diff --git a/paddle/cuda/include/hl_matrix.h b/paddle/cuda/include/hl_matrix.h
index abd5eb3a0cf338c689680dd0f7192be7b2530383..40828dd5cc76f4197e6cfbb1121f2eef2c1ac580 100644
--- a/paddle/cuda/include/hl_matrix.h
+++ b/paddle/cuda/include/hl_matrix.h
@@ -267,4 +267,16 @@ extern void hl_matrix_collect_shared_bias(real* B_d,
                                           const int dimN,
                                           real scale);
 
+/**
+ * @brief  Matrix rotation in 90 degrees
+ *
+ * @param[in]   mat       input matrix (M x N).
+ * @param[out]  matRot    output matrix (N x M).
+ * @param[in]   dimM      input matrix height.
+ * @param[in]   dimN      input matrix width.
+ * @param[in]   clockWise rotation direction
+ */
+extern void hl_matrix_rotate(
+    real* mat, real* matRot, int dimM, int dimN, bool clockWise);
+
 #endif /* HL_MATRIX_H_ */
diff --git a/paddle/cuda/include/stub/hl_matrix_stub.h b/paddle/cuda/include/stub/hl_matrix_stub.h
index 0b669f6735cb9771fd63ed8e3b45602db0db447c..a1712d1e4d2a5dc80526b7d7b5ad7bd4f5d8c1ed 100644
--- a/paddle/cuda/include/stub/hl_matrix_stub.h
+++ b/paddle/cuda/include/stub/hl_matrix_stub.h
@@ -106,4 +106,8 @@ inline void hl_matrix_collect_shared_bias(real* B_d,
                                           const int dimM,
                                           const int dimN,
                                           real scale) {}
+
+inline void hl_matrix_rotate(
+    real* mat, real* matRot, int dimM, int dimN, bool clockWise) {}
+
 #endif  // HL_MATRIX_STUB_H_
diff --git a/paddle/cuda/src/hl_cuda_matrix.cu b/paddle/cuda/src/hl_cuda_matrix.cu
index 2b4c6f7c39cff78c0e76cc1dfd41e1c7ef334f11..cd23bd31057c5c8cd10173bc5fa5fa67f2d0e422 100644
--- a/paddle/cuda/src/hl_cuda_matrix.cu
+++ b/paddle/cuda/src/hl_cuda_matrix.cu
@@ -840,3 +840,28 @@ void hl_matrix_collect_shared_bias(real* B_d,
       (B_d, A_d, channel, dimM, dimN, dim, limit, scale);
   CHECK_SYNC("hl_matrix_collect_shared_bias failed");
 }
+
+__global__ void keMatrixRotate(real* mat, real* matRot,
+                               int dimM, int dimN, bool clockWise) {
+    int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx < dimM * dimN) {
+        int i = idx / dimN;
+        int j = idx % dimN;
+        if (clockWise) {
+            matRot[j * dimM + i] = mat[(dimM - i - 1) * dimN + j];
+        } else {
+            matRot[j * dimM + i] = mat[i * dimN + (dimN - j - 1)];
+        }
+    }
+}
+
+void hl_matrix_rotate(real *mat, real* matRot,
+                      int dimM, int dimN, bool clockWise) {
+    CHECK_NOTNULL(mat);
+    CHECK_NOTNULL(matRot);
+    const int threads = 512;
+    const int blocks = DIVUP(dimM * dimN, threads);
+    keMatrixRotate<<< blocks, threads, 0, STREAM_DEFAULT >>>
+            (mat, matRot, dimM, dimN, clockWise);
+    CHECK_SYNC("hl_matrix_rotate failed");
+}
diff --git a/paddle/gserver/layers/FeatureMapExpandLayer.cpp b/paddle/gserver/layers/FeatureMapExpandLayer.cpp
index d023074c52167554358d0d4df7ec40cfba9da2a6..4b685812aac24782e4fe47fa85c7e91bd3494087 100644
--- a/paddle/gserver/layers/FeatureMapExpandLayer.cpp
+++ b/paddle/gserver/layers/FeatureMapExpandLayer.cpp
@@ -95,6 +95,9 @@ void FeatureMapExpandLayer::forward(PassType passType) {
 
 void FeatureMapExpandLayer::backward(const UpdateCallback& callback) {
   MatrixPtr inGrad = getInputGrad(0);
+  if (NULL == inGrad) {
+    return;
+  }
   MatrixPtr outGrad = getOutputGrad();
   size_t batchSize = getInput(0).getBatchSize();
   int imgSize = inGrad->getWidth();
diff --git a/paddle/gserver/layers/RotateLayer.cpp b/paddle/gserver/layers/RotateLayer.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..7c71088d786ab218bf0f71b577985c023dd1436f
--- /dev/null
+++ b/paddle/gserver/layers/RotateLayer.cpp
@@ -0,0 +1,102 @@
+/* Copyright (c) 2016 Baidu, Inc. 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 "RotateLayer.h"
+
+namespace paddle {
+
+REGISTER_LAYER(rotate, RotateLayer);
+
+bool RotateLayer::init(const LayerMap& layerMap,
+                       const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  CHECK_EQ(inputLayers_.size(), 1UL);
+  height_ = config_.height();
+  width_ = config_.width();
+  CHECK_GT(height_, 0);
+  CHECK_GT(width_, 0);
+  return true;
+}
+
+void RotateLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr input = getInputValue(0);
+  batchSize_ = input->getHeight();
+  size_ = input->getWidth();
+  CHECK_GE(size_, height_ * width_);
+  CHECK_EQ(size_ % (height_ * width_), 0)
+      << "total size_ is not dividable by (height_ * width_), i.e., "
+      << "channel number should be an integer";
+  channels_ = size_ / (height_ * width_);
+
+  resizeOutput(batchSize_, size_);
+
+  MatrixPtr outV = getOutputValue();
+  for (int b = 0; b < batchSize_; b++) {   // for each input feat map
+    for (int c = 0; c < channels_; c++) {  // for each feat channel
+      MatrixPtr inputSample =
+          Matrix::create(input->getData() + b * size_ + c * height_ * width_,
+                         height_,
+                         width_,
+                         false,
+                         useGpu_);
+      MatrixPtr outputSample =
+          Matrix::create(outV->getData() + b * size_ + c * height_ * width_,
+                         width_,
+                         height_,
+                         false,
+                         useGpu_);
+      inputSample->rotate(outputSample, false, true /* clock-wise */);
+    }
+  }
+
+  if (getInputGrad(0)) {
+    zeroGrad();
+  }
+}
+
+void RotateLayer::backward(const UpdateCallback& callback) {
+  (void)callback;
+
+  MatrixPtr outputGrad = getOutputGrad();
+  if (outputGrad == NULL) {
+    return;
+  }
+  // the grad should be rotated in the reverse direction
+  MatrixPtr preGrad = getInputGrad(0);
+
+  for (int b = 0; b < batchSize_; b++) {   // for each input feat map
+    for (int c = 0; c < channels_; c++) {  // for each feat channel
+      MatrixPtr inputSampleGrad =
+          Matrix::create(preGrad->getData() + b * size_ + c * height_ * width_,
+                         height_,
+                         width_,
+                         false,
+                         useGpu_);
+      MatrixPtr outputSampleGrad = Matrix::create(
+          outputGrad->getData() + b * size_ + c * height_ * width_,
+          width_,
+          height_,
+          false,
+          useGpu_);
+      MatrixPtr tmpGrad = nullptr;
+      outputSampleGrad->rotate(tmpGrad, true, false /* anti clock-wise */);
+      inputSampleGrad->add(*tmpGrad);
+    }
+  }
+}
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/RotateLayer.h b/paddle/gserver/layers/RotateLayer.h
new file mode 100644
index 0000000000000000000000000000000000000000..1a64d4d5a51d9c04df07861f02f1bb91eaec088e
--- /dev/null
+++ b/paddle/gserver/layers/RotateLayer.h
@@ -0,0 +1,51 @@
+/* Copyright (c) 2016 Baidu, Inc. 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. */
+
+#pragma once
+
+#include "Layer.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+/**
+ * A layer for rotating a multi-channel feature map (M x N x C) in the spatial
+ * domain
+ * The rotation is 90 degrees in clock-wise for each channel
+ * \f[
+ *   y(j,i,:) = x(M-i-1,j,:)
+ * \f]
+ * where \f$x\f$ is (M x N x C) input, and \f$y\f$ is (N x M x C) output.
+ *
+ * The config file api is rotate_layer
+ *
+*/
+
+class RotateLayer : public Layer {
+public:
+  explicit RotateLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+  void backward(const UpdateCallback& callback = nullptr);
+
+private:
+  int batchSize_;
+  int size_;
+  int height_;
+  int width_;
+  int channels_;
+};
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/TransLayer.h b/paddle/gserver/layers/TransLayer.h
index b43fa1ebfb003226daed724b4ede3006545e8b07..1d5a370b3414bf96076532858ff822080f40a2f5 100644
--- a/paddle/gserver/layers/TransLayer.h
+++ b/paddle/gserver/layers/TransLayer.h
@@ -20,7 +20,7 @@ limitations under the License. */
 
 namespace paddle {
 /**
- * A layer for transposition.
+ * A layer for transposing a minibatch matrix.
  * \f[
      y = x^\mathrm{T}
  * \f]
diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp
index 8c8e876bd64fb97e11bc04c26ec45358f3f808a1..14d9db52470b2828186eca04d303135910489266 100644
--- a/paddle/gserver/tests/test_LayerGrad.cpp
+++ b/paddle/gserver/tests/test_LayerGrad.cpp
@@ -1316,6 +1316,25 @@ TEST(Layer, ResizeLayer) {
   }
 }
 
+TEST(Layer, RotateLayer) {
+  TestConfig config;
+  config.biasSize = 0;
+  config.layerConfig.set_type("rotate");
+  const int CHANNEL = 2;
+  const int HEIGHT = 8;
+  const int WIDTH = 4;
+  const int INPUT_SIZE = HEIGHT * WIDTH * CHANNEL;
+  config.layerConfig.set_size(INPUT_SIZE);
+  config.layerConfig.set_height(HEIGHT);
+  config.layerConfig.set_width(WIDTH);
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", INPUT_SIZE, 0});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "rotate", 100, false, useGpu);
+  }
+}
+
 TEST(Layer, NCELayer) {
   TestConfig config;
   size_t numClasses = 4;
diff --git a/paddle/math/CpuSparseMatrix.cpp b/paddle/math/CpuSparseMatrix.cpp
index 82a482f701481267e564c7ad8179689deb65a75b..bf62229c03bb1d6e2bdf86d8c56a8157938fb832 100644
--- a/paddle/math/CpuSparseMatrix.cpp
+++ b/paddle/math/CpuSparseMatrix.cpp
@@ -372,7 +372,7 @@ MatrixPtr CpuSparseMatrix::subMatrix(size_t startRow, size_t numRows) {
 }
 
 /* mem MUST be alloced outside (memAlloc=false) */
-void CpuSparseMatrix::transpose(MatrixPtr matTrans, bool memAlloc) {
+void CpuSparseMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) {
   CHECK(!memAlloc);
   CpuSparseMatrix* mat = dynamic_cast<CpuSparseMatrix*>(matTrans.get());
   if (format_ == SPARSE_CSR) {
diff --git a/paddle/math/CpuSparseMatrix.h b/paddle/math/CpuSparseMatrix.h
index d3e8871cb5b320ce420d601bde7f18d85398dde7..860cad1047fc343b13efa901186ea218d0855151 100644
--- a/paddle/math/CpuSparseMatrix.h
+++ b/paddle/math/CpuSparseMatrix.h
@@ -201,7 +201,7 @@ public:
   void zeroMem();
 
   /// mem MUST be alloced outside (memAlloc=false)
-  void transpose(MatrixPtr matTrans, bool memAlloc);
+  void transpose(MatrixPtr& matTrans, bool memAlloc);
 
   void mul(const Matrix& A, const Matrix& B, real alpha, real beta);
 
diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp
index 3ae237bc7de895293c15eedc811cf8a2011a7c52..a8b53e2105b053399e62fba5321fd22c1fe4a50d 100644
--- a/paddle/math/Matrix.cpp
+++ b/paddle/math/Matrix.cpp
@@ -274,6 +274,18 @@ real GpuMatrix::getSum() {
   return sum;
 }
 
+real GpuMatrix::getMin() {
+  CHECK(isContiguous());
+  auto vec = GpuVector(height_ * width_, data_);
+  return vec.getMin();
+}
+
+real GpuMatrix::getMax() {
+  CHECK(isContiguous());
+  auto vec = GpuVector(height_ * width_, data_);
+  return vec.getMax();
+}
+
 void GpuMatrix::accumulateColSum(Matrix& src) {
   CHECK_EQ(getWidth(), src.getWidth());
   CHECK_EQ(getHeight(), (size_t)1);
@@ -371,11 +383,13 @@ MatrixPtr GpuMatrix::getTranspose() {
   }
 }
 
-void GpuMatrix::transpose(MatrixPtr matTrans, bool memAlloc) {
+void GpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) {
   if (memAlloc) {
     matTrans = std::make_shared<GpuMatrix>(width_, height_);
   } else {
     CHECK(matTrans != NULL);
+    CHECK_EQ(matTrans->getHeight(), width_);
+    CHECK_EQ(matTrans->getWidth(), height_);
   }
   real* dataTrans = matTrans->getData();
   real* data = getData();
@@ -385,13 +399,27 @@ void GpuMatrix::transpose(MatrixPtr matTrans, bool memAlloc) {
   hl_matrix_transpose(data, dataTrans, height_, width_, lda, ldc);
 }
 
+void GpuMatrix::rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) {
+  if (memAlloc) {
+    matRot = std::make_shared<GpuMatrix>(width_, height_);
+  } else {
+    CHECK(matRot != NULL);
+    CHECK_EQ(matRot->getHeight(), width_);
+    CHECK_EQ(matRot->getWidth(), height_);
+  }
+
+  real* dataRot = matRot->getData();
+  real* data = getData();
+  hl_matrix_rotate(data, dataRot, height_, width_, clockWise);
+}
+
 MatrixPtr GpuMatrix::getInverse() {
   MatrixPtr matInv;
   inverse(matInv, true);
   return matInv;
 }
 
-void GpuMatrix::inverse(MatrixPtr matInv, bool memAlloc) {
+void GpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) {
   CHECK_EQ(height_, width_);
 
   if (memAlloc) {
@@ -1690,11 +1718,13 @@ MatrixPtr CpuMatrix::getTranspose() {
   }
 }
 
-void CpuMatrix::transpose(MatrixPtr matTrans, bool memAlloc) {
+void CpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) {
   if (memAlloc) {
     matTrans = std::make_shared<CpuMatrix>(width_, height_);
   } else {
     CHECK(matTrans != NULL);
+    CHECK_EQ(matTrans->getHeight(), width_);
+    CHECK_EQ(matTrans->getWidth(), height_);
   }
   real* dataTrans = matTrans->getData();
   real* data = getData();
@@ -1708,13 +1738,35 @@ void CpuMatrix::transpose(MatrixPtr matTrans, bool memAlloc) {
   }
 }
 
+void CpuMatrix::rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) {
+  if (memAlloc) {
+    matRot = std::make_shared<CpuMatrix>(width_, height_);
+  } else {
+    CHECK(matRot != NULL);
+    CHECK_EQ(matRot->getHeight(), width_);
+    CHECK_EQ(matRot->getWidth(), height_);
+  }
+  real* dataRot = matRot->getData();
+  real* data = getData();
+
+  for (size_t i = 0; i < height_; i++) {
+    for (size_t j = 0; j < width_; j++) {
+      if (clockWise) {
+        dataRot[j * height_ + i] = data[(height_ - i - 1) * width_ + j];
+      } else {
+        dataRot[j * height_ + i] = data[i * width_ + (width_ - j - 1)];
+      }
+    }
+  }
+}
+
 MatrixPtr CpuMatrix::getInverse() {
   MatrixPtr matInv;
   inverse(matInv, true);
   return matInv;
 }
 
-void CpuMatrix::inverse(MatrixPtr matInv, bool memAlloc) {
+void CpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) {
   CHECK_EQ(height_, width_);
 
   if (memAlloc) {
diff --git a/paddle/math/Matrix.h b/paddle/math/Matrix.h
index 57c0c2fe40a95d75ca580196e6f9ae36ce6edcdc..c92c0a272d5a72868bd61035d77aa4ed0fad7a7c 100644
--- a/paddle/math/Matrix.h
+++ b/paddle/math/Matrix.h
@@ -372,7 +372,27 @@ public:
    * allocate matTrans' memory outside, then set memAlloc as false;
    * else set as true.
    */
-  virtual void transpose(MatrixPtr matTrans, bool memAlloc) {
+  virtual void transpose(MatrixPtr& matTrans, bool memAlloc) {
+    LOG(FATAL) << "Not implemented";
+  }
+
+  /**
+   * @brief  rotate 90 degrees in clock-wise if clockWise=true;
+   *         otherwise rotate in anti clock-wise
+   * clock-wise:
+   * \f[
+   *   y(j,i) = x(M-i-1,j)
+   * \f]
+   * anti clock-wise:
+   * \f[
+   *   y(j,i) = x(i, N-1-j)
+   * \f]
+   * where \f$x\f$ is (M x N) input, and \f$y\f$ is (N x M) output.
+   *
+   * allocate matRot' memory outside, then set memAlloc as false;
+   * else set as true.
+   */
+  virtual void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) {
     LOG(FATAL) << "Not implemented";
   }
 
@@ -387,7 +407,7 @@ public:
    * if allocate matInv's memory outside, then set memAlloc as false;
    * else set as true.
    */
-  virtual void inverse(MatrixPtr matInv, bool memAlloc) {
+  virtual void inverse(MatrixPtr& matInv, bool memAlloc) {
     LOG(FATAL) << "Not implemented";
   }
 
@@ -1169,11 +1189,15 @@ public:
   void accumulateColSum(Matrix& src);
   real getAbsSum();
 
+  real getMin();
+  real getMax();
+
   MatrixPtr getTranspose();
-  void transpose(MatrixPtr matTrans, bool memAlloc);
+  void transpose(MatrixPtr& matTrans, bool memAlloc);
+  void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise);
 
   MatrixPtr getInverse();
-  void inverse(MatrixPtr matInv, bool memAlloc);
+  void inverse(MatrixPtr& matInv, bool memAlloc);
 
   /// add b to each sample of this.
   void addBias(Matrix& b, real scale);
@@ -1485,10 +1509,11 @@ public:
   real getAbsSum();
 
   MatrixPtr getTranspose();
-  void transpose(MatrixPtr matTrans, bool memAlloc);
+  void transpose(MatrixPtr& matTrans, bool memAlloc);
+  void rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise);
 
   MatrixPtr getInverse();
-  void inverse(MatrixPtr matInv, bool memAlloc);
+  void inverse(MatrixPtr& matInv, bool memAlloc);
 
   void copyFrom(const Matrix& src);
 
diff --git a/paddle/math/SparseMatrix.cpp b/paddle/math/SparseMatrix.cpp
index 3bae6d373f240fcc773644386b290ef9874828ae..6370c77386688a334fa0de8b4e2b272882e9e2b0 100644
--- a/paddle/math/SparseMatrix.cpp
+++ b/paddle/math/SparseMatrix.cpp
@@ -497,7 +497,7 @@ void GpuSparseMatrix::setRow(size_t row,
 
 SparseValueType GpuSparseMatrix::getValueType() const { return valueType_; }
 
-void GpuSparseMatrix::transpose(MatrixPtr matTrans, bool memAlloc) {
+void GpuSparseMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) {
   CHECK_EQ(format_, SPARSE_CSC);
   int nnz = sMatrix_->nnz;
   if (memAlloc) {
diff --git a/paddle/math/SparseMatrix.h b/paddle/math/SparseMatrix.h
index 1d3801548e03a6ae679afb15bf7f620172d61c57..f6cd5df338965b55ca17636de097d2401dc057f9 100644
--- a/paddle/math/SparseMatrix.h
+++ b/paddle/math/SparseMatrix.h
@@ -109,7 +109,7 @@ public:
   MatrixPtr getTranspose();
 
   /// B = A'
-  void transpose(MatrixPtr matTrans, bool memAlloc);
+  void transpose(MatrixPtr& matTrans, bool memAlloc);
 
   void copyFrom(const Matrix& src);
   void copyFrom(const Matrix& src, hl_stream_t stream);
diff --git a/paddle/math/tests/test_SparseMatrix.cpp b/paddle/math/tests/test_SparseMatrix.cpp
index 9d3fbaef43d719d07577631d5df3ac4656610cc6..c0572dfdbf738a4dfad04811b3a3e1b65487ff6d 100644
--- a/paddle/math/tests/test_SparseMatrix.cpp
+++ b/paddle/math/tests/test_SparseMatrix.cpp
@@ -248,11 +248,13 @@ TEST(Matrix, SparseMatrixTranspose) {
             /*dense matrix transpose*/
             CpuMatrixPtr matC(new CpuMatrix(height, width));
             matC->copyFrom(*matA);
-            CpuMatrixPtr matD(new CpuMatrix(width, height));
+            MatrixPtr matD(new CpuMatrix(width, height));
             matC->transpose(matD, false);
+
             /*check result*/
             checkSMatrixEqual2Dense(
-                std::dynamic_pointer_cast<CpuSparseMatrix>(matB), matD);
+                std::dynamic_pointer_cast<CpuSparseMatrix>(matB),
+                std::dynamic_pointer_cast<CpuMatrix>(matD));
           }
         }
       }
diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp
index f0c49791d7e2a67220eafca3e1347f30958877a7..e024f2cf1b913f56301ac7b3380f0c382818f413 100644
--- a/paddle/math/tests/test_matrixCompare.cpp
+++ b/paddle/math/tests/test_matrixCompare.cpp
@@ -105,6 +105,21 @@ void testMatrixGetSum(int height, int width) {
   EXPECT_LE(fabs(cpuSum - gpuSum), err);
 }
 
+void testMatrixGetMinMax(int height, int width) {
+  MatrixPtr cpuInput = std::make_shared<CpuMatrix>(height, width);
+  MatrixPtr gpuInput = std::make_shared<GpuMatrix>(height, width);
+  cpuInput->randomizeUniform();
+  gpuInput->copyFrom(*cpuInput);
+
+  real cpuMin = cpuInput->getMin();
+  real gpuMin = gpuInput->getMin();
+  real cpuMax = cpuInput->getMax();
+  real gpuMax = gpuInput->getMax();
+
+  EXPECT_EQ(cpuMin, gpuMin);
+  EXPECT_EQ(cpuMax, gpuMax);
+}
+
 void testMatrixZeroAtOffset(int height, int width) {
   MatrixPtr cpuA = std::make_shared<CpuMatrix>(height, width);
   MatrixPtr gpuA = std::make_shared<GpuMatrix>(height, width);
@@ -161,11 +176,29 @@ void testMatrixTranspose(int height, int width) {
   cpu->randomizeUniform();
   gpu->copyFrom(*cpu);
   cpu->transpose(cpuT, false);
-  gpu->transpose(gpuT, false);
+  gpu->transpose(gpuT, true);
 
   TensorCheckEqual(*cpuT, *gpuT);
 }
 
+void testMatrixRotate(int height, int width) {
+  MatrixPtr cpu = std::make_shared<CpuMatrix>(height, width);
+  MatrixPtr gpu = std::make_shared<GpuMatrix>(height, width);
+  MatrixPtr cpuR = std::make_shared<CpuMatrix>(width, height);
+  MatrixPtr gpuR = std::make_shared<GpuMatrix>(width, height);
+
+  cpu->randomizeUniform();
+  gpu->copyFrom(*cpu);
+
+  cpu->rotate(cpuR, false, true);
+  gpu->rotate(gpuR, true, true);
+  TensorCheckEqual(*cpuR, *gpuR);
+
+  cpu->rotate(cpuR, true, false);
+  gpu->rotate(gpuR, false, false);
+  TensorCheckEqual(*cpuR, *gpuR);
+}
+
 void testMatrixInverse(int height) {
   MatrixPtr cpu = std::make_shared<CpuMatrix>(height, height);
   MatrixPtr gpu = std::make_shared<GpuMatrix>(height, height);
@@ -181,7 +214,7 @@ void testMatrixInverse(int height) {
   cpu->add(*outputCheck);
 
   gpu->copyFrom(*cpu);
-  cpu->inverse(cpuI, false);
+  cpu->inverse(cpuI, true);
   gpu->inverse(gpuI, false);
 
   TensorCheckErr(*cpuI, *gpuI);
@@ -200,6 +233,7 @@ TEST(Matrix, unary) {
       testMatrixZeroAtOffset(height, width);
       testMatrixGetSum(height, width);
       testMatrixTranspose(height, width);
+      testMatrixRotate(height, width);
     }
     // inverse
     testMatrixInverse(height);
diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto
index 0456404832c301d8ceb9338d32da0cea9eae5234..be4634d5103c0f219389823d132b1977963017e1 100644
--- a/proto/ModelConfig.proto
+++ b/proto/ModelConfig.proto
@@ -427,14 +427,14 @@ message LayerConfig {
   // bias size
   optional uint32 bias_size = 48 [default = 0];
 
-  // this parameter can be used as a user-defined parameter when necessary, 
+  // this parameter can be used as a user-defined parameter when necessary,
   // without changing the proto file.
-  // e.g., when a new layer with a user-defined parameter is implemented, 
+  // e.g., when a new layer with a user-defined parameter is implemented,
   // it can be used to pass that parameter, without modifying the proto file.
   // string type is used for flexibility: different types can be converted
-  // to string and reinterpreted in the user's own layer implementation.  
+  // to string and reinterpreted in the user's own layer implementation.
   optional string user_arg = 49;
-  
+
   // to indicate rectangle image data
   optional uint64 height = 50;
   optional uint64 width = 51;
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index 6701eced60d068312a1a866a6312002f9f5207f7..b02af991dc577e070dac813cfd18d35ab6dfc3e0 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -830,7 +830,6 @@ class Pool(Cfg):
             channels,
             size_x,
             size_y=None,
-            img_width=None,
             start=None,
             stride=None,  # 1 by defalut in protobuf
             stride_y=None,
@@ -1927,8 +1926,8 @@ class BatchNormLayer(LayerBase):
         image_conf = self.config.inputs[0].image_conf
         parse_image(self.inputs[0].image, input_layer.name, image_conf)
 
-        # Only pass the width and height of input to batch_norm layer 
-        # when either of it is non-zero. 
+        # Only pass the width and height of input to batch_norm layer
+        # when either of it is non-zero.
         if input_layer.width != 0 or input_layer.height != 0:
             self.set_cnn_layer(name, image_conf.img_size_y, image_conf.img_size,
                                image_conf.channels, False)
@@ -1968,6 +1967,18 @@ class ResizeLayer(LayerBase):
             'ResizeLayer must have one and only one input')
 
 
+@config_layer('rotate')
+class RotateLayer(LayerBase):
+    def __init__(self, name, inputs, height, width, device=None):
+        super(RotateLayer, self).__init__(
+            name, 'rotate', 0, inputs=inputs, device=device)
+        config_assert(
+            len(self.inputs) == 1,
+            'RotateLayer must have one and only one input')
+        self.set_layer_height_width(height, width)
+        self.set_layer_size(self.get_input_layer(0).size)
+
+
 @config_layer('blockexpand')
 class BlockExpandLayer(LayerBase):
     def __init__(self, name, inputs, **xargs):
diff --git a/python/paddle/trainer/recurrent_units.py b/python/paddle/trainer/recurrent_units.py
old mode 100644
new mode 100755
index ff7e8932dcad798815e8a4d8bc586fba335d5e7a..ef92107a1093d2ec2b2a41677e964fdaa60ac829
--- a/python/paddle/trainer/recurrent_units.py
+++ b/python/paddle/trainer/recurrent_units.py
@@ -15,10 +15,10 @@
 # recurrent_units.py
 # Version 2.0
 #
-# Some recurrent units can be used in recurrent layer group, 
+# Some recurrent units can be used in recurrent layer group,
 #   to use these units, import this module in your config_file:
-#     import trainer.recurrent_units 
-# 
+#     import trainer.recurrent_units
+#
 # The modules in this file are DEPRECATED.
 # If you would like to use lstm/gru
 # please use the functions defined in paddle.trainer_config_helpers.
@@ -29,7 +29,7 @@ from paddle.trainer.config_parser import *
 # long short term memory, can be used in recurrent machine
 # *inputs* must be a list of Projections, for example:
 #   inputs = [FullMatrixProjection("input_layer_name")],
-# *para_prefix* defines parameter names, if the *para_prefix* of 
+# *para_prefix* defines parameter names, if the *para_prefix* of
 #   two LstmRecurrentUnit is same, they share same parameters
 # *out_memory* can be defined outside if it's used outside
 def LstmRecurrentUnit(name,
@@ -197,7 +197,7 @@ def LstmRecurrentLayerGroup(name,
 # gated recurrent unit, can be used in recurrent machine
 # *inputs* should be a list of Projections, for example:
 #   inputs = [FullMatrixProjection("input_layer_name")],
-# *para_prefix* defines parameter names, if the *para_prefix* of 
+# *para_prefix* defines parameter names, if the *para_prefix* of
 #   two GatedRecurrentUnit is same, they share same parameters
 # *out_memory* can be defined outside if it's used outside
 
diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
old mode 100644
new mode 100755
index 14035ce3fd8aabe6a6800fa1be95193f053c1bba..66fa58ac91e33bfeac37d1bfbdad8dab4789c4bd
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -70,6 +70,7 @@ __all__ = [
     'interpolation_layer',
     'bilinear_interp_layer',
     'trans_layer',
+    'rotate_layer',
     'sum_to_one_norm_layer',
     'get_output_layer',
     'LayerType',
@@ -154,6 +155,7 @@ class LayerType(object):
     POWER_LAYER = 'power'
     SCALING_LAYER = 'scaling'
     TRANS_LAYER = 'trans'
+    ROTATE_LAYER = 'rotate'
     OUT_PROD_LAYER = 'out_prod'
     FEATURE_MAP_EXPAND_LAYER = 'featmap_expand'
 
@@ -1642,7 +1644,7 @@ def scaling_layer(input, weight, name=None, layer_attr=None):
 @layer_support()
 def trans_layer(input, name=None, layer_attr=None):
     """
-    A layer for transposition.
+    A layer for transposing a minibatch matrix.
 
     .. math::
        y = x^\mathrm{T}
@@ -1673,6 +1675,52 @@ def trans_layer(input, name=None, layer_attr=None):
         name, LayerType.TRANS_LAYER, parents=[input], size=input.size)
 
 
+@wrap_name_default()
+@layer_support()
+def rotate_layer(input, height, width, name=None, layer_attr=None):
+    """
+    A layer for rotating 90 degrees (clock-wise) for each feature channel,
+    usually used when the input sample is some image or feature map.
+
+    .. math::
+       y(j,i,:) = x(M-i-1,j,:)
+
+    where :math:`x` is (M x N x C) input, and :math:`y` is (N x M x C) output.
+
+    The example usage is:
+
+    .. code-block:: python
+
+       rot = rotate_layer(input=layer,
+                          height=100,
+                          width=100)
+
+    :param input: Input layer.
+    :type input: LayerOutput
+    :param height: The height of the sample matrix
+    :type height: int
+    :param name: Layer name.
+    :type name: basestring
+    :param layer_attr: extra layer attributes.
+    :type layer_attr: ExtraLayerAttribute.
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    assert isinstance(input, LayerOutput)
+    l = Layer(
+        name=name,
+        height=height,
+        width=width,
+        type=LayerType.ROTATE_LAYER,
+        inputs=[input.name],
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(
+        name=name,
+        layer_type=LayerType.ROTATE_LAYER,
+        parents=[input],
+        size=l.config.size)
+
+
 @wrap_name_default()
 @layer_support()
 def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None):
@@ -1826,14 +1874,14 @@ def img_conv_layer(input,
                    trans=False,
                    layer_type=None):
     """
-    Convolution layer for image. Paddle can support both square and non-square 
+    Convolution layer for image. Paddle can support both square and non-square
     input currently.
 
     The details of convolution layer, please refer UFLDL's `convolution
     <http://ufldl.stanford.edu/tutorial/supervised/
     FeatureExtractionUsingConvolution/>`_ .
 
-    Convolution Transpose (deconv) layer for image. Paddle can support both square 
+    Convolution Transpose (deconv) layer for image. Paddle can support both square
     and non-square input currently.
 
     The details of convolution transpose layer,
@@ -1892,7 +1940,7 @@ def img_conv_layer(input,
     :param trans: true if it is a convTransLayer, false if it is a convLayer
     :type trans: bool
     :param layer_type: specify the layer_type, default is None. If trans=True,
-                       layer_type has to be "exconvt", otherwise layer_type 
+                       layer_type has to be "exconvt", otherwise layer_type
                        has to be either "exconv" or "cudnn_conv"
     :type layer_type: String
     :return: LayerOutput object.
@@ -3626,9 +3674,9 @@ def pad_layer(input,
     input data and 3 zeros after the input data in channel dimension.
     pad_h means padding zeros in height dimension. pad_w means padding zeros
     in width dimension.
-    
+
     For example,
-    
+
     .. code-block::
 
       input(2,2,2,3)  = [
@@ -3637,7 +3685,7 @@ def pad_layer(input,
                           [ [[4,3,1], [1,8,7]],
                             [[3,8,9], [2,3,5]] ]
                         ]
- 
+
       pad_c=[1,1], pad_h=[0,0], pad_w=[0,0]
       output(2,4,2,3) = [
                           [ [[0,0,0], [0,0,0]],
@@ -4746,6 +4794,7 @@ def cross_entropy_with_selfnorm(input,
                                 layer_attr=None):
     """
     A loss layer for multi class entropy with selfnorm.
+    Input should be a vector of positive numbers, without normalization.
 
     .. code-block:: python
 
diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py
old mode 100644
new mode 100755
index d0cbe3891f60f4b153b5bf659901b63dd178ae19..cadde11ff81658cb309cd1bf7a44bac6374c1e44
--- a/python/paddle/trainer_config_helpers/networks.py
+++ b/python/paddle/trainer_config_helpers/networks.py
@@ -957,22 +957,22 @@ def simple_gru(input,
     use one complete layer to implement rnn (including simple rnn, gru and lstm)
     with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But,
     the multiplication operation :math:`W x_t` is not computed in these layers.
-    See details in their interfaces in layers.py. 
+    See details in their interfaces in layers.py.
     The other implementation is to use an recurrent group which can ensemble a
     series of layers to compute rnn step by step. This way is flexible for
     attenion mechanism or other complex connections.
 
     - gru_step_layer: only compute rnn by one step. It needs an memory as input
       and can be used in recurrent group.
-    - gru_unit: a wrapper of gru_step_layer with memory. 
+    - gru_unit: a wrapper of gru_step_layer with memory.
     - gru_group: a GRU cell implemented by a combination of multiple layers in
       recurrent group.
-      But :math:`W x_t` is not done in group.  
+      But :math:`W x_t` is not done in group.
     - gru_memory: a GRU cell implemented by one layer, which does same calculation
-      with gru_group and is faster than gru_group. 
-    - simple_gru: a complete GRU implementation inlcuding :math:`W x_t` and 
+      with gru_group and is faster than gru_group.
+    - simple_gru: a complete GRU implementation inlcuding :math:`W x_t` and
       gru_group. :math:`W` contains :math:`W_r`, :math:`W_z` and :math:`W`, see
-      formula in grumemory. 
+      formula in grumemory.
 
     The computational speed is that, grumemory is relatively better than
     gru_group, and gru_group is relatively better than simple_gru.
diff --git a/python/paddle/trainer_config_helpers/tests/layers_test_config.py b/python/paddle/trainer_config_helpers/tests/layers_test_config.py
index ae275735aa2b852b3b226a4a0e5b2d4d000ba199..e6cd35ee761d1acd0b5c1943554c7ea1de6a13f5 100644
--- a/python/paddle/trainer_config_helpers/tests/layers_test_config.py
+++ b/python/paddle/trainer_config_helpers/tests/layers_test_config.py
@@ -39,6 +39,7 @@ z1 = mixed_layer(
 assert z1.size > 0
 
 y2 = fc_layer(input=y, size=15)
+z2 = rotate_layer(input=y2, height=5, width=3)
 
 cos1 = cos_sim(a=x1, b=y1)
 cos3 = cos_sim(a=x1, b=y2, size=3)
@@ -46,7 +47,7 @@ cos3 = cos_sim(a=x1, b=y2, size=3)
 linear_comb = linear_comb_layer(weights=x1, vectors=y2, size=3)
 
 out = fc_layer(
-    input=[cos1, cos3, linear_comb, z, z1],
+    input=[cos1, cos3, linear_comb, z, z1, z2],
     size=num_classes,
     act=SoftmaxActivation())