the pooling 
+    operation. Note that for sequence with sub-sequence, the default value
+    of stride is -1.
+
     The example usage is:
 
     .. code-block:: python
@@ -1268,6 +1277,8 @@ def pooling_layer(input,
     :param pooling_type: Type of pooling, MaxPooling(default), AvgPooling,
                          SumPooling, SquareRootNPooling.
     :type pooling_type: BasePoolingType|None
+    :param stride: The step size between successive pooling regions.
+    :type stride: Int
     :param bias_attr: Bias parameter attribute. False if no bias.
     :type bias_attr: ParameterAttribute|None|False
     :param layer_attr: The Extra Attributes for layer, such as dropout.
@@ -1285,12 +1296,16 @@ def pooling_layer(input,
         extra_dict['output_max_index'] = pooling_type.output_max_index
     extra_dict.update(ExtraLayerAttribute.to_kwargs(layer_attr))
 
+    if agg_level == AggregateLevel.TO_SEQUENCE:
+        assert stride == -1
+
     Layer(
         name=name,
         type=pooling_type.name,
         inputs=[Input(input.name)],
         bias=ParamAttr.to_bias(bias_attr),
         trans_type=agg_level,
+        stride=stride,
         **extra_dict)
 
     return LayerOutput(
@@ -1552,7 +1567,7 @@ def last_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
-    :param stride: window size.
+    :param stride: The step size between successive pooling regions.
     :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
@@ -1608,7 +1623,7 @@ def first_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
-    :param stride: window size.
+    :param stride: The step size between successive pooling regions.
     :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.

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

@@ -14,6 +14,7 @@ layers {
     input_layer_name: "dat_in"
   }
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_1__"
@@ -24,6 +25,7 @@ layers {
     input_layer_name: "dat_in"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_2__"
@@ -35,6 +37,7 @@ layers {
   }
   average_strategy: "average"
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_3__"
@@ -46,6 +49,7 @@ layers {
   }
   average_strategy: "average"
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_4__"
@@ -57,6 +61,7 @@ layers {
   }
   average_strategy: "sum"
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_5__"
@@ -68,6 +73,7 @@ layers {
   }
   average_strategy: "sum"
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__seq_pooling_6__"
@@ -77,8 +83,44 @@ layers {
   inputs {
     input_layer_name: "dat_in"
   }
+  trans_type: "non-seq"
+  seq_pool_stride: 5
+}
+layers {
+  name: "__seq_pooling_7__"
+  type: "average"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "dat_in"
+  }
+  average_strategy: "average"
+  trans_type: "non-seq"
+  seq_pool_stride: 5
+}
+layers {
+  name: "__seq_pooling_8__"
+  type: "average"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "dat_in"
+  }
+  average_strategy: "sum"
+  trans_type: "non-seq"
+  seq_pool_stride: 5
+}
+layers {
+  name: "__seq_pooling_9__"
+  type: "max"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "dat_in"
+  }
   output_max_index: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 input_layer_names: "dat_in"
 output_layer_names: "__seq_pooling_0__"
@@ -88,6 +130,9 @@ output_layer_names: "__seq_pooling_3__"
 output_layer_names: "__seq_pooling_4__"
 output_layer_names: "__seq_pooling_5__"
 output_layer_names: "__seq_pooling_6__"
+output_layer_names: "__seq_pooling_7__"
+output_layer_names: "__seq_pooling_8__"
+output_layer_names: "__seq_pooling_9__"
 sub_models {
   name: "root"
   layer_names: "dat_in"
@@ -98,6 +143,9 @@ sub_models {
   layer_names: "__seq_pooling_4__"
   layer_names: "__seq_pooling_5__"
   layer_names: "__seq_pooling_6__"
+  layer_names: "__seq_pooling_7__"
+  layer_names: "__seq_pooling_8__"
+  layer_names: "__seq_pooling_9__"
   input_layer_names: "dat_in"
   output_layer_names: "__seq_pooling_0__"
   output_layer_names: "__seq_pooling_1__"
@@ -106,6 +154,9 @@ sub_models {
   output_layer_names: "__seq_pooling_4__"
   output_layer_names: "__seq_pooling_5__"
   output_layer_names: "__seq_pooling_6__"
+  output_layer_names: "__seq_pooling_7__"
+  output_layer_names: "__seq_pooling_8__"
+  output_layer_names: "__seq_pooling_9__"
   is_recurrent_layer_group: false
 }
 

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

@@ -14,6 +14,14 @@ for pt in POOL_TYPE:
     for al in AGG_LEVEL:
         opts.append(pooling_layer(input=din, agg_level=al, pooling_type=pt()))
 
+for pt in POOL_TYPE:
+    opts.append(
+        pooling_layer(
+            input=din,
+            agg_level=AggregateLevel.TO_NO_SEQUENCE,
+            pooling_type=pt(),
+            stride=5))
+
 opts.append(
     pooling_layer(
         input=din, pooling_type=MaxPooling(output_max_index=True)))