add ClipLayer

paddle/gserver/layers/ClipLayer.cpp

@@ -0,0 +1,78 @@
+/* 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 "Layer.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * A layer for clipping the input value by the threshold.
+ * \f[
+ *   out[i] = \min\left(\max\left(in[i],p_{1}\right),p_{2}\right)
+ * \f]
+ */
+
+class ClipLayer : public Layer {
+protected:
+  real clipThresholdLow_;
+  real clipThresholdHigh_;
+
+public:
+  explicit ClipLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(clip, ClipLayer);
+
+bool ClipLayer::init(const LayerMap& layerMap,
+                     const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  CHECK_EQ(inputLayers_.size(), 1U);
+  auto layerConf = config_.inputs(0).clip_conf();
+  clipThresholdLow_ = layerConf.clip_threshold_low();
+  clipThresholdHigh_ = layerConf.clip_threshold_high();
+  CHECK_LT(clipThresholdLow_, clipThresholdHigh_);
+  return true;
+}
+
+void ClipLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr inV = getInputValue(0);
+  resetOutput(inV->getHeight(), inV->getWidth());
+  MatrixPtr outV = getOutputValue();
+  outV->copyFrom(*inV);
+  outV->clip(clipThresholdLow_, clipThresholdHigh_);
+}
+
+void ClipLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inV = getInputValue(0);
+  MatrixPtr inG = getInputGrad(0);
+  MatrixPtr outV = getOutputValue();
+  MatrixPtr outG = getOutputGrad();
+  MatrixPtr tmpMtx;
+  Matrix::resizeOrCreate(
+      tmpMtx, outG->getHeight(), outG->getWidth(), false, useGpu_);
+  tmpMtx->clipDerivative(*inV, clipThresholdLow_, clipThresholdHigh_);
+  inG->addDotMul(*outG, *tmpMtx, 1, 1);
+}
+
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1879,6 +1879,21 @@ TEST(Layer, CropLayer) {
   }
 }
 
+TEST(Layer, ClipLayer) {
+  const size_t batchSize = 128;
+  const size_t size = 512;
+  TestConfig config;
+  config.layerConfig.set_type("clip");
+  config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ClipConfig* layerConf = input->mutable_clip_conf();
+  layerConf->set_clip_threshold_low(std::rand() / (real)RAND_MAX);
+  layerConf->set_clip_threshold_high(std::rand() / (real)RAND_MAX);
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "clip", batchSize, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/math/BaseMatrix.cu

@@ -442,6 +442,12 @@ DEFINE_MATRIX_UNARY_PARAMETER_OP(Clip, TWO_PARAMETER,
 template<class T>
 void BaseMatrixT<T>::clip(T p1, T p2) { applyUnary(unary::Clip<T>(p1, p2)); }
 
+DEFINE_MATRIX_BINARY_PARAMETER_OP(ClipDerivative, TWO_PARAMETER, a = b < p1 ? 0 : (b > p2 ? 0 : 1));
+template<class T>
+void BaseMatrixT<T>::clipDerivative(BaseMatrixT& b, T p1, T p2) {
+  applyBinary(binary::ClipDerivative<T>(p1, p2), b);
+}
+
 DEFINE_MATRIX_UNARY_PARAMETER_OP(BiggerThanScalar, ONE_PARAMETER,
                                  a = a > p ? 1.0f : 0.0f);
 template<class T>

paddle/math/BaseMatrix.h

@@ -488,6 +488,13 @@ public:
    */
   void clip(T p1, T p2);
 
+  /**
+   * this = b < low ? 0 : 1
+   *
+   * this = b > high ? 0 : 1
+   */
+  void clipDerivative(BaseMatrixT& b, T p1, T p2);
+
   /**
    * @code
    * a = a > p ? 1.0f : 0.0f

proto/ModelConfig.proto

@@ -289,6 +289,11 @@ message DetectionOutputConfig {
   optional uint32 width = 9 [default = 1];
 }
 
+message ClipConfig {
+  required float clip_threshold_low = 1;
+  required float clip_threshold_high = 2;
+}
+
 message LayerInputConfig {
   required string input_layer_name = 1;
   optional string input_parameter_name = 2;
@@ -309,6 +314,7 @@ message LayerInputConfig {
   optional RowConvConfig row_conv_conf = 15;
   optional MultiBoxLossConfig multibox_loss_conf = 16;
   optional DetectionOutputConfig detection_output_conf = 17;
+  optional ClipConfig clip_conf = 18;
 }
 
 message LayerConfig {

python/paddle/trainer/config_parser.py

@@ -2169,6 +2169,23 @@ class RowConvLayer(LayerBase):
         self.create_input_parameter(0, psize, dims)
 
 
+@config_layer('clip')
+class ClipLayer(LayerBase):
+    def __init__(self, name, inputs, clip_threshold_low, clip_threshold_high):
+        super(ClipLayer, self).__init__(name, 'clip', 0, inputs=inputs)
+        config_assert(
+            len(self.inputs) == 1,
+            'ClipLayer layer must have one and only one input.')
+        config_assert(
+            clip_threshold_low < clip_threshold_high,
+            'clip_threshold_low must be less than clip_threshold_high.')
+        input_layer = self.get_input_layer(0)
+        self.set_layer_size(input_layer.size)
+        self.config.inputs[0].clip_conf.clip_threshold_low = clip_threshold_low
+        self.config.inputs[
+            0].clip_conf.clip_threshold_high = clip_threshold_high
+
+
 # key: cost type
 # value: cost class
 g_cost_map = {}

python/paddle/trainer_config_helpers/layers.py

@@ -31,103 +31,33 @@ except ImportError:
 import copy
 
 __all__ = [
-    'full_matrix_projection',
-    'AggregateLevel',
-    'ExpandLevel',
-    'identity_projection',
-    'dotmul_projection',
-    'dotmul_operator',
-    'repeat_layer',
-    'seq_reshape_layer',
-    'table_projection',
-    'mixed_layer',
-    'data_layer',
-    'embedding_layer',
-    'fc_layer',
-    'grumemory',
-    'pooling_layer',
-    'lstmemory',
-    'last_seq',
-    'first_seq',
-    'cos_sim',
-    'hsigmoid',
-    'conv_projection',
-    'mse_cost',
-    'regression_cost',
-    'classification_cost',
-    'LayerOutput',
-    'img_conv_layer',
-    'img_pool_layer',
-    'batch_norm_layer',
-    'img_cmrnorm_layer',
-    'addto_layer',
-    'concat_layer',
-    'seq_concat_layer',
-    'lstm_step_layer',
-    'recurrent_group',
-    'memory',
-    'StaticInput',
-    'expand_layer',
-    'scaling_layer',
-    'scaling_projection',
-    'power_layer',
-    'interpolation_layer',
-    'bilinear_interp_layer',
-    'trans_layer',
-    'rotate_layer',
-    'sum_to_one_norm_layer',
-    'get_output_layer',
-    'LayerType',
-    'context_projection',
-    'beam_search',
-    'maxid_layer',
-    'GeneratedInput',
-    'SubsequenceInput',
-    'gru_step_layer',
-    'gru_step_naive_layer',
-    'recurrent_layer',
-    'BaseGeneratedInput',
-    'conv_operator',
-    'conv_shift_layer',
-    'tensor_layer',
-    'selective_fc_layer',
-    'sampling_id_layer',
-    'slope_intercept_layer',
-    'trans_full_matrix_projection',
-    'linear_comb_layer',
-    'convex_comb_layer',
-    'ctc_layer',
-    'warp_ctc_layer',
-    'crf_layer',
-    'crf_decoding_layer',
-    'nce_layer',
-    'cross_entropy_with_selfnorm',
-    'cross_entropy',
-    'multi_binary_label_cross_entropy',
-    'sum_cost',
-    'rank_cost',
-    'lambda_cost',
-    'huber_cost',
-    'block_expand_layer',
-    'maxout_layer',
-    'out_prod_layer',
-    'printer_layer',
-    'print_layer',
-    'priorbox_layer',
-    'cross_channel_norm_layer',
-    'multibox_loss_layer',
-    'detection_output_layer',
-    'spp_layer',
-    'pad_layer',
-    'eos_layer',
-    'smooth_l1_cost',
-    'layer_support',
-    'multiplex_layer',
-    'row_conv_layer',
-    'dropout_layer',
-    'prelu_layer',
-    'gated_unit_layer',
-    'crop_layer',
+    'full_matrix_projection', 'AggregateLevel', 'ExpandLevel',
+    'identity_projection', 'dotmul_projection', 'dotmul_operator',
+    'repeat_layer', 'seq_reshape_layer', 'table_projection', 'mixed_layer',
+    'data_layer', 'embedding_layer', 'fc_layer', 'grumemory', 'pooling_layer',
+    'lstmemory', 'last_seq', 'first_seq', 'cos_sim', 'hsigmoid',
+    'conv_projection', 'mse_cost', 'regression_cost', 'classification_cost',
+    'LayerOutput', 'img_conv_layer', 'img_pool_layer', 'batch_norm_layer',
+    'img_cmrnorm_layer', 'addto_layer', 'concat_layer', 'seq_concat_layer',
+    'lstm_step_layer', 'recurrent_group', 'memory', 'StaticInput',
+    'expand_layer', 'scaling_layer', 'scaling_projection', 'power_layer',
+    'interpolation_layer', 'bilinear_interp_layer', 'trans_layer',
+    'rotate_layer', 'sum_to_one_norm_layer', 'get_output_layer', 'LayerType',
+    'context_projection', 'beam_search', 'maxid_layer', 'GeneratedInput',
+    'SubsequenceInput', 'gru_step_layer', 'gru_step_naive_layer',
+    'recurrent_layer', 'BaseGeneratedInput', 'conv_operator',
+    'conv_shift_layer', 'tensor_layer', 'selective_fc_layer',
+    'sampling_id_layer', 'slope_intercept_layer',
+    'trans_full_matrix_projection', 'linear_comb_layer', 'convex_comb_layer',
+    'ctc_layer', 'warp_ctc_layer', 'crf_layer', 'crf_decoding_layer',
+    'nce_layer', 'cross_entropy_with_selfnorm', 'cross_entropy',
+    'multi_binary_label_cross_entropy', 'sum_cost', 'rank_cost', 'lambda_cost',
+    'huber_cost', 'block_expand_layer', 'maxout_layer', 'out_prod_layer',
+    'printer_layer', 'print_layer', 'priorbox_layer',
+    'cross_channel_norm_layer', 'multibox_loss_layer', 'detection_output_layer',
+    'spp_layer', 'pad_layer', 'eos_layer', 'smooth_l1_cost', 'layer_support',
+    'multiplex_layer', 'row_conv_layer', 'dropout_layer', 'prelu_layer',
+    'gated_unit_layer', 'crop_layer', 'clip_layer'
 ]
 
 
@@ -220,6 +150,7 @@ class LayerType(object):
 
     PRELU = 'prelu'
     CROP_LAYER = 'crop'
+    CLIP_LAYER = 'clip'
 
     @staticmethod
     def is_layer_type(type_name):
@@ -6006,3 +5937,36 @@ def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):
         layer_type=LayerType.CROP_LAYER,
         parents=input,
         size=l.config.size)
+
+
+@wrap_name_default("clip")
+def clip_layer(input, clip_threshold_low, clip_threshold_high, name=None):
+    """
+    A layer for clipping the input value by the threshold.
+
+    .. math::
+
+        out[i] = \min\left(\max\left(in[i],p_{1}\right),p_{2}\right)
+
+    .. code-block:: python
+
+        clip = clip_layer(input=input_layer, clip_threshold_low=-10, clip_threshold_high=10)
+
+    :param name: The Layer Name.
+    :type name: basestring
+    :param input: The input layer.
+    :type input: LayerOutput.
+    :param clip_threshold_low: The lower threshold for clipping.
+    :type clip_threshold_low: float
+    :param clip_threshold_high: The upper threshold for clipping.
+    :type clip_threshold_high: float
+    :return: LayerOutput
+    """
+    Layer(
+        name=name,
+        type=LayerType.CLIP_LAYER,
+        inputs=[input.name],
+        clip_threshold_low=clip_threshold_low,
+        clip_threshold_high=clip_threshold_high)
+    return LayerOutput(
+        name, LayerType.CLIP_LAYER, parents=[input], size=input.size)