Remove redundancy codes

doc/api/v2/config/layer.rst

@@ -109,6 +109,12 @@ sum_to_one_norm
     :members: sum_to_one_norm
     :noindex:
 
+cross_channel_norm
+---------------
+..  automodule:: paddle.v2.layer
+    :members: cross_channel_norm
+    :noindex:
+
 Recurrent Layers
 ================
 

paddle/gserver/layers/CrossChannelNormLayer.cpp

@@ -40,7 +40,6 @@ void CrossChannelNormLayer::forward(PassType passType) {
   normBuffer_->addScalar(*normBuffer_, 1e-6);
   inV->square2(*dataBuffer_);
   for (size_t i = 0; i < batchSize; i++) {
-    spatialBuffer_->zeroMem();
     MatrixPtr inTmp = Matrix::create(
         inV->getData() + i * dataDim, channels_, spatialDim, false, useGpu_);
     MatrixPtr dataTmp = Matrix::create(dataBuffer_->getData() + i * dataDim,
@@ -80,7 +79,6 @@ void CrossChannelNormLayer::backward(const UpdateCallback& callback) {
   scaleDiff_->zeroMem();
   for (size_t i = 0; i < batchSize; i++) {
     spatialBuffer_->zeroMem();
-    channelBuffer_->zeroMem();
     // propagate to param.
     MatrixPtr dataBufferTmp =
         Matrix::create(dataBuffer_->getData() + i * dataDim,

paddle/gserver/layers/NormLayer.h

@@ -66,11 +66,10 @@ class ResponseNormLayer : public NormLayer {
 };
 
 /**
- * This layer applys normalize across the channels of each sample to a
- * conv layer's output and scale the output by a group of trainable factors
- * which dimensions equal to the channel's number.
- * - Input: One and only one input layer are accepted. The input layer must be
- *        be a data output layer.
+ * This layer applys normalization across the channels of each sample to a
+ * conv layer's output, and scales the output by a group of trainable factors
+ * whose equal to the number of channels.
+ * - Input: One and only one input layer are accepted.
  * - Output: The normalized data of the input data.
  * Reference:
  *    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed,

python/paddle/trainer_config_helpers/layers.py

@@ -1015,6 +1015,7 @@ def cross_channel_norm_layer(input, name=None, param_attr=None):
     This layer applys normalize across the channels of each sample to
     a conv layer's output and scale the output by a group of trainable
     factors which dimensions equal to the channel's number.
+
     :param name: The Layer Name.
     :type name: basestring
     :param input: The input layer.