Merge pull request #3126 from hedaoyuan/slice

Slice Layer

doc/api/v2/config/layer.rst

@@ -198,6 +198,10 @@ identity_projection
 ..  autoclass:: paddle.v2.layer.identity_projection
     :noindex:
 
+slice_projection
+-------------------
+..  autoclass:: paddle.v2.layer.slice_projection
+    :noindex:
 
 table_projection
 ----------------

paddle/gserver/layers/SliceProjection.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "Projection.h"
+
+namespace paddle {
+
+/**
+ * SliceProjection can slice the input value into multiple parts,
+ * and then select some of them to merge into a new output.
+ *
+ * First, calculate the slices that need to be merged into the output.
+ * slices = input.slices().for_output()
+ *
+ * Second, merge each slice into the output.
+ * for(auto slice: slices) {
+ *   out.addAtOffset(slice, offset);
+ * }
+ *
+ * Input slices as output: s0, s1, ...:
+ *   -----------------------
+ *   |///|   |//////|      |
+ *   |/s0|   |//s1//|      |
+ *   |///|   |//////|      |
+ *   -----------------------
+ * Output, merge s0, s1, ... into one output:
+ *   ----------------
+ *   |///|//////|   |
+ *   |/s0|//s1//|...|
+ *   |///|//////|   |
+ *   ----------------
+ *
+ * The config file api is slice_projection.
+ */
+class SliceProjection : public Projection {
+public:
+  SliceProjection(const ProjectionConfig& config,
+                  const ParameterPtr& parameter,
+                  bool useGpu);
+  virtual void forward();
+  virtual void backward(const UpdateCallback& callback);
+
+protected:
+  std::vector<std::pair<size_t, size_t>> slices_;
+};
+
+REGISTER_PROJECTION(slice, SliceProjection);
+
+/**
+ * Constructed function.
+ * @note SliceProjection should not have any parameter.
+ */
+SliceProjection::SliceProjection(const ProjectionConfig& config,
+                                 const ParameterPtr& parameter,
+                                 bool useGpu)
+    : Projection(config, parameter, useGpu) {
+  CHECK(!parameter) << "'slice' projection should not have any parameter";
+
+  slices_.reserve(config.slices_size());
+  for (const auto& slice : config.slices()) {
+    slices_.push_back(std::make_pair(slice.start(), slice.end()));
+  }
+}
+
+void SliceProjection::forward() {
+  size_t offset = 0;
+  for (auto& slice : slices_) {
+    auto slice_out = in_->value->subColMatrix(slice.first, slice.second);
+    out_->value->addAtOffset(*slice_out, offset);
+    offset += slice_out->getWidth();
+  }
+}
+
+void SliceProjection::backward(const UpdateCallback& callback) {
+  if (in_->grad) {
+    size_t offset = 0;
+    for (auto& slice : slices_) {
+      auto slice_out = in_->grad->subColMatrix(slice.first, slice.second);
+      slice_out->addAtOffset(*out_->grad, offset);
+      offset += slice_out->getWidth();
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/tests/concat_slice_a.conf

+#edit-mode: -*- python -*-
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from paddle.trainer_config_helpers import *
+
+settings(batch_size=10)
+
+data = data_layer(name ="input", size=8*16*16)
+
+conv1 = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                       num_channels=8,
+                       num_filters=16, stride=1,
+                       bias_attr=False,
+                       act=ReluActivation())
+conv2 = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                       num_channels=8,
+                       num_filters=16, stride=1,
+                       bias_attr=False,
+                       act=ReluActivation())
+
+proj1 = slice_projection(input=conv1, slices=[(0, 4), (4, 12)])
+
+proj2 = slice_projection(input=conv2, slices=[(1, 5), (5, 15)])
+
+concat = concat_layer(input=[proj1, proj2])
+
+outputs(concat)
+

paddle/gserver/tests/concat_slice_b.conf

+#edit-mode: -*- python -*-
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from paddle.trainer_config_helpers import *
+
+settings(batch_size=10)
+
+data = data_layer(name ="input", size=8*16*16)
+
+conv1 = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                       num_channels=8,
+                       num_filters=16, stride=1,
+                       bias_attr=False,
+                       act=ReluActivation())
+conv2 = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                       num_channels=8,
+                       num_filters=16, stride=1,
+                       bias_attr=False,
+                       act=ReluActivation())
+
+proj1 = slice_projection(input=conv1, slices=[(0, 12)])
+
+proj2 = slice_projection(input=conv2, slices=[(1, 15)])
+
+concat = concat_layer(input=[proj1, proj2])
+
+outputs(concat)
+

paddle/gserver/tests/test_LayerGrad.cpp

@@ -152,6 +152,26 @@ TEST(Projection, identity) {
   }
 }
 
+TEST(Projection, slice) {
+  ProjectionConfig conf;
+  conf.set_type("slice");
+  conf.set_input_size(100);
+  SliceConfig& slice1 = *conf.add_slices();
+  slice1.set_start(10);
+  slice1.set_end(20);
+  SliceConfig& slice2 = *conf.add_slices();
+  slice2.set_start(50);
+  slice2.set_end(70);
+  conf.set_output_size(30);
+  for (auto useGpu : {false, true}) {
+    testProjectionGrad(conf,
+                       INPUT_DATA,
+                       /* parameterSize */ 0,
+                       /* batchSize */ 10,
+                       useGpu);
+  }
+}
+
 TEST(Projection, scaling) {
   ProjectionConfig conf;
   conf.set_type("scaling");

paddle/gserver/tests/test_NetworkCompare.cpp

@@ -237,6 +237,12 @@ TEST(Compare, concat_table) {
   compareNetwork(config_file_a, config_file_b);
 }
 
+TEST(Compare, concat_slice) {
+  std::string config_file_a = "./gserver/tests/concat_slice_a.conf";
+  std::string config_file_b = "./gserver/tests/concat_slice_b.conf";
+  compareNetwork(config_file_a, config_file_b);
+}
+
 #ifndef PADDLE_ONLY_CPU
 TEST(Compare, img_pool) {
   std::string config_file_a = "./gserver/tests/img_pool_a.conf";

proto/ModelConfig.proto

@@ -198,6 +198,11 @@ message RowConvConfig {
   required uint32 context_length = 1;
 }
 
+message SliceConfig {
+  required uint32 start = 1;
+  required uint32 end = 2;
+}
+
 message ProjectionConfig {
   required string type = 1;
   required string name = 2;
@@ -218,6 +223,10 @@ message ProjectionConfig {
 
   // For pool
   optional PoolConfig pool_conf = 12;
+
+  // For slice
+  // Each slice output is the input[start, end)
+  repeated SliceConfig slices = 13;
 }
 
 message OperatorConfig {

python/paddle/trainer/config_parser.py

@@ -565,6 +565,35 @@ class IdentityOffsetProjection(Projection):
         return []
 
 
+@config_class
+class SliceProjection(Projection):
+    type = 'slice'
+
+    def __init__(self, input_layer_name, slices, **xargs):
+        super(SliceProjection, self).__init__(input_layer_name, **xargs)
+        input = g_layer_map[input_layer_name]
+        if input.type in ["exconv", "cudnn_conv"]:
+            # the slice operator is for the channel dimension
+            assert input.num_filters is not None
+            channels = input.num_filters
+            image_size = input.size / channels
+            assert slices[len(slices) - 1][1] <= channels
+            for i in xrange(len(slices)):
+                slice = self.proj_conf.slices.add()
+                slice.start = slices[i][0] * image_size
+                slice.end = slices[i][1] * image_size
+                self.size += slice.end - slice.start
+        else:
+            config_assert(False,
+                          'Currently the input should be convolution layer')
+
+    def calc_parameter_size(self, input_size, output_size):
+        return 0
+
+    def calc_parameter_dims(self, input_size, output_size):
+        return []
+
+
 # DotMulProjection performs element-wise multiplication with weight
 @config_class
 class DotMulProjection(Projection):

python/paddle/trainer_config_helpers/layers.py

@@ -128,6 +128,7 @@ __all__ = [
     'prelu_layer',
     'gated_unit_layer',
     'crop_layer',
+    'slice_projection',
 ]
 
 
@@ -536,6 +537,45 @@ def identity_projection(input, offset=None, size=None):
     return proj
 
 
+def slice_projection(input, slices):
+    """
+    slice_projection can slice the input value into multiple parts,
+    and then select some of them to merge into a new output.
+
+    .. math::
+       output = [input.slices()]
+
+    The example usage is:
+
+    .. code-block:: python
+
+       proj = slice_projection(input=layer, slices=[(0, 10), (20, 30)])
+
+    Note that slice_projection should not have any parameter.
+
+    :param input: Input Layer.
+    :type input: LayerOutput
+    :param slices: An array of slice parameters.
+                   Each slice contains the start and end offsets based
+                   on the input.
+    :type slices: pair of int
+    :return: A SliceProjection object
+    :rtype: SliceProjection
+    """
+    assert len(slices) >= 1
+    start = 0
+    for i in xrange(len(slices)):
+        assert len(slices[i]) == 2
+        # The start position of the next slice needs to be greater than
+        # or equal to the end position of the previous slice.
+        assert slices[i][0] >= start
+        assert slices[i][1] >= slices[i][0]
+        start = slices[i][1]
+    proj = SliceProjection(input_layer_name=input.name, slices=slices)
+    proj.origin = input
+    return proj
+
+
 @wrap_param_attr_default()
 def scaling_projection(input, param_attr=None):
     """