Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into bug

.travis.yml

@@ -54,7 +54,9 @@ before_install:
     fi
   - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then paddle/scripts/travis/before_install.osx.sh; fi
   - if [[ "$JOB" == "PRE_COMMIT" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
-  - pip install numpy wheel protobuf sphinx recommonmark sphinx_rtd_theme virtualenv pre-commit requests==2.9.2 LinkChecker
+  # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python 
+  # protobuf version.
+  - pip install numpy wheel 'protobuf==3.1' sphinx recommonmark sphinx_rtd_theme virtualenv pre-commit requests==2.9.2 LinkChecker
 script:
   - paddle/scripts/travis/main.sh
 notifications:

doc/design/api.md

+# PaddlePaddle Design Doc
+
+## Ingredients
+
+As the first step of our design, we list important concepts in deep
+learning and try to figure their relationship, as shown below:
+
+```
+Model = {topology, parameters}
+
+Evaluator = {Model*, activations}
+- forward
+- test(cost, ...)
+
+GradientMachine = {Evaluator*, gradients}
+- backward
+
+Optimizer = {GradientMachine*}
+- train(cost, ...)
+- update
+- checkpoint
+```
+
+where the pair of curly braces `{` and `}` indicate *composition*, `*`
+indicates a *reference*, and `-` marks a "class method".
+
+
+### Model
+
+We used to think that parameters are part of the topology (or layers).
+But that is not true because multiple layers could share the same
+parameter matrix.  An example is a network that compares two text
+segments in a semantic space:
+
+```
+          semantic
+text A -> projection ---\
+          layer A        \
+                          cosine
+                          similarity -> output
+                          layer
+          semantic       /
+text B -> projection ---/
+          layer B
+```
+
+In this network, the two semantic projection layers (A and B) share
+the same parameter matrix.
+
+For more information about our API that specifies topology and
+parameter sharing, please refer to [TODO: API].
+
+
+### Evaluator
+
+Supposed that we have a trained ranking model, we should be able to
+use it in our search engine.  The search engine's Web server is a
+concurrent program so to serve many HTTP requests simultaneously.  It
+doesn't make sense for each of these threads to have its own copy of the model because that would duplicate topologies and parameters.
+However, each thread should be able to record layer outputs, i.e.,
+activations, computed from an input, derived from the request.  With
+*Evaluator* that saves activations, we can write the over-simplified
+server program as:
+
+```python
+m = paddle.model.load("trained.model")
+
+http.handle("/",
+            lambda req:
+                e = paddle.evaluator.create(m)
+                e.forward(req)
+                e.activation(layer="output")) # returns activations of layer "output"
+```
+
+### GradientMachine
+
+Similar to the evaluation, the training needs to compute gradients so
+to update model parameters.  Because an [optimizer](#optimizer) might
+run multiple simultaneous threads to update the same model, gradients
+should be separated from the model.  Because gradients are only used
+in training, but not serving, they should be separate from Evaluator.
+Hence the `GradientMachine`.
+
+### Optimizer
+
+None of Model, Evaluator, nor GradientMachine implements the training
+loop, hence Optimizer.  We can define a concurrent optimizer that runs
+multiple simultaneous threads to train a model -- just let each
+thread has its own GradientMachine object.
+
+Most models should be able to be trained using the
+`paddle.optimizer.SGD` by calling its `train` method.  Many
+customizations to the SGD algorithm happens with the update equation,
+e.g., momentum and the Adam SGD algorithm.  We make `train` calls
+`update` to do an update, so that we can derive a `paddle.optimizer.Adam`
+from `paddle.optimizer.SGD` by overrides only the `update` method.
+
+
+## Programming Interface
+
+A fictive example of PaddlePaddle program looks like the following:
+
+```python
+import paddle
+
+def read(args):
+    f = open_file(args["filename"])
+    mb = read_a_minibatch(f)
+    end_pass = eof(f)
+    if end_pass:
+       f = open_file(args["filename"]) # rewind for reading again
+    yield mb, end_pass
+
+input = paddle.layer.data(...)
+intermediate = paddle.layers.fc(input)
+output = paddle.layer.softmax(intermediate)
+
+model = paddle.model.create(output)
+
+paddle.train(model, data_provider=read)
+```
+
+This shows some important part of a program:
+
+1. Define how to read (and augment) data by defining a function, in
+   this example, `read`, that `yields` a minibatch and a boolean flag
+   `eof_of_pass`.
+
+1. Define the topology, `input`, `intermediate`, and `output` in this
+   example.
+
+1. Create parameters from the topology thus forms the model by calling
+   `paddel.model.create`.
+
+1. Train the model by calling `paddle.train`.
+
+
+### Reader
+
+Not all programming frameworks allow users to define I/O functions.
+An example is Google MapReduce, which can only read from text,
+SSTable, and RecordIO files.  Hadoop MapReduce allows users to define
+readers and writers by deriving from base classes `Reader` and
+`Writer`.  The former is less flexible but also less error-prone.  We
+decide to provide the flexibility to users to define their readers.
+
+
+#### A Synthetic Data Reader
+
+Sometimes we want to test a topology and/or a training algorithm using
+synthetic data.  We can do this by defining the reader a synthesizer:
+
+```python
+def read(args):
+    x = sample_from_uniform(0.0, 1.0)
+    y = sample_from_gauss(2 * x, sigma)
+    yield {x, y}, False # no end-of-file so no end-of-pass
+```
+
+#### A Reader for Online Learning
+
+Readers can also read an infinite data stream, e.g., a log stream from
+a search engine and collected by Kafka:
+
+```python
+def read(args):
+    log_stream = kafka.open_channel(args["kafka channel name"])
+    yeild log_stream.read(), False # no end-of-pass in online learning
+```
+
+### Topology
+
+By default, layers don't have names.  But if we want to refer to a
+layer later some time, for example, when we do serving using the model
+and wants activations/outputs of a layer, we should give it a name.
+
+```python
+input = paddle.layer.data(...)
+intermediate = paddle.layer.fc(input, name="inter", ...)
+output = paddle.layer.softmax(intermediate, name="output", ...)
+
+m = paddle.model.create(output)
+e = paddle.evaluator.create(model)
+e.forward(read_an_input()) # compute activations of all layers.
+print e.activations(layer="inter")  # retrieve the activations of layer "inter"
+print e.activations(layer="output") # retrieve the activations of layer "output"
+```
+
+#### Sharing Parameters
+
+In [above section](#model) we shows a network whose two layers share
+the same parameter matrix.  To specify such cases, we give "parameter
+names" to layers.  If some layers have the same paraemter names,
+`paddle.model.create` creates a single parameter matrix for these
+layers:
+
+```python
+text1 = paddle.layer.data(...)
+sematic1 = paddle.layer.fc(text1, ..., parameter_name="sematic_projection")
+text2 = paddle.layer.data(...)
+sematic2 = paddle.layer.fc(text2, ..., parameter_name="sematic_projection")
+out = paddle.layer.cosine(semantic1, semantic2)
+```
+
+We can also share parameter matrices between layers in different
+models.  To do this, we need an additional parameter that refers to a
+model:
+
+```python
+model1_input = paddle.layer.data(...)
+model1_output = paddle.layer.softmax(model1_input, ...,
+                                     parameter_name="a_parameter_matrix")
+model1 = paddle.model.create(model1_output)
+
+# Another model
+model2_semantic = paddle.layer.fc(text2, ...,
+                                  parameter_name="a_parameter_matrix",
+                                  parameter_model=model1)
+```
+
+### Training
+
+The recommended way to training a model is to call `paddle.train`,
+which simply calls `paddle.optimizer.Default`, a global variable of
+type `paddle.optimizer.SGD`.  Equivalently, we can do
+
+```python
+opt = paddle.optimizer.SGD(...)
+opt.train(model, reader=read, ...)
+```
+
+#### Distributed Training
+
+If users want to do distributed training on a cluster, s/he should
+call `paddle.dist_train` and provides access tokens to the cluster as
+a parameter.
+
+For example, if the user has a TLS certificate that allows him to
+access a Kubernetes cluster, s/he should be able to call
+
+```python
+paddle.dist_train(model,
+                  reader=read,
+                  optimizer=paddle.optimizer.SGDOptimizer(...),
+                  k8s_user="yi",
+                  k8s_token="kube_cluster_tls.pem",
+                  k8s_job="hello",
+                  num_parameter_servers=15)
+```
+
+The pseudo code if `paddle.dist_train` is as follows:
+
+```python
+def dist_train():
+    if os.getenv("KUBERNETES_SERVICE_HOST") == None:
+        image_name = k8s_user + '/' + k8s_job
+        docker_build(image_name)
+        docker_push()
+        kube_ctrl_start_job(image_name, k8s_user, k8s_token)
+    else:
+        rank = kube_list_containers_in_job_and_return_current_containers_rank()
+        if rank == 0:
+            master()
+        elif rank < 15:
+            parameter_server()
+        else:
+            optimizer.train(model, reader=read)
+```
+
+Please be aware that if a process is running on the Kubernetes
+cluster, it will have some environment variables pre-defined.
+
+If `dist_train` doesn't see these environment variables, it knowns
+that it's running on users' personal computer, and it should work as a
+*launcher*.  Otherwise, it knows that it's running on the cluster and
+need to figure out its role as either the master, or a trainer, or a
+parameter server.

paddle/cuda/include/hl_matrix.h

@@ -188,48 +188,6 @@ extern void hl_param_relu_backward_diff(real* grad_o,
                                         int width,
                                         int height,
                                         int partial_sum);
-/**
- * @brief cos sim forward
- *
- * @param[out]    output         output data
- * @param[in]     input1         input1 data(matrix)
- * @param[in]     input2         input2 data(matrix or vector)
- * @param[in]     width          matrix width
- * @param[in]     input1_height  input1_height
- * @param[in]     input2_height  input2_height
- * @param[in]     scale          scale factor
- */
-extern void hl_cossim(real* output,
-                      real* input1,
-                      real* input2,
-                      int width,
-                      int input1_height,
-                      int input2_height,
-                      real scale);
-/**
- * @brief cos sim derivate
- *
- * @param[in]     grad             output grad
- * @param[in]     output           output data
- * @param[in]     prevOutX         input1 data
- * @param[in]     prevOutY         input2 data
- * @param[out]    prevGradX        input1 grad
- * @param[out]    prevGradY        input2 grad
- * @param[in]     width            matrix width
- * @param[in]     input1_height    input1 height
- * @param[in]     input2_height    input2 height
- * @param[in]     scale            scale factor
- */
-extern void hl_cossim_derivative(real* grad,
-                                 real* output,
-                                 real* prevOutX,
-                                 real* prevOutY,
-                                 real* prevGradX,
-                                 real* prevGradY,
-                                 int width,
-                                 int input1_height,
-                                 int input2_height,
-                                 real scale);
 
 /**
  * @brief   Matrix addition: A_d[i][j] += scale * B_d[j/channel].

paddle/cuda/include/stub/hl_matrix_stub.h

@@ -74,25 +74,6 @@ inline void hl_param_relu_backward_diff(real* grad_o,
                                         int height,
                                         int partial_sum) {}
 
-inline void hl_cossim(real* output,
-                      real* input1,
-                      real* input2,
-                      int width,
-                      int input1_height,
-                      int input2_height,
-                      real scale) {}
-
-inline void hl_cossim_derivative(real* grad,
-                                 real* output,
-                                 real* prevOutX,
-                                 real* prevOutY,
-                                 real* prevGradX,
-                                 real* prevGradY,
-                                 int width,
-                                 int input1_height,
-                                 int input2_height,
-                                 real scale) {}
-
 inline void hl_matrix_add_shared_bias(real* A_d,
                                       real* B_d,
                                       const int channel,

paddle/cuda/src/hl_cuda_matrix.cu

@@ -584,177 +584,6 @@ void hl_param_relu_backward_diff(real* grad_o,
   CHECK_SYNC("hl_param_relu_backward_diff failed");
 }
 
-template<int blockSize>
-__global__ void KeCosSim(real* output,
-                         real* input1,
-                         real* input2,
-                         int width,
-                         int input1_height,
-                         int input2_height,
-                         real scale) {
-  const int ty = blockIdx.y;
-  int tid = threadIdx.x;
-
-  __shared__ real xx[blockSize];
-  __shared__ real yy[blockSize];
-  __shared__ real xy[blockSize];
-
-  xx[tid] = 0.0;
-  yy[tid] = 0.0;
-  xy[tid] = 0.0;
-  __syncthreads();
-
-  input1 += ty * width;
-  if (input2_height > 1) {
-    input2 += ty * width;
-  }
-  for (int index = tid; index < width; index += blockSize) {
-    real x = input1[index];
-    real y = input2[index];
-    xx[tid] += x * x;
-    yy[tid] += y * y;
-    xy[tid] += x * y;
-  }
-  __syncthreads();
-
-  for (int s = blockSize / 2; s > 0; s >>= 1) {
-    if (tid < s) {
-      xx[tid] += xx[tid + s];
-      yy[tid] += yy[tid + s];
-      xy[tid] += xy[tid + s];
-    }
-    __syncthreads();
-  }
-  if (tid == 0) {
-    output[ty] = scale * xy[0] / (sqrt(xx[0]) * sqrt(yy[0]));
-  }
-}
-
-void hl_cossim(real* output,
-               real* input1,
-               real* input2,
-               int width,
-               int input1_height,
-               int input2_height,
-               real scale) {
-  CHECK_NOTNULL(output);
-  CHECK_NOTNULL(input1);
-  CHECK_NOTNULL(input2);
-  const int blockSize = 256;
-  dim3 threads(blockSize, 1);
-  dim3 grid(1, input1_height);
-
-  KeCosSim<blockSize><<<grid, threads, 0, STREAM_DEFAULT>>>
-    (output, input1, input2, width, input1_height, input2_height, scale);
-  CHECK_SYNC("hl_cossim failed");
-}
-
-template<int blockSize>
-__global__ void KeCosSimDerivative(real* grad,
-                                   real* output,
-                                   real* prevOutX,
-                                   real* prevOutY,
-                                   real* prevGradX,
-                                   real* prevGradY,
-                                   int width,
-                                   int input1_height,
-                                   int input2_height,
-                                   real scale) {
-  const int ty = blockIdx.y;
-  int tid = threadIdx.x;
-
-  __shared__ real xx[blockSize];
-  __shared__ real yy[blockSize];
-  __shared__ real xy[blockSize];
-
-  xx[tid] = 0.0;
-  yy[tid] = 0.0;
-  xy[tid] = 0.0;
-  __syncthreads();
-
-  prevOutX += ty * width;
-  prevGradX += ty * width;
-  if (input2_height > 1) {
-    prevOutY += ty * width;
-    prevGradY += ty * width;
-  }
-  for (int index = tid; index < width; index += blockSize) {
-    real x = prevOutX[index];
-    real y = prevOutY[index];
-    xx[tid] += x * x;
-    yy[tid] += y * y;
-    xy[tid] += x * y;
-  }
-  __syncthreads();
-
-  for (int s = blockSize / 2; s > 0; s >>= 1) {
-    if (tid < s) {
-      xx[tid] += xx[tid + s];
-      yy[tid] += yy[tid + s];
-      xy[tid] += xy[tid + s];
-    }
-    __syncthreads();
-  }
-  if (xy[0] == 0) {
-    real reciprocal = 1.0 / (sqrt(xx[0]) * sqrt(yy[0]));
-    for (int index = tid; index < width; index += blockSize) {
-      prevGradX[index] +=
-        scale * grad[ty] * prevOutY[index] * reciprocal;
-      if (input2_height > 1) {
-        prevGradY[index] +=
-          scale * grad[ty] * prevOutX[index] * reciprocal;
-      } else {
-        paddle::paddleAtomicAdd(prevGradY + index,
-          scale * grad[ty] * prevOutX[index] * reciprocal);
-      }
-    }
-  } else {
-    real reciprocalXY = 1.0 / xy[0];
-    real reciprocalSquareSumX = 1.0 / xx[0];
-    real reciprocalSquareSumY = 1.0 / yy[0];
-    for (int index = tid; index < width; index += blockSize) {
-      prevGradX[index] += output[ty] * grad[ty] *
-        (prevOutY[index] * reciprocalXY -
-         prevOutX[index] * reciprocalSquareSumX);
-      if (input2_height > 1) {
-        prevGradY[index] += output[ty] * grad[ty] *
-          (prevOutX[index] * reciprocalXY -
-           prevOutY[index] * reciprocalSquareSumY);
-      } else {
-        paddle::paddleAtomicAdd(prevGradY + index, output[ty] * grad[ty] *
-          (prevOutX[index] * reciprocalXY -
-           prevOutY[index] * reciprocalSquareSumY));
-      }
-    }
-  }
-}
-
-
-void hl_cossim_derivative(real* grad,
-                          real* output,
-                          real* prevOutX,
-                          real* prevOutY,
-                          real* prevGradX,
-                          real* prevGradY,
-                          int width,
-                          int input1_height,
-                          int input2_height,
-                          real scale) {
-  CHECK_NOTNULL(grad);
-  CHECK_NOTNULL(output);
-  CHECK_NOTNULL(prevOutX);
-  CHECK_NOTNULL(prevOutY);
-  CHECK_NOTNULL(prevGradX);
-  CHECK_NOTNULL(prevGradY);
-  const int blockSize = 256;
-  dim3 threads(blockSize, 1);
-  dim3 grid(1, input1_height);
-  KeCosSimDerivative<blockSize><<<grid, threads, 0, STREAM_DEFAULT>>>
-    (grad, output, prevOutX, prevOutY, prevGradX, prevGradY, width,
-        input1_height, input2_height, scale);
-  CHECK_SYNC("hl_cossim_derivate failed");
-}
-
 __global__ void KeMatrixAddSharedBias(real* A,
                                       real* B,
                                       const int channel,

paddle/function/BufferArg.h

@@ -190,7 +190,7 @@ public:
       : BufferArg(VALUE_TYPE_INT32, shape, argType) {
     bufferType_ = TENSOR_SEQUENCE_ID;
     CHECK_EQ(shape_.ndims(), 1UL);
-    CHECK_GT(shape_[0], 1UL);
+    CHECK_GE(shape_[0], 1UL);
     numSeqs_ = shape_[0] - 1;
   }
 
@@ -226,7 +226,8 @@ public:
   SequenceArg(ValueType valueType,
               const TensorShape& shape,
               ArgType argType = UNSPECIFIED)
-      : BufferArg(valueType, shape, argType), startPositions_(TensorShape()) {
+      : BufferArg(valueType, shape, argType),
+        startPositions_(TensorShape({shape[0]})) {
     bufferType_ = TENSOR_SEQUENCE_DATA;
   }
 

paddle/function/CMakeLists.txt

@@ -27,6 +27,7 @@ if(WITH_TESTING)
     add_simple_unittest(ContextProjectionOpTest)
     add_simple_unittest(PadOpTest)
     add_simple_unittest(MulOpTest)
+    add_simple_unittest(CosSimOpTest)
 endif()
 endif()
 

paddle/function/ContextProjectionOp.cpp

@@ -108,26 +108,23 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK(1 == inputs.size() || 2 == inputs.size());
-    CHECK_EQ((size_t)1, outputs.size());
+    CHECK(1UL == inputs.size() || 2UL == inputs.size());
+    CHECK_EQ(1UL, outputs.size());
     CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
         << "SequenceArg required here";
     const auto val_seqs = dynamic_cast<const SequenceArg&>(inputs[0]);
     auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
 
     CHECK(out_seq.data() && val_seqs.data() && val_seqs.getSequenceId().data());
-    CHECK_EQ(out_seq.shape().ndims(), (size_t)2);
-    CHECK_EQ(val_seqs.shape().ndims(), (size_t)2);
-    CHECK_EQ(val_seqs.getSequenceId().shape().ndims(), (size_t)1);
-    if (2 == inputs.size()) {
-      CHECK_EQ(inputs[1].shape().ndims(), (size_t)2);
-    }
+    CHECK_EQ(out_seq.shape().ndims(), 2UL);
+    CHECK_EQ(val_seqs.shape().ndims(), 2UL);
     /// dim of output = dim of input * context_length
     CHECK_EQ(out_seq.shape()[1], val_seqs.shape()[1] * context_length_);
     /// input and output has the same batch_size
     CHECK_EQ(val_seqs.shape()[0], out_seq.shape()[0]);
-    /// dim of input == dim of weight
-    if (2 == inputs.size()) {
+    if (2UL == inputs.size()) {
+      CHECK_EQ(inputs[1].shape().ndims(), 2UL);
+      /// dim of input == dim of weight
       CHECK_EQ(val_seqs.shape()[1], inputs[1].shape()[1]);
     }
 
@@ -135,10 +132,11 @@ public:
     auto out_mat = out_seq.matrix<Device>();
     const auto in_mat = val_seqs.matrix<Device>();
     const auto w_mat =
-        (2 == inputs.size())
+        (2UL == inputs.size() && inputs[1].data())
             ? inputs[1].matrix<Device>()
             : typename Tensor<real, Device>::Matrix(nullptr, 0, 0);
     const auto seq_vec = val_seqs.getSequenceId().vector<int, Device>();
+
     ContextProjectionForward<Device>(out_mat,
                                      in_mat,
                                      w_mat,
@@ -235,36 +233,40 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ((size_t)1, inputs.size());
-    CHECK_EQ((size_t)2, outputs.size());
+    CHECK_EQ(1UL, inputs.size());
+    CHECK(1UL == outputs.size() || 2UL == outputs.size());
     CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
         << "SequenceArg required here";
     const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
     auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
     CHECK(in_seq.data() && in_seq.getSequenceId().data());
-    CHECK_EQ(in_seq.shape().ndims(), (size_t)2);
-    CHECK_EQ(in_seq.getSequenceId().shape().ndims(), (size_t)1);
-    CHECK_EQ(out_seq.shape().ndims(), (size_t)2);
-    CHECK_EQ(out_seq.getSequenceId().shape().ndims(), (size_t)1);
-    CHECK_EQ(outputs[1].shape().ndims(), (size_t)2);
+    CHECK_EQ(in_seq.shape().ndims(), 2UL);
+    CHECK_EQ(out_seq.shape().ndims(), 2UL);
+    CHECK_EQ(out_seq.getSequenceId().shape().ndims(), 1UL);
 
-    /// dim of input grad == dim of weight
-    CHECK_EQ(out_seq.shape()[1], outputs[1].shape()[1]);
     /// input and output grad has the same batch_size
     CHECK_EQ(out_seq.shape()[0], in_seq.shape()[0]);
     /// dim of output grad = dim of input grad * context_length
     CHECK_EQ(in_seq.shape()[1], out_seq.shape()[1] * context_length_);
     CHECK_EQ(out_seq.getArgType(), ADD_TO);
-    CHECK_EQ(outputs[1].getArgType(), ADD_TO);
+
+    if (2UL == outputs.size()) {
+      CHECK_EQ(outputs[1].shape().ndims(), 2UL);
+      /// dim of input grad == dim of weight
+      CHECK_EQ(out_seq.shape()[1], outputs[1].shape()[1]);
+      CHECK_EQ(outputs[1].getArgType(), ADD_TO);
+    }
 
     const auto seq_vec = in_seq.getSequenceId().vector<int, Device>();
     const auto out_grad_mat = in_seq.matrix<Device>();
     auto in_grad_mat =
         !out_seq.data() ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
                         : out_seq.matrix<Device>();
-    auto w_grad_mat = !outputs[1].data()
-                          ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
-                          : outputs[1].matrix<Device>();
+    auto w_grad_mat =
+        (2UL == outputs.size() && outputs[1].data())
+            ? outputs[1].matrix<Device>()
+            : typename Tensor<real, Device>::Matrix(nullptr, 0, 0);
+
     ContextProjectionBackward<Device>(out_grad_mat,
                                       in_grad_mat,
                                       w_grad_mat,
@@ -304,17 +306,17 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ(1, static_cast<int>(inputs.size()));
-    CHECK_EQ(1, static_cast<int>(outputs.size()));
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
     CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
         << "SequenceArg required here";
     const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
     const auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
 
     CHECK(in_seq.data() && out_seq.data() && in_seq.getSequenceId().data());
-    CHECK_EQ(static_cast<int>(out_seq.shape().ndims()), 2);
-    CHECK_EQ(static_cast<int>(in_seq.shape().ndims()), 2);
-    CHECK_EQ(static_cast<int>(in_seq.getSequenceId().shape().ndims()), 1);
+    CHECK_EQ(out_seq.shape().ndims(), 2UL);
+    CHECK_EQ(in_seq.shape().ndims(), 2UL);
+    CHECK_EQ(in_seq.getSequenceId().shape().ndims(), 1UL);
     /// output layer grad dim == input layer grad dim * context_length_
     CHECK_EQ(in_seq.shape().ndims(), out_seq.shape().ndims() * context_length_);
     /// input and output has the same batch_size
@@ -355,14 +357,14 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ(1, static_cast<int>(inputs.size()));
-    CHECK_EQ(1, static_cast<int>(outputs.size()));
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
     CHECK(inputs[0].isSequenceArg()) << "SequenceArg required here";
     const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
     CHECK(in_seq.data() && in_seq.getSequenceId().data() && outputs[0].data());
-    CHECK_EQ(static_cast<int>(outputs[0].shape().ndims()), 2);
-    CHECK_EQ(static_cast<int>(in_seq.shape().ndims()), 2);
-    CHECK_EQ(static_cast<int>(in_seq.getSequenceId().shape().ndims()), 1);
+    CHECK_EQ(outputs[0].shape().ndims(), 2UL);
+    CHECK_EQ(in_seq.shape().ndims(), 2UL);
+    CHECK_EQ(in_seq.getSequenceId().shape().ndims(), 1UL);
     CHECK_EQ(in_seq.shape()[0], outputs[0].shape()[0]);
     /// output layer grad dim == weight dim * context_length_
     CHECK_EQ(in_seq.shape()[1], outputs[0].shape()[1] * context_length_);

paddle/function/ContextProjectionOp.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-
 #include "Function.h"
 
 namespace paddle {

paddle/function/ContextProjectionOpTest.cpp

@@ -28,55 +28,26 @@ void testMatrixProjectionForward(int context_start,
                std::max(0, (int)(context_start + context_length - 1));
   if (pad == 0) is_padding = false;
 
-  FunctionCompare compare("ContextProjectionForward",
-                          FuncConfig()
-                              .set("context_length", context_length)
-                              .set("context_start", context_start)
-                              .set("begin_pad", std::max(0, -context_start)));
-
-  CpuMatrix cpu_in(batch_size, input_dim);
-  cpu_in.randomizeUniform();
-  GpuMatrix gpu_in(batch_size, input_dim);
-  gpu_in.copyFrom(cpu_in);
-  auto cpu_weight =
-      is_padding ? std::make_shared<CpuMatrix>(pad, input_dim) : nullptr;
-  auto gpu_weight =
-      is_padding ? std::make_shared<GpuMatrix>(pad, input_dim) : nullptr;
-  if (is_padding) {
-    cpu_weight->randomizeUniform();
-    gpu_weight->copyFrom(*cpu_weight);
+  FunctionCompare test("ContextProjectionForward",
+                       FuncConfig()
+                           .set("context_length", context_length)
+                           .set("context_start", context_start)
+                           .set("begin_pad", std::max(0, -context_start)));
+
+  // prepare input arguments
+  test.addSequence(SequenceIdArg(TensorShape{batch_size}));
+  test.addInputs(
+      SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim}));
+  if (is_padding) {  // weight
+    test.addInputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{pad, input_dim}));
   }
-  IVectorPtr cpu_seq;
-  generateSequenceStartPositions(batch_size, cpu_seq);
-  IVectorPtr gpu_seq = IVector::create(cpu_seq->getSize(), true);
-  gpu_seq->copyFrom(*cpu_seq);
-
-  CpuMatrix cpu_out(batch_size, input_dim * context_length);
-  GpuMatrix gpu_out(batch_size, input_dim * context_length);
-  cpu_out.randomizeUniform();
-  gpu_out.copyFrom(cpu_out);
-
-  BufferArgs cpu_inputs;
-  BufferArgs cpu_outputs;
-  cpu_inputs.addArg(cpu_in, *cpu_seq);
-  if (cpu_weight) {
-    cpu_inputs.addArg(*cpu_weight, *cpu_seq);
-  }
-  cpu_outputs.addArg(cpu_out, *cpu_seq, ADD_TO);
-
-  compare.getCpuFunction()->calc(cpu_inputs, cpu_outputs);
+  test.addOutputs(
+      SequenceArg(VALUE_TYPE_FLOAT,
+                  TensorShape{batch_size, input_dim * context_length}),
+      ADD_TO);
 
-  BufferArgs gpu_inputs;
-  BufferArgs gpu_outputs;
-  gpu_inputs.addArg(gpu_in, *gpu_seq);
-  if (gpu_weight) {
-    gpu_inputs.addArg(*gpu_weight, *gpu_seq);
-  }
-  gpu_outputs.addArg(gpu_out, *gpu_seq, ADD_TO);
-
-  compare.getGpuFunction()->calc(gpu_inputs, gpu_outputs);
-
-  autotest::TensorCheckEqual(cpu_out, gpu_out);
+  // run Function
+  test.run();
 }
 
 void testMatrixProjectionBackward(int context_start,
@@ -88,63 +59,31 @@ void testMatrixProjectionBackward(int context_start,
                std::max(0, (int)(context_start + context_length - 1));
   if (pad == 0) is_padding = false;
 
-  FunctionCompare compare("ContextProjectionBackward",
-                          FuncConfig()
-                              .set("context_length", context_length)
-                              .set("context_start", context_start)
-                              .set("begin_pad", std::max(0, -context_start))
-                              .set("is_padding", is_padding)
-                              .set("total_pad", pad));
-
-  CpuMatrix cpu_in_grad(batch_size, input_dim);
-  cpu_in_grad.randomizeUniform();
-  GpuMatrix gpu_in_grad(batch_size, input_dim);
-  gpu_in_grad.copyFrom(cpu_in_grad);
-
-  CpuMatrix cpu_out_grad(batch_size, input_dim * context_length);
-  cpu_out_grad.randomizeUniform();
-  GpuMatrix gpu_out_grad(batch_size, input_dim * context_length);
-  gpu_out_grad.copyFrom(cpu_out_grad);
-
-  IVectorPtr cpu_seq;
-  generateSequenceStartPositions(batch_size, cpu_seq);
-  IVectorPtr gpu_seq = IVector::create(cpu_seq->getSize(), true);
-  gpu_seq->copyFrom(*cpu_seq);
-
-  auto cpu_w_grad =
-      is_padding ? std::make_shared<CpuMatrix>(pad, input_dim) : nullptr;
-  auto gpu_w_grad =
-      is_padding ? std::make_shared<GpuMatrix>(pad, input_dim) : nullptr;
-  if (is_padding) {
-    cpu_w_grad->randomizeUniform();
-    gpu_w_grad->copyFrom(*cpu_w_grad);
+  FunctionCompare test("ContextProjectionBackward",
+                       FuncConfig()
+                           .set("context_length", context_length)
+                           .set("context_start", context_start)
+                           .set("begin_pad", std::max(0, -context_start))
+                           .set("is_padding", is_padding)
+                           .set("total_pad", pad));
+
+  // prepare input arguments
+  test.addSequence(SequenceIdArg(TensorShape{batch_size}));
+  test.addInputs(SequenceArg(
+      VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim * context_length}));
+  test.addOutputs(
+      SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim}),
+      ADD_TO);
+  if (is_padding) {  // weight
+    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{pad, input_dim}),
+                    ADD_TO);
   }
 
-  BufferArgs cpu_inputs;
-  BufferArgs cpu_outputs;
-  cpu_inputs.addArg(cpu_out_grad, *cpu_seq);
-  cpu_outputs.addArg(cpu_in_grad, *cpu_seq, ADD_TO);
-  cpu_outputs.addArg(
-      cpu_w_grad ? *cpu_w_grad : CpuMatrix(nullptr, 0, input_dim), ADD_TO);
-
-  compare.getCpuFunction()->calc(cpu_inputs, cpu_outputs);
-
-  BufferArgs gpu_inputs;
-  BufferArgs gpu_outputs;
-  gpu_inputs.addArg(gpu_out_grad, *gpu_seq);
-  gpu_outputs.addArg(gpu_in_grad, *gpu_seq, ADD_TO);
-  gpu_outputs.addArg(
-      gpu_w_grad ? *gpu_w_grad : GpuMatrix(nullptr, 0, input_dim), ADD_TO);
-
-  compare.getGpuFunction()->calc(gpu_inputs, gpu_outputs);
-
-  autotest::TensorCheckErr(cpu_in_grad, gpu_in_grad);
-  if (is_padding) {
-    autotest::TensorCheckErr(*cpu_w_grad, *gpu_w_grad);
-  }
+  // run Function
+  test.run();
 }
 
-TEST(ContextProjection, projection) {
+TEST(ContextProjection, Projection) {
   for (auto context_start : {-5, -3, -1, 0, 3}) {
     for (auto context_length : {1, 2, 5, 7}) {
       for (auto trainable_padding : {false, true}) {

paddle/function/CosSimOp.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 "CosSimOp.h"
+#include "paddle/math/Matrix.h"
+#include "paddle/math/Vector.h"
+
+namespace paddle {
+/**
+ * Cosine Similarity for CpuMatrix
+ *
+ * \param out_mat, output value, size: nSamples * 1.
+ * \param in1_mat, input value 1, size: nSamples * dim.
+ * \param in2_mat, input value 2, size: n2 * dim (n2 == 1 or n2 == nSamples).
+ * \param scale, default 1.0
+ *
+ */
+template <>
+void CosSimForward<DEVICE_TYPE_CPU>(CpuMatrix& out_mat,
+                                    const CpuMatrix& in1_mat,
+                                    const CpuMatrix& in2_mat,
+                                    real scale) {
+  CHECK(out_mat.getData() && in1_mat.getData() && in2_mat.getData());
+  size_t num_samples = out_mat.getHeight();
+  size_t dim = in1_mat.getWidth();
+  /// column vector [nSamples, 1]
+  real* out = out_mat.getData();
+  const real* x = in1_mat.getData();
+  const real* y = in2_mat.getData();
+
+  /// in2 might only have one row or full rows
+  CHECK(in2_mat.getHeight() == 1LU || in2_mat.getHeight() == num_samples);
+  size_t inc = (in2_mat.getHeight() == 1LU) ? 0 : dim;
+  for (size_t i = 0; i < num_samples; ++i, x += dim, y += inc) {
+    real square_sum_x = 0;
+    real square_sum_y = 0;
+    real xy = 0;
+    for (size_t j = 0; j < dim; ++j) {
+      square_sum_x += x[j] * x[j];
+      square_sum_y += y[j] * y[j];
+      xy += x[j] * y[j];
+    }
+    CHECK(square_sum_x > 0 && square_sum_y > 0);
+    out[i] = scale * xy / (std::sqrt(square_sum_x) * std::sqrt(square_sum_y));
+  }
+}
+
+/**
+ * Cosine Similarity
+ * for each row i,
+ *   out[i] = scale * cos(input1[i], input2[i])
+ *      = scale * <input1[i], input2[i]>/sqrt(|input1[i]|^2 * |input2[i]|^2)
+ * when input2 only has one row, then for each row i,
+ *   out[i] = cos(input1[i], input2[0])
+ *
+ * \param inputs[0] input matrix 1, size: nSamples * dim.
+ * \param inputs[1] input matrix 2, size: n2 * dim (n2 == 1 or n2 == nSamples).
+ * \param outputs[0] output matrix, size : nSamples * 1.
+ */
+
+template <DeviceType Device>
+class CosSimForwardFunc : public FunctionBase {
+  void init(const FuncConfig& config) override {
+    scale_ = config.get<real>("scale");
+  }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(inputs.size(), 2UL);
+    CHECK_EQ(outputs.size(), 1UL);
+
+    CHECK_EQ(inputs[0].shape().ndims(), 2UL);
+    CHECK_EQ(inputs[1].shape().ndims(), 2UL);
+    CHECK_EQ(outputs[0].shape().ndims(), 2UL);
+
+    CHECK_EQ(inputs[0].shape()[0], outputs[0].shape()[0]);
+    CHECK_EQ(inputs[0].shape()[1], inputs[1].shape()[1]);
+    CHECK_EQ(outputs[0].shape()[1], 1UL);
+
+    CHECK(outputs[0].data() && inputs[0].data() && inputs[1].data());
+
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
+    auto out_mat = outputs[0].matrix<Device>();
+    const auto in1_mat = inputs[0].matrix<Device>();
+    const auto in2_mat = inputs[1].matrix<Device>();
+
+    CosSimForward<Device>(out_mat, in1_mat, in2_mat, scale_);
+  }
+
+private:
+  real scale_;
+};
+
+/**
+ * Cosine Similarity Derivative for CpuMatrix
+ *
+ * \param in1_grad  forward input grad 1, size: nSamples * dim.
+ * \param in2_grad  forward input grad 2,
+ *                  size: n2 * dim (n2 == 1 or n2 == nSamples).
+ *
+ * \param out_grad  backward loss output grad, size : nSamples * 1.
+ * \param out_val   forward output value, size: nSamples * 1.
+ * \param in1_val   forward input value 1, size: nSamples * dim.
+ * \param in2_val   forward input value 2,
+ *                  size: n2 * dim (n2 == 1 or n2 == nSamples).
+ * \param scale,    default 1.0
+ */
+template <>
+void CosSimBackward<DEVICE_TYPE_CPU>(const CpuMatrix& out_grad,
+                                     const CpuMatrix& out_val,
+                                     const CpuMatrix& in1_val,
+                                     const CpuMatrix& in2_val,
+                                     CpuMatrix& in1_grad,
+                                     CpuMatrix& in2_grad,
+                                     real scale) {
+  CHECK(out_grad.getData() && out_val.getData() && in1_val.getData() &&
+        in2_val.getData() && in1_grad.getData() && in2_grad.getData());
+  CHECK_EQ(out_val.useGpu_, false) << "Matrix type are GPU, CPU required";
+
+  const real* grad = out_grad.getData();
+  const real* out = out_val.getData();
+  const real* prev_out_x = in1_val.getData();
+  const real* prev_out_y = in2_val.getData();
+  real* prev_grad_x = in1_grad.getData();
+  real* prev_grad_y = in2_grad.getData();
+
+  size_t num_samples = out_grad.getHeight();
+  size_t dim = in1_val.getWidth();
+  CHECK_EQ(in2_val.getHeight(), in2_grad.getHeight());
+  CHECK(in2_val.getHeight() == 1LU || in2_val.getHeight() == num_samples);
+  size_t inc = (in2_val.getHeight() == 1LU) ? 0 : dim;
+  for (size_t i = 0; i < num_samples; ++i,
+              prev_out_x += dim,
+              prev_out_y += inc,
+              prev_grad_x += dim,
+              prev_grad_y += inc) {
+    real square_sum_x = 0;
+    real square_sum_y = 0;
+    real xy = 0;
+    for (size_t j = 0; j < dim; ++j) {
+      square_sum_x += prev_out_x[j] * prev_out_x[j];
+      square_sum_y += prev_out_y[j] * prev_out_y[j];
+      xy += prev_out_x[j] * prev_out_y[j];
+    }
+    CHECK(square_sum_x > 0 && square_sum_y > 0);
+    if (xy == 0) {
+      real reciprocal =
+          1.0f / (std::sqrt(square_sum_x) * std::sqrt(square_sum_y));
+      for (size_t j = 0; j < dim; ++j) {
+        prev_grad_x[j] += scale * grad[i] * prev_out_y[j] * reciprocal;
+        prev_grad_y[j] += scale * grad[i] * prev_out_x[j] * reciprocal;
+      }
+    } else {
+      real reciprocal_xy = 1.0f / xy;
+      real reciprocal_square_sum_x = 1.0f / square_sum_x;
+      real reciprocal_square_sum_y = 1.0f / square_sum_y;
+      for (size_t j = 0; j < dim; ++j) {
+        prev_grad_x[j] +=
+            out[i] * grad[i] * (prev_out_y[j] * reciprocal_xy -
+                                prev_out_x[j] * reciprocal_square_sum_x);
+        prev_grad_y[j] +=
+            out[i] * grad[i] * (prev_out_x[j] * reciprocal_xy -
+                                prev_out_y[j] * reciprocal_square_sum_y);
+      }
+    }
+  }
+}
+
+/**
+ * Cosine Similarity backward Derivative
+ *
+ * \param outputs[0] forward input grad 1, size: nSamples * dim.
+ * \param outputs[1] forward input grad 2,
+ *                  size: n2 * dim (n2 == 1 or n2 == nSamples).
+ *
+ * \param inputs[0] backward loss output grad, size : nSamples * 1.
+ * \param inputs[1] forward output value, size: nSamples * 1.
+ * \param inputs[2] forward input value 1, size: nSamples * dim.
+ * \param inputs[3] forward input value 2,
+ *                  size: n2 * dim (n2 == 1 or n2 == nSamples).
+ */
+template <DeviceType Device>
+class CosSimBackwardFunc : public FunctionBase {
+  void init(const FuncConfig& config) override {
+    scale_ = config.get<real>("scale");
+  }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(inputs.size(), 4UL);
+    CHECK_EQ(outputs.size(), 2UL);
+    /// dim of out_grad and out_val == 1, column vector
+    CHECK_EQ(inputs[0].shape()[1], 1UL);
+    CHECK_EQ(inputs[1].shape()[1], 1UL);
+    /// nSamples of out_grad == out_val == in_val1 == in_grad1
+    CHECK_EQ(inputs[1].shape()[0], inputs[0].shape()[0]);
+    CHECK_EQ(inputs[0].shape()[0], inputs[0].shape()[0]);
+    CHECK_EQ(outputs[0].shape()[0], inputs[0].shape()[0]);
+    /// dim of in1_val1 == in_val2 == in_grad1 == in_grad2
+    CHECK_EQ(inputs[3].shape()[1], inputs[2].shape()[1]);
+    CHECK_EQ(outputs[0].shape()[1], inputs[2].shape()[1]);
+    CHECK_EQ(outputs[1].shape()[1], inputs[2].shape()[1]);
+
+    CHECK(inputs[0].data() && inputs[1].data() && inputs[2].data() &&
+          inputs[3].data() && outputs[0].data() && outputs[1].data());
+
+    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
+    CHECK_EQ(outputs[1].getArgType(), ADD_TO);
+
+    const auto out_grad = inputs[0].matrix<Device>();
+    const auto out_val = inputs[1].matrix<Device>();
+    const auto in1_val = inputs[2].matrix<Device>();
+    const auto in2_val = inputs[3].matrix<Device>();
+    auto in1_grad = outputs[0].matrix<Device>();
+    auto in2_grad = outputs[1].matrix<Device>();
+
+    CosSimBackward<Device>(
+        out_grad, out_val, in1_val, in2_val, in1_grad, in2_grad, scale_);
+  }
+
+private:
+  real scale_;
+};
+
+REGISTER_TYPED_FUNC(CosSimForward, CPU, CosSimForwardFunc);
+REGISTER_TYPED_FUNC(CosSimBackward, CPU, CosSimBackwardFunc);
+#ifndef PADDLE_ONLY_CPU
+REGISTER_TYPED_FUNC(CosSimForward, GPU, CosSimForwardFunc);
+REGISTER_TYPED_FUNC(CosSimBackward, GPU, CosSimBackwardFunc);
+#endif
+}  // namespace paddle

paddle/function/CosSimOp.h

+/* 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. */
+
+#pragma once
+
+#include "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief   Cosine Similarity Forward.
+ * for each row i,
+ * out[i] = scale * cos(in1[i], in2[i])
+ *        = scale * \sum_j (in1[i][j] * in2[i][j]) /
+ *                  sqrt(sum_j (in1[i][j]^2) * sum_j (in2[i][j])^2)
+ *
+ * \param[out]  output            output value.
+ * \param[in]   intput1           input value.
+ * \param[in]   intput2           input value.
+ * \param[in]   scale             default 1.0.
+ *
+ */
+template <DeviceType Device>
+void CosSimForward(typename Tensor<real, Device>::Matrix& output,
+                   const typename Tensor<real, Device>::Matrix& input1,
+                   const typename Tensor<real, Device>::Matrix& input2,
+                   real scale);
+
+/**
+ * \brief   Cosine Similarity BackWard for Derivative.
+ *
+ * \param[in]       output grad           backward loss output grad.
+ * \param[in]       output val            forward-output value.
+ * \param[in]       input val1            forward input value 1.
+ * \param[in]       input val2            forward input value 2.
+ * \param[in/out]   input grad            forward input grad 1.
+ * \param[in/out]   input grad            forward input grad 2.
+ * \param[in]       scale                 default 1.0.
+ *
+ */
+template <DeviceType Device>
+void CosSimBackward(const typename Tensor<real, Device>::Matrix& out_grad,
+                    const typename Tensor<real, Device>::Matrix& out_value,
+                    const typename Tensor<real, Device>::Matrix& in1_value,
+                    const typename Tensor<real, Device>::Matrix& in2_value,
+                    typename Tensor<real, Device>::Matrix& in1_grad,
+                    typename Tensor<real, Device>::Matrix& in2_grad,
+                    real scale);
+
+}  // namespace paddle

paddle/function/CosSimOpGpu.cu

+/* 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 "hl_base.h"
+#include "hl_device_functions.cuh"
+#include "CosSimOp.h"
+
+namespace paddle {
+
+template<int block_size>
+__global__ void KeCosSim(real* output,
+                         const real* input1,
+                         const real* input2,
+                         int width,
+                         int input1_height,
+                         int input2_height,
+                         real scale) {
+  const int ty = blockIdx.y;
+  int tid = threadIdx.x;
+
+  __shared__ real xx[block_size];
+  __shared__ real yy[block_size];
+  __shared__ real xy[block_size];
+
+  xx[tid] = 0.0;
+  yy[tid] = 0.0;
+  xy[tid] = 0.0;
+  __syncthreads();
+
+  input1 += ty * width;
+  if (input2_height > 1) {
+    input2 += ty * width;
+  }
+  for (int index = tid; index < width; index += block_size) {
+    real x = input1[index];
+    real y = input2[index];
+    xx[tid] += x * x;
+    yy[tid] += y * y;
+    xy[tid] += x * y;
+  }
+  __syncthreads();
+
+  for (int s = block_size / 2; s > 0; s >>= 1) {
+    if (tid < s) {
+      xx[tid] += xx[tid + s];
+      yy[tid] += yy[tid + s];
+      xy[tid] += xy[tid + s];
+    }
+    __syncthreads();
+  }
+  if (tid == 0) {
+    output[ty] = scale * xy[0] / (sqrt(xx[0]) * sqrt(yy[0]));
+  }
+}
+
+void hlCossim(real* output,
+              const real* input1,
+              const real* input2,
+              size_t width,
+              size_t input1_height,
+              size_t input2_height,
+              real scale) {
+  CHECK_NOTNULL(output);
+  CHECK_NOTNULL(input1);
+  CHECK_NOTNULL(input2);
+  const int block_size = 256;
+  dim3 threads(block_size, 1);
+  dim3 grid(1, input1_height);
+
+  KeCosSim<block_size><<<grid, threads, 0, STREAM_DEFAULT>>>
+    (output, input1, input2, width, input1_height, input2_height, scale);
+  CHECK_SYNC("hlCossim failed");
+}
+
+template <>
+void CosSimForward<DEVICE_TYPE_GPU>(GpuMatrix& out_mat,
+                                    const GpuMatrix& in1_mat,
+                                    const GpuMatrix& in2_mat,
+                                    real scale) {
+  CHECK(out_mat.getData() && in1_mat.getData() && in2_mat.getData());
+  CHECK(in1_mat.useGpu_ == true && in2_mat.useGpu_ == true)
+      << "Matrix type are not GPU";
+
+  size_t num_samples = out_mat.getHeight();
+  size_t dim = in1_mat.getWidth();
+  real* out = out_mat.getData();
+  const real* x = in1_mat.getData();
+  const real* y = in2_mat.getData();
+  hlCossim(out, x, y, dim, in1_mat.getHeight(), in2_mat.getHeight(), scale);
+}
+
+template<int block_size>
+__global__ void KeCosSimDerivative(const real* grad,
+                                   const real* output,
+                                   const real* prev_out_x,
+                                   const real* prev_out_y,
+                                   real* prev_grad_x,
+                                   real* prev_grad_y,
+                                   size_t width,
+                                   size_t input1_height,
+                                   size_t input2_height,
+                                   real scale) {
+  const int ty = blockIdx.y;
+  int tid = threadIdx.x;
+
+  __shared__ real xx[block_size];
+  __shared__ real yy[block_size];
+  __shared__ real xy[block_size];
+
+  xx[tid] = 0.0;
+  yy[tid] = 0.0;
+  xy[tid] = 0.0;
+  __syncthreads();
+
+  prev_out_x += ty * width;
+  prev_grad_x += ty * width;
+  if (input2_height > 1) {
+    prev_out_y += ty * width;
+    prev_grad_y += ty * width;
+  }
+  for (int index = tid; index < width; index += block_size) {
+    real x = prev_out_x[index];
+    real y = prev_out_y[index];
+    xx[tid] += x * x;
+    yy[tid] += y * y;
+    xy[tid] += x * y;
+  }
+  __syncthreads();
+
+  for (int s = block_size / 2; s > 0; s >>= 1) {
+    if (tid < s) {
+      xx[tid] += xx[tid + s];
+      yy[tid] += yy[tid + s];
+      xy[tid] += xy[tid + s];
+    }
+    __syncthreads();
+  }
+  if (xy[0] == 0) {
+    real reciprocal = 1.0 / (sqrt(xx[0]) * sqrt(yy[0]));
+    for (int index = tid; index < width; index += block_size) {
+      prev_grad_x[index] +=
+        scale * grad[ty] * prev_out_y[index] * reciprocal;
+      if (input2_height > 1) {
+        prev_grad_y[index] +=
+          scale * grad[ty] * prev_out_x[index] * reciprocal;
+      } else {
+        paddle::paddleAtomicAdd(prev_grad_y + index,
+          scale * grad[ty] * prev_out_x[index] * reciprocal);
+      }
+    }
+  } else {
+    real reciprocalXY = 1.0 / xy[0];
+    real reciprocalSquareSumX = 1.0 / xx[0];
+    real reciprocalSquareSumY = 1.0 / yy[0];
+    for (int index = tid; index < width; index += block_size) {
+      prev_grad_x[index] += output[ty] * grad[ty] *
+        (prev_out_y[index] * reciprocalXY -
+         prev_out_x[index] * reciprocalSquareSumX);
+      if (input2_height > 1) {
+        prev_grad_y[index] += output[ty] * grad[ty] *
+          (prev_out_x[index] * reciprocalXY -
+           prev_out_y[index] * reciprocalSquareSumY);
+      } else {
+        paddle::paddleAtomicAdd(prev_grad_y + index, output[ty] * grad[ty] *
+          (prev_out_x[index] * reciprocalXY -
+           prev_out_y[index] * reciprocalSquareSumY));
+      }
+    }
+  }
+}
+
+void hlCossimDerivative(const real* grad,
+                        const real* output,
+                        const real* prev_out_x,
+                        const real* prev_out_y,
+                        real* prev_grad_x,
+                        real* prev_grad_y,
+                        size_t width,
+                        size_t input1_height,
+                        size_t input2_height,
+                        real scale) {
+  CHECK_NOTNULL(grad);
+  CHECK_NOTNULL(output);
+  CHECK_NOTNULL(prev_out_x);
+  CHECK_NOTNULL(prev_out_y);
+  CHECK_NOTNULL(prev_grad_x);
+  CHECK_NOTNULL(prev_grad_y);
+  const int block_size = 256;
+  dim3 threads(block_size, 1);
+  dim3 grid(1, input1_height);
+  KeCosSimDerivative<block_size><<<grid, threads, 0, STREAM_DEFAULT>>>
+    (grad, output, prev_out_x, prev_out_y, prev_grad_x, prev_grad_y, width,
+        input1_height, input2_height, scale);
+  CHECK_SYNC("hlCossimDerivate failed");
+}
+
+template <>
+void CosSimBackward<DEVICE_TYPE_GPU>(const GpuMatrix& out_grad,
+                                     const GpuMatrix& out_val,
+                                     const GpuMatrix& in1_val,
+                                     const GpuMatrix& in2_val,
+                                     GpuMatrix& in1_grad,
+                                     GpuMatrix& in2_grad,
+                                     real scale) {
+  CHECK(out_grad.getData() && out_val.getData() && in1_val.getData() &&
+        in2_val.getData() && in1_grad.getData() && in2_grad.getData());
+  CHECK(out_grad.useGpu_ && out_val.useGpu_ && in1_val.useGpu_
+        && in2_val.useGpu_ && in1_grad.useGpu_ && in2_grad.useGpu_)
+        << "Matrix types are not equally GPU";
+
+  size_t dim = in1_val.getWidth();
+  const real* grad = out_grad.getData();
+  const real* out = out_val.getData();
+  const real* prev_out_x = in1_val.getData();
+  const real* prev_out_y = in2_val.getData();
+  real* prev_grad_x = in1_grad.getData();
+  real* prev_grad_y = in2_grad.getData();
+  hlCossimDerivative(grad,
+                     out,
+                     prev_out_x,
+                     prev_out_y,
+                     prev_grad_x,
+                     prev_grad_y,
+                     dim,
+                     in1_val.getHeight(),
+                     in2_val.getHeight(),
+                     scale);
+}
+
+}  // namespace paddle

paddle/function/CosSimOpTest.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 <gtest/gtest.h>
+#include "FunctionTest.h"
+#include "paddle/math/Matrix.h"
+
+using namespace paddle;  // NOLINT
+
+void testCosSimForward(size_t height_x,
+                       size_t height_y,
+                       size_t width,
+                       real scale) {
+  FunctionCompare test("CosSimForward", FuncConfig().set("scale", scale));
+  // prepare input arguments
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width}));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1}),
+                  ASSIGN_TO);
+  // run Function
+  test.run();
+}
+
+void testCosSimBackward(size_t height_x,
+                        size_t height_y,
+                        size_t width,
+                        real scale) {
+  FunctionCompare test("CosSimBackward", FuncConfig().set("scale", scale));
+  // prepare input arguments
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width}));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width}),
+                  ADD_TO);
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width}),
+                  ADD_TO);
+  // run Function
+  test.run();
+}
+
+TEST(Matrix, cosSim) {
+  for (auto height_x : {10, 100, 1000}) {
+    for (auto height_y : {1, height_x}) {
+      for (auto width : {10, 100, 1000}) {
+        for (auto scale : {1.0, 2.0}) {
+          testCosSimForward(height_x, height_y, width, scale);
+          testCosSimBackward(height_x, height_y, width, scale);
+        }
+      }
+    }
+  }
+}

paddle/function/FunctionTest.h

@@ -69,6 +69,54 @@ public:
         gpuMemory_.back()->getBuf(), input.valueType(), input.shape()));
   }
 
+  // assume one copy of sequence is shared by different SequenceArgs
+  void addSequence(const SequenceIdArg& input) {
+    CHECK_EQ(input.shape().ndims(), 1UL);
+    size_t batchSize = input.shape()[0];
+    size_t numSeqs = batchSize / 10 + 1;
+    size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32);
+    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(sizeId));
+    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(sizeId));
+    cpuSeq_ = std::make_shared<SequenceIdArg>(cpuMemory_.back()->getBuf(),
+                                              TensorShape{numSeqs + 1});
+    gpuSeq_ = std::make_shared<SequenceIdArg>(gpuMemory_.back()->getBuf(),
+                                              TensorShape{numSeqs + 1});
+    /// init sequence Id
+    initArg(*cpuSeq_, batchSize);
+
+    // todo(tianbing), delete it
+    CHECK_EQ(cpuSeq_->shape().getElements(), cpuSeq_->numSeqs() + 1);
+
+    CpuIVector cpuSeq(cpuSeq_->shape().getElements(), (int*)cpuSeq_->data());
+    GpuIVector gpuSeq(gpuSeq_->shape().getElements(), (int*)gpuSeq_->data());
+    gpuSeq.copyFrom(cpuSeq);
+  }
+
+  void addInputs(const SequenceArg& input) {
+    CHECK_EQ(input.shape().ndims(), 2UL);
+    size_t batchSize = input.shape()[0];
+    if (!cpuSeq_ || !gpuSeq_) {  // sequence not exist
+      addSequence(SequenceIdArg(TensorShape{batchSize}));
+    }
+
+    size_t size =
+        input.shape().getElements() * sizeOfValuType(input.valueType());
+    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
+    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
+
+    /// SequenceArg
+    cpuInputs_.emplace_back(
+        std::make_shared<SequenceArg>(cpuMemory_.back()->getBuf(),
+                                      input.valueType(),
+                                      input.shape(),
+                                      *cpuSeq_));
+    gpuInputs_.emplace_back(
+        std::make_shared<SequenceArg>(gpuMemory_.back()->getBuf(),
+                                      input.valueType(),
+                                      input.shape(),
+                                      *gpuSeq_));
+  }
+
   // output need only contains shape, do not contains data.
   void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) {
     size_t size =
@@ -116,24 +164,31 @@ public:
         std::make_shared<SparseMatrixArg>(*gpuSparse_, argType));
   }
 
-  void addInputs(const SequenceArg& input) {
-    size_t batchSize = input.shape()[0];
-    size_t numSeqs = batchSize / 10 + 1;
-
-    size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32);
-    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(sizeId));
-    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(sizeId));
-
-    TensorShape seqsId({numSeqs + 1});
-    // void* cpuBuffer = cpuMemory_.back()->getBuf();
-    // void* gpuBuffer = gpuMemory_.back()->getBuf();
+  void addOutputs(const SequenceArg& output, ArgType argType = ASSIGN_TO) {
+    CHECK_EQ(output.shape().ndims(), 2UL);
+    size_t batchSize = output.shape()[0];
 
+    if (!cpuSeq_ || !gpuSeq_) {  // sequence not exist
+      addSequence(SequenceIdArg(TensorShape{batchSize}));
+    }
     size_t size =
-        input.shape().getElements() * sizeOfValuType(input.valueType());
+        output.shape().getElements() * sizeOfValuType(output.valueType());
     cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
     gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
 
-    // TODO: need be implemented.
+    /// SequenceArg
+    cpuOutputs_.emplace_back(
+        std::make_shared<SequenceArg>(cpuMemory_.back()->getBuf(),
+                                      output.valueType(),
+                                      output.shape(),
+                                      *cpuSeq_,
+                                      argType));
+    gpuOutputs_.emplace_back(
+        std::make_shared<SequenceArg>(gpuMemory_.back()->getBuf(),
+                                      output.valueType(),
+                                      output.shape(),
+                                      *gpuSeq_,
+                                      argType));
   }
 
   void addInputs(const SparseMatrixArg& input) {
@@ -193,14 +248,44 @@ public:
   std::shared_ptr<FunctionBase> getGpuFunction() const { return gpuFunc_; }
 
 protected:
+  // only init cpu argument, gpu argument copy from cpu argument.
+  void initArg(BufferArg& arg) {
+    CpuVector vector(arg.shape().getElements(), (real*)arg.data());
+    vector.uniform(0.001, 1);
+  }
+
+  void initArg(SequenceArg& arg) {
+    /// init only matrix
+    CpuVector vector(arg.shape().getElements(), (real*)arg.data());
+    vector.uniform(0.001, 1);
+  }
+
+  void initArg(SequenceIdArg& arg, size_t batchSize) {
+    size_t numSeqs = arg.numSeqs();
+    int* buf = reinterpret_cast<int*>(arg.data());
+    int pos = 0;
+    size_t maxLen = 2 * batchSize / numSeqs;
+    for (int i = 0; i < (int)numSeqs; ++i) {
+      int len = 1 + uniformRandom(std::min<int64_t>(
+                        maxLen, batchSize - pos - numSeqs + i));
+      buf[i] = pos;
+      pos += len;
+      VLOG(1) << " len=" << len;
+    }
+    buf[numSeqs] = batchSize;
+  }
+
   void initInputs() {
     for (size_t i = 0; i < cpuInputs_.size(); i++) {
       if (cpuInputs_[i]->isSparseArg()) {
         continue;  /// sparse matrix already init
       }
 
-      initArg(*cpuInputs_[i]);
-
+      if (cpuInputs_[i]->isSequenceArg()) {
+        initArg(dynamic_cast<SequenceArg&>(*cpuInputs_[i]));
+      } else {
+        initArg(*cpuInputs_[i]);
+      }
       // TODO: Need a BufferCopy used to copy from one BufferArg to another.
       CpuVector cpuVector(cpuInputs_[i]->shape().getElements(),
                           (real*)cpuInputs_[i]->data());
@@ -217,7 +302,11 @@ protected:
         continue;  /// sparse matrix already init
       }
 
-      initArg(*cpuOutputs_[i]);
+      if (cpuOutputs_[i]->isSequenceArg()) {
+        initArg(dynamic_cast<SequenceArg&>(*cpuOutputs_[i]));
+      } else {
+        initArg(*cpuOutputs_[i]);
+      }
 
       // TODO: Need a BufferCopy used to copy from one BufferArg to another.
       CpuVector cpuVector(cpuOutputs_[i]->shape().getElements(),
@@ -241,28 +330,6 @@ protected:
     }
   }
 
-  // only init cpu argument, gpu argument copy from cpu argument.
-  void initArg(BufferArg& arg) {
-    CpuVector vector(arg.shape().getElements(), (real*)arg.data());
-    vector.uniform(0.001, 1);
-  }
-
-  void initArg(SequenceIdArg& arg, size_t batchSize) {
-    size_t numSeqs = arg.numSeqs();
-    int* buf = reinterpret_cast<int*>(arg.data());
-    int pos = 0;
-    size_t maxLen = 2 * batchSize / numSeqs;
-    for (int i = 0; i < (int)numSeqs; ++i) {
-      int len = uniformRandom(
-                    std::min<int64_t>(maxLen, batchSize - pos - numSeqs + i)) +
-                1;
-      buf[i] = pos;
-      pos += len;
-      VLOG(1) << " len=" << len;
-    }
-    buf[numSeqs] = batchSize;
-  }
-
 protected:
   std::shared_ptr<FunctionBase> cpuFunc_;
   std::shared_ptr<FunctionBase> gpuFunc_;
@@ -274,6 +341,8 @@ protected:
   std::vector<BufferArgPtr> gpuOutputs_;
   std::shared_ptr<CpuSparseMatrix> cpuSparse_;
   std::shared_ptr<GpuSparseMatrix> gpuSparse_;
+  std::shared_ptr<SequenceIdArg> cpuSeq_;
+  std::shared_ptr<SequenceIdArg> gpuSeq_;
 };
 
 }  // namespace paddle

paddle/gserver/dataproviders/PyDataProvider2.cpp

@@ -647,7 +647,7 @@ public:
       DataBatch& gpuBatch = *batch;
       std::vector<Argument>& gpuArguments = gpuBatch.getStreams();
       gpuArguments.resize(cpuArguments.size());
-      gpuBatch.setSize(size);
+      gpuBatch.setSize(bsize);
       for (size_t i = 0; i < headers_.size(); ++i) {
         gpuArguments[i].resizeAndCopyFrom(
             cpuArguments[i], useGpu_, HPPL_STREAM_1);

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -155,7 +155,8 @@ protected:
 public:
   explicit BootBiasLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     if (!Layer::init(layerMap, parameterMap)) return false;
 
     if (biasParameter_) {
@@ -174,7 +175,7 @@ public:
     }
   }
 
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     if (biases_) {
       MatrixPtr outV = getOutputValue();
       outV->addBias(*(biases_->getW()), 1);
@@ -182,7 +183,7 @@ public:
     }
   }
 
-  virtual void backward(const UpdateCallback& callback) {
+  void backward(const UpdateCallback& callback) override {
     if (biases_) {
       backwardActivation();
       biases_->getWGrad()->collectBias(*getOutputGrad(), 1);

paddle/gserver/layers/AddtoLayer.h

@@ -44,19 +44,20 @@ public:
   /**
    * Intialization of AddtoLayer.
    */
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * Forward propagation.
    * @note There is no weight matrix for each input,
    *       because it just a simple add operation.
    */
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
   /**
    * Backward propagation.
    */
-  void backward(const UpdateCallback& callback = nullptr);
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/AgentLayer.h

@@ -35,7 +35,8 @@ public:
 
   ~AgentLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   // if *numSamples* set,
   // real layer output will only use first *numSamples* rows
@@ -44,8 +45,8 @@ public:
     numSamples_ = numSamples;
   }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr) {}
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override {}
 };
 
 /**
@@ -56,8 +57,8 @@ public:
   explicit SequenceAgentLayer(const LayerConfig& config) : AgentLayer(config) {}
   ~SequenceAgentLayer() {}
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr) {}
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override {}
 };
 
 /**
@@ -78,7 +79,8 @@ public:
 
   virtual ~GatherAgentLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   // call before addRealLayer
   void copyIdAndSequenceInfo(const Argument& input,
@@ -88,8 +90,8 @@ public:
   // add one real layer, can call many times
   void addRealLayer(LayerPtr layer) { realLayers_.push_back(layer); }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 };
 
 /**
@@ -133,7 +135,8 @@ public:
 
   virtual ~ScatterAgentLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * @brief set real layer in generation
@@ -182,8 +185,8 @@ public:
     numSequences_ = numSequences;
   }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 };
 
 /**

paddle/gserver/layers/AverageLayer.h

@@ -38,12 +38,11 @@ public:
   explicit AverageLayer(const LayerConfig& config)
       : SequencePoolLayer(config) {}
 
-  ~AverageLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   MatrixPtr outMtx_;

paddle/gserver/layers/BatchNormBaseLayer.h

@@ -52,7 +52,8 @@ public:
    */
   static Layer* create(const LayerConfig& config);
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * @brief Calculate feature map size. Some input uses frameHeight and

paddle/gserver/layers/BatchNormalizationLayer.h

@@ -33,9 +33,10 @@ public:
 
   ~BatchNormalizationLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   /// Epsilon value used in the batch normalization formula.
@@ -58,7 +59,7 @@ protected:
   /// to batch, channels* imagePixels.
   void shrinkMat(const MatrixPtr& in, MatrixPtr& out);
 
-  void onPassEnd() { firstTest_ = true; }
+  void onPassEnd() override { firstTest_ = true; }
 
   MatrixPtr tmpMat_, tmpGrad_;
   MatrixPtr expandedIn_, expandedOut_;

paddle/gserver/layers/BilinearInterpLayer.h

@@ -38,9 +38,10 @@ public:
   virtual ~BilinearInterpLayer() {}
 
   size_t getSize();
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/BlockExpandLayer.h

@@ -58,10 +58,11 @@ public:
 
   ~BlockExpandLayer() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/CRFDecodingLayer.h

@@ -32,9 +32,10 @@ namespace paddle {
 class CRFDecodingLayer : public CRFLayer {
 public:
   explicit CRFDecodingLayer(const LayerConfig& config) : CRFLayer(config) {}
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   std::unique_ptr<LinearChainCRF> crf_;

paddle/gserver/layers/CRFLayer.h

@@ -29,9 +29,10 @@ namespace paddle {
 class CRFLayer : public Layer {
 public:
   explicit CRFLayer(const LayerConfig& config) : Layer(config) {}
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   size_t numClasses_;

paddle/gserver/layers/CTCLayer.h

@@ -22,10 +22,11 @@ namespace paddle {
 class CTCLayer : public Layer {
 public:
   explicit CTCLayer(const LayerConfig& config) : Layer(config) {}
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
   void forwardImp(const Argument& softmaxSeqs, const Argument& labelSeqs);
-  virtual void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
   void backwardImp(const UpdateCallback& callback,
                    const Argument& softmaxSeqs,
                    const Argument& labelSeqs);

paddle/gserver/layers/ConcatenateLayer.cpp

@@ -28,10 +28,11 @@ public:
 
   ~ConcatenateLayer() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(concat, ConcatenateLayer);
@@ -101,10 +102,11 @@ public:
 
   ~ConcatenateLayer2() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   std::vector<std::unique_ptr<Projection>> projections_;

paddle/gserver/layers/ConvBaseLayer.h

@@ -80,7 +80,8 @@ protected:
 public:
   explicit ConvBaseLayer(const LayerConfig& config) : Layer(config) {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * imgSizeH_ and imgSizeW_ will be set according to the previous input layers

paddle/gserver/layers/ConvShiftLayer.cpp

@@ -47,10 +47,11 @@ public:
 
   ~ConvShiftLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(conv_shift, ConvShiftLayer);

paddle/gserver/layers/ConvexCombinationLayer.cpp

@@ -49,10 +49,11 @@ public:
 
   ~ConvexCombinationLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(convex_comb, ConvexCombinationLayer);

paddle/gserver/layers/CosSimLayer.cpp

@@ -26,15 +26,23 @@ bool CosSimLayer::init(const LayerMap& layerMap,
   Layer::init(layerMap, parameterMap);
 
   CHECK_EQ(inputLayers_.size(), 2LU);
+
+  createFunction(forward_,
+                 "CosSimForward",
+                 FuncConfig().set("scale", (real)config_.cos_scale()));
+  createFunction(backward_,
+                 "CosSimBackward",
+                 FuncConfig().set("scale", (real)config_.cos_scale()));
+
   return true;
 }
 
 void CosSimLayer::forward(PassType passType) {
   Layer::forward(passType);
-
   /* malloc memory for the output_ if necessary */
   int batchSize = getInputValue(0)->getHeight();
   int size = getSize();
+  CHECK_EQ(forward_.size(), 1) << "Only one forward function needed";
 
   {
     REGISTER_TIMER_INFO("CosFwResetTimer", getName().c_str());
@@ -42,26 +50,43 @@ void CosSimLayer::forward(PassType passType) {
   }
 
   MatrixPtr outV = getOutputValue();
-
   /* activation */ {
     REGISTER_TIMER_INFO("CosFwAtvTimer", getName().c_str());
     MatrixPtr prevOut1 = getInputValue(0);
     MatrixPtr prevOut2 = getInputValue(1);
-    outV->cosSim(*prevOut1, *prevOut2, config_.cos_scale());
+
+    CHECK(outV && prevOut1 && prevOut2);
+    BufferArgs inputs;
+    BufferArgs outputs;
+    inputs.addArg(*prevOut1);
+    inputs.addArg(*prevOut2);
+    outputs.addArg(*outV, ASSIGN_TO);
+    forward_[0]->calc(inputs, outputs);
   }
 }
 
 void CosSimLayer::backward(const UpdateCallback& callback) {
   /* activation */ {
     REGISTER_TIMER_INFO("CosBpAtvTimer", getName().c_str());
-    MatrixPtr outG = this->getOutputGrad();
-
-    outG->cosSimDerivative(*this->getOutputValue(),
-                           *getInputValue(0),
-                           *getInputValue(1),
-                           *getInputGrad(0),
-                           *getInputGrad(1),
-                           config_.cos_scale());
+    CHECK_EQ(backward_.size(), 1) << "Only one backward function needed";
+
+    const auto outG = this->getOutputGrad();
+    const auto outV = this->getOutputValue();
+    const auto inV1 = this->getInputValue(0);
+    const auto inV2 = this->getInputValue(1);
+    auto inG1 = this->getInputGrad(0);
+    auto inG2 = this->getInputGrad(1);
+    CHECK(outG && outV && inV1 && inV2 && inG1 && inG2);
+    BufferArgs inputs;
+    BufferArgs outputs;
+    inputs.addArg(*outG);
+    inputs.addArg(*outV);
+    inputs.addArg(*inV1);
+    inputs.addArg(*inV2);
+    outputs.addArg(*inG1, ADD_TO);
+    outputs.addArg(*inG2, ADD_TO);
+
+    backward_[0]->calc(inputs, outputs);
   }
 }
 

paddle/gserver/layers/CosSimLayer.h

@@ -28,7 +28,7 @@ namespace paddle {
  *
  * - Input1: A vector (batchSize * dataDim) *
  * - Input2: A vector (batchSize * dataDim) or (1 * dataDim) *
- * - Output: A vector (dataDim * 1)
+ * - Output: A vector (batchSize * 1)
  *
  * The config file api is cos_sim.
  */
@@ -38,10 +38,11 @@ public:
 
   ~CosSimLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/CosSimVecMatLayer.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include "paddle/utils/Stat.h"
 
 namespace paddle {
-
 /**
  * @brief A layer for computing cosine similarity between a vector
  * and each row of a matrix
@@ -46,10 +45,11 @@ public:
 
   ~CosSimVecMatLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(cos_vm, CosSimVecMatLayer);
@@ -97,11 +97,22 @@ bool CosSimVecMatLayer::init(const LayerMap& layerMap,
                            dataDim,
                            /* trans= */ false,
                            useGpu_);
+
+  CHECK(tmpRow0 && tmpRow1 && tmpRow2 && tmpRow3 && tmpMtx0 && tmpMtx1);
+
+  createFunction(forward_,
+                 "CosSimForward",
+                 FuncConfig().set("scale", (real)config_.cos_scale()));
+  createFunction(backward_,
+                 "CosSimBackward",
+                 FuncConfig().set("scale", (real)config_.cos_scale()));
+
   return true;
 }
 
 void CosSimVecMatLayer::forward(PassType passType) {
   Layer::forward(passType);
+  CHECK_EQ(forward_.size(), 1) << "Only one forward function needed";
 
   MatrixPtr inV0 = getInputValue(0);
   MatrixPtr inV1 = getInputValue(1);
@@ -117,17 +128,25 @@ void CosSimVecMatLayer::forward(PassType passType) {
   }
 
   MatrixPtr outV = getOutputValue();
-
+  CHECK(outV && inV0 && inV1);
   REGISTER_TIMER_INFO("FwCosVMTimer", getName().c_str());
   for (size_t i = 0; i < batchSize; i++) {
     tmpRow0->setData(inV0->rowBuf(i));
     tmpMtx0->setData(inV1->rowBuf(i));
     tmpRow2->setData(outV->rowBuf(i));
-    tmpRow2->cosSim(*(tmpMtx0), *(tmpRow0), config_.cos_scale());
+
+    BufferArgs inputs;
+    BufferArgs outputs;
+    inputs.addArg(*tmpMtx0);
+    inputs.addArg(*tmpRow0);
+    outputs.addArg(*tmpRow2, ASSIGN_TO);
+    forward_[0]->calc(inputs, outputs);
   }
 }
 
 void CosSimVecMatLayer::backward(const UpdateCallback& callback) {
+  CHECK_EQ(backward_.size(), 1) << "Only one forward function needed";
+
   MatrixPtr inV0 = getInputValue(0);
   MatrixPtr inV1 = getInputValue(1);
   MatrixPtr inG0 = getInputGrad(0);
@@ -136,27 +155,27 @@ void CosSimVecMatLayer::backward(const UpdateCallback& callback) {
   MatrixPtr outG = getOutputGrad();
 
   size_t batchSize = inV0->getHeight();
-
+  CHECK(inV0 && inV1 && inG0 && inG1 && outV && outG);
   REGISTER_TIMER_INFO("BwCosVMTimer", getName().c_str());
 
-  if (inG0 && inG1) {
-    for (size_t i = 0; i < batchSize; i++) {
-      tmpRow0->setData(inV0->rowBuf(i));
-      tmpRow1->setData(inG0->rowBuf(i));
-      tmpMtx0->setData(inV1->rowBuf(i));
-      tmpMtx1->setData(inG1->rowBuf(i));
-      tmpRow2->setData(outV->rowBuf(i));
-      tmpRow3->setData(outG->rowBuf(i));
-
-      tmpRow3->cosSimDerivative(*(tmpRow2),
-                                *(tmpMtx0),
-                                *(tmpRow0),
-                                *(tmpMtx1),
-                                *(tmpRow1),
-                                config_.cos_scale());
-    }
-  } else {
-    CHECK(!inG0 || !inG1) << "Not supported";
+  for (size_t i = 0; i < batchSize; i++) {
+    tmpRow0->setData(inV0->rowBuf(i));
+    tmpRow1->setData(inG0->rowBuf(i));
+    tmpMtx0->setData(inV1->rowBuf(i));
+    tmpMtx1->setData(inG1->rowBuf(i));
+    tmpRow2->setData(outV->rowBuf(i));
+    tmpRow3->setData(outG->rowBuf(i));
+
+    BufferArgs inputs;
+    BufferArgs outputs;
+    inputs.addArg(*tmpRow3);
+    inputs.addArg(*tmpRow2);
+    inputs.addArg(*tmpMtx0);
+    inputs.addArg(*tmpRow0);
+    outputs.addArg(*tmpMtx1, ADD_TO);
+    outputs.addArg(*tmpRow1, ADD_TO);
+
+    backward_[0]->calc(inputs, outputs);
   }
 }
 

paddle/gserver/layers/CostLayer.cpp

@@ -367,8 +367,6 @@ void LambdaCost::backward(const UpdateCallback& callback) {
   getInputGrad(0)->add(*marginGrad_);
 }
 
-void LambdaCost::onPassEnd() {}
-
 void LambdaCost::calcGrad(const real* outputScore,
                           const real* score,
                           real* gradData,
@@ -611,14 +609,15 @@ class SumCostLayer : public Layer {
 public:
   explicit SumCostLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     bool ret = Layer::init(layerMap, parameterMap);
     if (!ret) return ret;
     CHECK_EQ(inputLayers_.size(), 1UL);
     return true;
   }
 
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
     const MatrixPtr& input = getInputValue(0);
 
@@ -629,7 +628,7 @@ public:
     output_.value->sumRows(*input, /* scaleSum= */ 1, /* scaleDest= */ 0);
   }
 
-  virtual void backward(const UpdateCallback& callback = nullptr) {
+  void backward(const UpdateCallback& callback = nullptr) override {
     getInputGrad(0)->add((real)1);
   }
 };

paddle/gserver/layers/CostLayer.h

@@ -32,15 +32,16 @@ class CostLayer : public Layer {
 public:
   explicit CostLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   LayerPtr getOutputLayer() { return inputLayers_[0]; }
 
   LayerPtr getLabelLayer() { return inputLayers_[1]; }
 
-  virtual void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void backward(const UpdateCallback& callback = nullptr) override;
 
   virtual void forwardImp(Matrix& outputValue,
                           Argument& label,
@@ -68,11 +69,14 @@ public:
   explicit MultiClassCrossEntropy(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 };
 
 /**
@@ -95,11 +99,14 @@ public:
   explicit MultiClassCrossEntropyWithSelfNorm(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 
 protected:
   MatrixPtr sftMaxSum_;
@@ -117,11 +124,14 @@ public:
   explicit SoftBinaryClassCrossEntropy(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 
 protected:
   MatrixPtr targetPerDim_;
@@ -139,11 +149,14 @@ public:
   explicit SumOfSquaresCostLayer(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 };
 
 /**
@@ -162,17 +175,18 @@ class RankingCost : public Layer {
 public:
   explicit RankingCost(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   LayerPtr getOutputLayer(size_t i) { return inputLayers_[i]; }
 
   LayerPtr getLabelLayer() { return inputLayers_[2]; }
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback = nullptr);
+  void backward(const UpdateCallback& callback = nullptr) override;
 
-  void onPassEnd();
+  void onPassEnd() override;
 
   void forwardImp(Matrix& output, Argument& label, Matrix& cost) {
     (void)output;
@@ -214,17 +228,16 @@ class LambdaCost : public Layer {
 public:
   explicit LambdaCost(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   LayerPtr getOutputLayer() { return inputLayers_[0]; }
 
   LayerPtr getScoreLayer() { return inputLayers_[1]; }
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback = nullptr);
-
-  void onPassEnd();
+  void backward(const UpdateCallback& callback = nullptr) override;
 
   real calcNDCG(const real* outputScore, const real* score, int size);
   void calcGrad(const real* outputScore,
@@ -256,11 +269,14 @@ public:
   explicit MultiBinaryLabelCrossEntropy(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 };
 
 /**
@@ -282,13 +298,16 @@ class HuberTwoClass : public CostLayer {
 public:
   explicit HuberTwoClass(const LayerConfig& config) : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
   void forwardImpIn(Matrix& output, Argument& label, Matrix& cost);
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 
   void backwardImpIn(Matrix& outputValue, Argument& label, Matrix& outputGrad);
 };

paddle/gserver/layers/CudnnBatchNormLayer.h

@@ -35,14 +35,15 @@ public:
 
   ~CudnnBatchNormLayer();
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
   /**
    * reshape tensor of ioDesc_.
    */
   void reshape(int batchSize);
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   /**

paddle/gserver/layers/CudnnConvLayer.h

@@ -45,9 +45,10 @@ public:
 
   ~CudnnConvLayer();
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
   void addBiases();
   void bpropBiases();
 };

paddle/gserver/layers/CudnnPoolLayer.h

@@ -45,7 +45,8 @@ public:
                         hl_pooling_mode_t* mode = nullptr);
   explicit CudnnPoolLayer(const LayerConfig& config);
   ~CudnnPoolLayer();
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * Reshape input and output tensor descriptor.
@@ -53,8 +54,8 @@ public:
    * So reshaping is needed.
    */
   void reshape(int batchSize);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/DataLayer.h

@@ -33,13 +33,13 @@ public:
   /**
    * Prefetch sparse matrix/ids only.
    */
-  void prefetch() { output_ = data_; }
+  void prefetch() override { output_ = data_; }
 
   /**
    * Forward propagation. Copy data_ (value, in, grad, ids, cpuSequenceDims,
    * sequenceStartPositions, subSequenceStartPositions, strs) to output_.
    */
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
     copyDataToOutput(output_);
     if (FLAGS_show_layer_stat) {
@@ -50,9 +50,9 @@ public:
   /**
    * Data layer's backward propagation do nothing.
    */
-  virtual void backward(const UpdateCallback& callback) { (void)callback; }
+  void backward(const UpdateCallback& callback) override { (void)callback; }
 
-  virtual void copyOutputToOtherDevice() {
+  void copyOutputToOtherDevice() override {
     for (size_t i = 0; i != outputOtherDevice_.size(); i++) {
       copyDataToOutput(outputOtherDevice_[i]);
     }

paddle/gserver/layers/DataNormLayer.h

@@ -44,10 +44,11 @@ public:
 
   ~DataNormLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   int mode_;

paddle/gserver/layers/EosIdCheckLayer.cpp

@@ -27,14 +27,14 @@ class EosIdCheckLayer : public Layer {
 public:
   explicit EosIdCheckLayer(const LayerConfig& config) : Layer(config) {}
 
-  virtual bool init(const LayerMap& layerMap,
-                    const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     bool ret = Layer::init(layerMap, parameterMap);
     CHECK_EQ(1UL, inputLayers_.size());
     return ret;
   }
 
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
 
     const Argument& input = getInput(0);
@@ -42,7 +42,7 @@ public:
     output_.ids->isEqualTo(*input.ids, config_.eos_id());
   }
 
-  virtual void backward(const UpdateCallback& callback) {}
+  void backward(const UpdateCallback& callback) override {}
 };
 
 REGISTER_LAYER(eos_id, EosIdCheckLayer);

paddle/gserver/layers/ExpandConvBaseLayer.h

@@ -48,7 +48,8 @@ public:
 
   ~ExpandConvBaseLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   size_t getOutputSize();
   /**

paddle/gserver/layers/ExpandConvLayer.h

@@ -35,10 +35,11 @@ public:
 
   ~ExpandConvLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/ExpandConvTransLayer.h

@@ -34,10 +34,11 @@ public:
 
   ~ExpandConvTransLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/ExpandLayer.h

@@ -53,10 +53,11 @@ public:
 
   ~ExpandLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/FeatureMapExpandLayer.cpp

@@ -46,10 +46,11 @@ public:
 
   ~FeatureMapExpandLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(featmap_expand, FeatureMapExpandLayer);

paddle/gserver/layers/FullyConnectedLayer.h

@@ -36,13 +36,14 @@ public:
   explicit FullyConnectedLayer(const LayerConfig& config) : Layer(config) {}
   ~FullyConnectedLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   Weight& getWeight(int idx) { return *weights_[idx]; }
 
-  void prefetch();
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void prefetch() override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/GatedRecurrentLayer.h

@@ -50,17 +50,18 @@ class GatedRecurrentLayer : public Layer, public GruCompute {
 public:
   explicit GatedRecurrentLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
 
-  void resetState();
+  void resetState() override;
 
-  void setState(LayerStatePtr state);
+  void setState(LayerStatePtr state) override;
 
-  LayerStatePtr getState();
+  LayerStatePtr getState() override;
 
 protected:
   void forwardSequence(int batchSize,

paddle/gserver/layers/GetOutputLayer.cpp

@@ -22,17 +22,18 @@ public:
 
   ~GetOutputLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     if (!Layer::init(layerMap, parameterMap)) return false;
     CHECK_EQ(1U, inputLayers_.size());
     CHECK_NE(inputArgument_[0], "");
     return true;
   }
 
-  void forward(PassType passType) {
+  void forward(PassType passType) override {
     output_ = getPrev(0)->getOutput(inputArgument_[0]);
   }
-  void backward(const UpdateCallback& callback = nullptr) {}
+  void backward(const UpdateCallback& callback = nullptr) override {}
 };
 
 REGISTER_LAYER(get_output, GetOutputLayer);

paddle/gserver/layers/GruStepLayer.cpp

@@ -55,10 +55,11 @@ public:
 
   ~GruStepLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(gru_step, GruStepLayer);

paddle/gserver/layers/HierarchicalSigmoidLayer.h

@@ -61,9 +61,10 @@ class HierarchicalSigmoidLayer : public Layer {
 public:
   explicit HierarchicalSigmoidLayer(const LayerConfig& config)
       : Layer(config) {}
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   /**

paddle/gserver/layers/InterpolationLayer.cpp

@@ -43,10 +43,11 @@ public:
 
   ~InterpolationLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(interpolation, InterpolationLayer);

paddle/gserver/layers/LstmLayer.h

@@ -74,17 +74,18 @@ class LstmLayer : public Layer, public LstmCompute {
 public:
   explicit LstmLayer(const LayerConfig &config) : Layer(config) {}
 
-  bool init(const LayerMap &layerMap, const ParameterMap &parameterMap);
+  bool init(const LayerMap &layerMap,
+            const ParameterMap &parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback &callback);
+  void backward(const UpdateCallback &callback) override;
 
-  void resetState();
+  void resetState() override;
 
-  void setState(LayerStatePtr state);
+  void setState(LayerStatePtr state) override;
 
-  LayerStatePtr getState();
+  LayerStatePtr getState() override;
 
 protected:
   /**

paddle/gserver/layers/LstmStepLayer.cpp

@@ -35,10 +35,11 @@ public:
 
   ~LstmStepLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(lstm_step, LstmStepLayer);

paddle/gserver/layers/MDLstmLayer.cpp

@@ -181,11 +181,12 @@ class MDLstmLayer : public LstmLayer {
 public:
   explicit MDLstmLayer(const LayerConfig& config) : LstmLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   void forwardOneSequence(int start, CoordIterator& coordIter);

paddle/gserver/layers/MaxIdLayer.cpp

@@ -30,8 +30,8 @@ private:
 public:
   explicit MaxIdLayer(const LayerConfig& config) : Layer(config) {}
 
-  virtual bool init(const LayerMap& layerMap,
-                    const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     bool ret = Layer::init(layerMap, parameterMap);
     CHECK_EQ(1UL, inputLayers_.size());
 
@@ -40,7 +40,7 @@ public:
     return ret;
   }
 
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
     const Argument& input = getInput(0);
     size_t batchSize = input.getBatchSize();
@@ -54,7 +54,7 @@ public:
     input.value->rowMax(*output_.ids, *output_.in);
   }
 
-  virtual void backward(const UpdateCallback& callback) {}
+  void backward(const UpdateCallback& callback) override {}
 };
 
 REGISTER_LAYER(maxid, MaxIdLayer);

paddle/gserver/layers/MaxLayer.h

@@ -42,14 +42,13 @@ protected:
 public:
   explicit MaxLayer(const LayerConfig& config) : SequencePoolLayer(config) {}
 
-  ~MaxLayer() {}
-
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     return SequencePoolLayer::init(layerMap, parameterMap);
   }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/MaxOutLayer.h

@@ -45,10 +45,11 @@ public:
   explicit MaxOutLayer(const LayerConfig& config) : Layer(config) {}
   virtual ~MaxOutLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/MixedLayer.h

@@ -35,21 +35,22 @@ public:
 
   ~MixedLayer() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  virtual void prefetch();
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
-  virtual void resetState();
+  void prefetch() override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+  void resetState() override;
   /**
    * setState() should be called after getState().
    * Argument state consists of all projections states.
    */
-  virtual void setState(LayerStatePtr state);
+  void setState(LayerStatePtr state) override;
   /**
    * Return state which consists of all projections states.
    */
-  virtual LayerStatePtr getState();
+  LayerStatePtr getState() override;
 
 protected:
   std::vector<std::unique_ptr<Projection>> projections_;

paddle/gserver/layers/MultiplexLayer.cpp

@@ -69,10 +69,11 @@ public:
 
   ~MultiplexLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 private:
   /**

paddle/gserver/layers/NCELayer.cpp

@@ -61,7 +61,8 @@ public:
         rand_(0, config.num_classes() - 1),
         prepared_(false) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     /* Initialize the basic parent class */
     Layer::init(layerMap, parameterMap);
 
@@ -146,7 +147,7 @@ public:
     prepared_ = true;
   }
 
-  void prefetch() {
+  void prefetch() override {
     prepareSamples();
     IVector::resizeOrCreate(labelIds_, samples_.size(), useGpu_);
     int* ids = labelIds_->getData();
@@ -163,7 +164,7 @@ public:
     }
   }
 
-  void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
 
     CHECK(!useGpu_) << "GPU is not supported";
@@ -199,7 +200,7 @@ public:
     forwardCost();
   }
 
-  void backward(const UpdateCallback& callback) {
+  void backward(const UpdateCallback& callback) override {
     Matrix::resizeOrCreate(sampleOut_.grad,
                            1,
                            samples_.size(),

paddle/gserver/layers/NormLayer.h

@@ -30,7 +30,8 @@ class NormLayer : public Layer {
 public:
   explicit NormLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     Layer::init(layerMap, parameterMap);
     return true;
   }
@@ -56,9 +57,10 @@ protected:
 public:
   explicit ResponseNormLayer(const LayerConfig& config) : NormLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType) { LOG(FATAL) << "Not implemented"; }
-  void backward(const UpdateCallback& callback = nullptr) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override { LOG(FATAL) << "Not implemented"; }
+  void backward(const UpdateCallback& callback = nullptr) override {
     LOG(FATAL) << "Not implemented";
   }
 };

paddle/gserver/layers/NormProjectionLayer.h

@@ -36,9 +36,10 @@ public:
 
   size_t getSize();
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   TensorShape shape_;

paddle/gserver/layers/OuterProdLayer.cpp

@@ -38,10 +38,11 @@ public:
 
   ~OuterProdLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(out_prod, OuterProdLayer);

paddle/gserver/layers/PadLayer.h

@@ -29,9 +29,10 @@ public:
 
   ~PadLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   void setOutDims(const size_t batchSize);

paddle/gserver/layers/ParameterReluLayer.h

@@ -56,9 +56,10 @@ public:
 
   ~ParameterReluLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 }  // namespace paddle

paddle/gserver/layers/PoolLayer.h

@@ -46,7 +46,8 @@ public:
    */
   static Layer* create(const LayerConfig& config);
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/PoolProjectionLayer.h

@@ -40,7 +40,7 @@ public:
 
   size_t getSize();
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 }  // namespace paddle

paddle/gserver/layers/PowerLayer.cpp

@@ -40,10 +40,11 @@ public:
 
   ~PowerLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(power, PowerLayer);

paddle/gserver/layers/PrintLayer.cpp

@@ -19,8 +19,8 @@ namespace paddle {
 class PrintLayer : public Layer {
 public:
   explicit PrintLayer(const LayerConfig& config) : Layer(config) {}
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback) {}
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override {}
 };
 
 void PrintLayer::forward(PassType passType) {

paddle/gserver/layers/PriorBox.cpp

@@ -30,10 +30,11 @@ namespace paddle {
 class PriorBoxLayer : public Layer {
 public:
   explicit PriorBoxLayer(const LayerConfig& config) : Layer(config) {}
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback) {}
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override {}
 
 protected:
   int numPriors_;

paddle/gserver/layers/RecurrentLayer.cpp

@@ -45,17 +45,18 @@ class RecurrentLayer : public Layer {
 public:
   explicit RecurrentLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
 
-  void resetState();
+  void resetState() override;
 
-  void setState(LayerStatePtr state);
+  void setState(LayerStatePtr state) override;
 
-  LayerStatePtr getState();
+  LayerStatePtr getState() override;
 
 protected:
   /**

paddle/gserver/layers/RecurrentLayerGroup.cpp

@@ -33,15 +33,15 @@ public:
   void initSubNetwork(NeuralNetwork* rootNetwork,
                       const ModelConfig& config,
                       const std::vector<ParameterType>& parameterTypes,
-                      bool useGpu);
+                      bool useGpu) override;
 
-  void forward(PassType passType) {
+  void forward(PassType passType) override {
     REGISTER_TIMER_INFO("RecurrentGroupFwTime", getName().c_str());
     const std::vector<Argument> inArgs;
     std::vector<Argument> outArgs;
     network_->forward(inArgs, &outArgs, passType);
   }
-  void backward(const UpdateCallback& callback) {
+  void backward(const UpdateCallback& callback) override {
     REGISTER_TIMER_INFO("RecurrentGroupBwTime", getName().c_str());
     network_->backward(nullptr);
 
@@ -53,7 +53,8 @@ public:
   /**
    * @see Layer.accessSubNetwork
    */
-  void accessSubNetwork(const std::function<void(NeuralNetwork&)>& callback) {
+  void accessSubNetwork(
+      const std::function<void(NeuralNetwork&)>& callback) override {
     callback(*network_);
   }
 

paddle/gserver/layers/ResizeLayer.cpp

@@ -20,18 +20,19 @@ namespace paddle {
 /**
  * @brief A layer for resizing a minibatch matrix h*w to h'*w'
  * @note
- * origin matrix height * witdth)
+ * origin matrix height * width)
  * resize matrix: (height * width / size) * size
  */
 class ResizeLayer : public Layer {
 public:
   explicit ResizeLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
 };
 
 REGISTER_LAYER(resize, ResizeLayer);

paddle/gserver/layers/SamplingIdLayer.cpp

@@ -35,8 +35,8 @@ public:
   explicit SamplingIdLayer(const LayerConfig& config)
       : Layer(config), rand1_(0, 1) {}
 
-  virtual bool init(const LayerMap& layerMap,
-                    const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     bool ret = Layer::init(layerMap, parameterMap);
     CHECK_EQ(1UL, inputLayers_.size());
     if (useGpu_) {
@@ -48,7 +48,7 @@ public:
     return ret;
   }
 
-  void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
     if (useGpu_) {
       for (size_t i = 0; i < inputLayers_.size(); i++) {
@@ -83,7 +83,7 @@ public:
     output_.ids->copyFrom(ids.data(), batchSize);
   }
 
-  virtual void backward(const UpdateCallback& callback) {}
+  void backward(const UpdateCallback& callback) override {}
 };
 
 REGISTER_LAYER(sampling_id, SamplingIdLayer);

paddle/gserver/layers/ScalingLayer.cpp

@@ -37,10 +37,11 @@ public:
 
   ~ScalingLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(scaling, ScalingLayer);

paddle/gserver/layers/SelectiveFullyConnectedLayer.h

@@ -65,9 +65,10 @@ public:
       : Layer(config), selCols_(nullptr) {}
 
   ~SelectiveFullyConnectedLayer() {}
-  void prefetch();
+  void prefetch() override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   Weight& getWeight(int idx) { return *weights_[idx]; }
 
@@ -90,8 +91,8 @@ public:
   void fillSelectiveData(
       const std::shared_ptr<std::vector<std::pair<int*, size_t>>>& candidates);
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 private:
   /**

paddle/gserver/layers/SequenceConcatLayer.cpp

@@ -35,10 +35,11 @@ public:
 
   ~SequenceConcatLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(seqconcat, SequenceConcatLayer);

paddle/gserver/layers/SequenceLastInstanceLayer.cpp

@@ -42,12 +42,11 @@ public:
   explicit SequenceLastInstanceLayer(const LayerConfig& config)
       : SequencePoolLayer(config) {}
 
-  ~SequenceLastInstanceLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(seqlastins, SequenceLastInstanceLayer);

paddle/gserver/layers/SequencePoolLayer.h

@@ -46,12 +46,11 @@ protected:
 public:
   explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {}
 
-  virtual ~SequencePoolLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/SequenceReshapeLayer.cpp

@@ -34,12 +34,11 @@ protected:
 public:
   explicit SequenceReshapeLayer(const LayerConfig& config) : Layer(config) {}
 
-  ~SequenceReshapeLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(seqreshape, SequenceReshapeLayer);

paddle/gserver/layers/SlopeInterceptLayer.cpp

@@ -39,12 +39,11 @@ class SlopeInterceptLayer : public Layer {
 public:
   explicit SlopeInterceptLayer(const LayerConfig& config) : Layer(config) {}
 
-  ~SlopeInterceptLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(slope_intercept, SlopeInterceptLayer);

paddle/gserver/layers/SpatialPyramidPoolLayer.h

@@ -43,9 +43,8 @@ protected:
 public:
   explicit SpatialPyramidPoolLayer(const LayerConfig& config) : Layer(config) {}
 
-  ~SpatialPyramidPoolLayer() {}
-
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   ProjectionConfig getConfig(size_t sizeX_,
                              size_t sizeY_,
@@ -54,7 +53,7 @@ public:
                              std::string& poolType_);
   size_t getSize();
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 }  // namespace paddle

paddle/gserver/layers/SubSequenceLayer.cpp

@@ -35,12 +35,11 @@ protected:
 public:
   explicit SubSequenceLayer(const LayerConfig& config) : Layer(config) {}
 
-  ~SubSequenceLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(subseq, SubSequenceLayer);

paddle/gserver/layers/SumToOneNormLayer.cpp

@@ -41,12 +41,11 @@ protected:
 public:
   explicit SumToOneNormLayer(const LayerConfig& config) : Layer(config) {}
 
-  ~SumToOneNormLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(sum_to_one_norm, SumToOneNormLayer);

paddle/gserver/layers/TensorLayer.h

@@ -44,13 +44,12 @@ protected:
 public:
   explicit TensorLayer(const LayerConfig& config) : Layer(config) {}
 
-  ~TensorLayer() {}
-
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   Weight& getWeight(int idx) { return *weights_[idx]; }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 }  // namespace paddle

paddle/gserver/layers/TransLayer.h

@@ -32,9 +32,10 @@ class TransLayer : public Layer {
 public:
   explicit TransLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 }  // namespace paddle

paddle/gserver/layers/ValidationLayer.h

@@ -26,7 +26,8 @@ class ValidationLayer : public Layer {
 public:
   explicit ValidationLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   LayerPtr getOutputLayer() { return inputLayers_[0]; }
 
@@ -37,13 +38,13 @@ public:
     return inputLayers_[2];
   }
 
-  virtual void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void backward(const UpdateCallback& callback = nullptr) override;
 
   virtual void validationImp(MatrixPtr outputValue, IVectorPtr label) = 0;
 
-  virtual void onPassEnd() = 0;
+  void onPassEnd() override = 0;
 };
 
 /*
@@ -57,11 +58,12 @@ public:
         cpuLabel_(nullptr),
         cpuWeight_(nullptr) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void validationImp(MatrixPtr outputValue, IVectorPtr label);
+  void validationImp(MatrixPtr outputValue, IVectorPtr label) override;
 
-  void onPassEnd();
+  void onPassEnd() override;
 
   struct PredictionResult {
     PredictionResult(real __out, int __label) : out(__out), label(__label) {}
@@ -86,11 +88,12 @@ public:
   explicit PnpairValidation(const LayerConfig& config)
       : ValidationLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void validationImp(MatrixPtr outputValue, IVectorPtr label);
+  void validationImp(MatrixPtr outputValue, IVectorPtr label) override;
 
-  void onPassEnd();
+  void onPassEnd() override;
 
 private:
   bool passBegin_;

paddle/gserver/layers/WarpCTCLayer.h

@@ -30,9 +30,10 @@ public:
   explicit WarpCTCLayer(const LayerConfig& config) : Layer(config) {}
   ~WarpCTCLayer() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   /**

paddle/math/Matrix.cpp

@@ -941,59 +941,6 @@ void GpuMatrix::softreluDerivative(Matrix& output) {
 void GpuMatrix::scaledTanh(Matrix& output, real p1, real p2) {
   BaseMatrix::scaledTanh(output, p1, p2);
 }
-void GpuMatrix::cosSim(Matrix& output1, Matrix& output2, real scale) {
-  CHECK(output1.useGpu_ == true && output2.useGpu_ == true)
-      << "Matrix type are not equal";
-  size_t numSamples = getHeight();
-  size_t dim = output1.getWidth();
-  CHECK_EQ(getWidth(), 1UL);
-  CHECK_EQ(output1.getHeight(), numSamples);
-  CHECK_EQ(output1.getWidth(), output2.getWidth());
-  real* out = getData();
-  real* x = output1.getData();
-  real* y = output2.getData();
-  hl_cossim(out, x, y, dim, output1.getHeight(), output2.getHeight(), scale);
-}
-void GpuMatrix::cosSimDerivative(Matrix& output,
-                                 Matrix& prevOut1,
-                                 Matrix& prevOut2,
-                                 Matrix& prevGrad1,
-                                 Matrix& prevGrad2,
-                                 real scale) {
-  CHECK(output.useGpu_ == true && prevOut1.useGpu_ == true &&
-        prevOut2.useGpu_ == true && prevGrad1.useGpu_ == true &&
-        prevGrad2.useGpu_ == true)
-      << "Matrix type are not equal";
-  CHECK_EQ(getWidth(), 1UL);
-  CHECK_EQ(output.getWidth(), 1UL);
-
-  size_t numSamples = getHeight();
-  CHECK_EQ(output.getHeight(), numSamples);
-  CHECK_EQ(prevOut1.getHeight(), numSamples);
-  CHECK_EQ(prevGrad1.getHeight(), numSamples);
-
-  size_t dim = prevOut1.getWidth();
-  CHECK_EQ(prevOut2.getWidth(), dim);
-  CHECK_EQ(prevGrad1.getWidth(), dim);
-  CHECK_EQ(prevGrad2.getWidth(), dim);
-
-  real* grad = getData();
-  real* out = output.getData();
-  real* prevOutX = prevOut1.getData();
-  real* prevOutY = prevOut2.getData();
-  real* prevGradX = prevGrad1.getData();
-  real* prevGradY = prevGrad2.getData();
-  hl_cossim_derivative(grad,
-                       out,
-                       prevOutX,
-                       prevOutY,
-                       prevGradX,
-                       prevGradY,
-                       dim,
-                       prevOut1.getHeight(),
-                       prevOut2.getHeight(),
-                       scale);
-}
 
 void GpuMatrix::randomizeUniform() {
   CHECK(isContiguous());
@@ -3470,105 +3417,6 @@ void CpuMatrix::softmaxDerivative(Matrix& output, Matrix& sftmaxSum) {
   }
 }
 
-void CpuMatrix::cosSim(Matrix& output1, Matrix& output2, real scale) {
-  size_t numSamples = getHeight();
-  size_t dim = output1.getWidth();
-  CHECK_EQ(getWidth(), 1UL);
-  CHECK_EQ(output1.getHeight(), numSamples);
-  CHECK_EQ(output1.getWidth(), output2.getWidth());
-
-  real* out = getData();
-  const real* x = output1.getData();
-  const real* y = output2.getData();
-  size_t yInc = dim;
-  if (output2.getHeight() == 1LU) {
-    yInc = 0;
-  } else {
-    CHECK_EQ(output2.getHeight(), numSamples);
-  }
-  for (size_t i = 0; i < numSamples; ++i, x += dim, y += yInc) {
-    real squareSumX = 0;
-    real squareSumY = 0;
-    real xy = 0;
-    for (size_t j = 0; j < dim; ++j) {
-      squareSumX += _square(x[j]);
-      squareSumY += _square(y[j]);
-      xy += x[j] * y[j];
-    }
-    CHECK(squareSumX > 0 && squareSumY > 0);
-    out[i] = scale * xy / (std::sqrt(squareSumX) * std::sqrt(squareSumY));
-  }
-}
-
-void CpuMatrix::cosSimDerivative(Matrix& output,
-                                 Matrix& prevOut1,
-                                 Matrix& prevOut2,
-                                 Matrix& prevGrad1,
-                                 Matrix& prevGrad2,
-                                 real scale) {
-  CHECK(output.useGpu_ == false) << "Matrix type are not equal";
-
-  CHECK_EQ(getWidth(), 1UL);
-  CHECK_EQ(output.getWidth(), 1UL);
-
-  size_t numSamples = getHeight();
-  CHECK_EQ(output.getHeight(), numSamples);
-  CHECK_EQ(prevOut1.getHeight(), numSamples);
-  CHECK_EQ(prevGrad1.getHeight(), numSamples);
-
-  size_t dim = prevOut1.getWidth();
-  CHECK_EQ(prevOut2.getWidth(), dim);
-  CHECK_EQ(prevGrad1.getWidth(), dim);
-  CHECK_EQ(prevGrad2.getWidth(), dim);
-
-  const real* grad = getData();
-  const real* out = output.getData();
-  const real* prevOutX = prevOut1.getData();
-  const real* prevOutY = prevOut2.getData();
-  real* prevGradX = prevGrad1.getData();
-  real* prevGradY = prevGrad2.getData();
-  size_t yInc = dim;
-  if (prevOut2.getHeight() == 1LU) {
-    yInc = 0;
-    CHECK_EQ(prevGrad2.getHeight(), 1LU);
-  } else {
-    CHECK_EQ(prevOut2.getHeight(), numSamples);
-    CHECK_EQ(prevGrad2.getHeight(), numSamples);
-  }
-  for (size_t i = 0; i < numSamples; ++i,
-              prevOutX += dim,
-              prevOutY += yInc,
-              prevGradX += dim,
-              prevGradY += yInc) {
-    real squareSumX = 0;
-    real squareSumY = 0;
-    real xy = 0;
-    for (size_t j = 0; j < dim; ++j) {
-      squareSumX += _square(prevOutX[j]);
-      squareSumY += _square(prevOutY[j]);
-      xy += prevOutX[j] * prevOutY[j];
-    }
-    CHECK(squareSumX > 0 && squareSumY > 0);
-    if (xy == 0) {
-      real reciprocal = 1.0f / (std::sqrt(squareSumX) * std::sqrt(squareSumY));
-      for (size_t j = 0; j < dim; ++j) {
-        prevGradX[j] += scale * grad[i] * prevOutY[j] * reciprocal;
-        prevGradY[j] += scale * grad[i] * prevOutX[j] * reciprocal;
-      }
-    } else {
-      real reciprocalXY = 1.0f / xy;
-      real reciprocalSquareSumX = 1.0f / squareSumX;
-      real reciprocalSquareSumY = 1.0f / squareSumY;
-      for (size_t j = 0; j < dim; ++j) {
-        prevGradX[j] += out[i] * grad[i] * (prevOutY[j] * reciprocalXY -
-                                            prevOutX[j] * reciprocalSquareSumX);
-        prevGradY[j] += out[i] * grad[i] * (prevOutX[j] * reciprocalXY -
-                                            prevOutY[j] * reciprocalSquareSumY);
-      }
-    }
-  }
-}
-
 void CpuMatrix::sumOfSquares(Matrix& output, Matrix& label) {
   CHECK(output.useGpu_ == false && label.useGpu_ == false)
       << "Matrix type are not equal";

paddle/math/Matrix.h

@@ -799,26 +799,6 @@ public:
     LOG(FATAL) << "Not implemented";
   }
 
-  /**
-   * cosine similarity, for each row i,
-   *   this[i] = cos(output1[i], output2[i])
-   *
-   * output2 can only have one row, then for each row i,
-   *   this[i] = cos(output1[i], output2[0])
-   */
-  virtual void cosSim(Matrix& output1, Matrix& output2, real scale = 1.0f) {
-    LOG(FATAL) << "Not implemented";
-  }
-
-  virtual void cosSimDerivative(Matrix& output,
-                                Matrix& prevOut1,
-                                Matrix& prevOut2,
-                                Matrix& prevGrad1,
-                                Matrix& prevGrad2,
-                                real scale = 1.0f) {
-    LOG(FATAL) << "Not implemented";
-  }
-
   /// print out the values of elements to os
   virtual void print(std::ostream& os) const {
     LOG(FATAL) << "Not implemented";
@@ -1324,14 +1304,6 @@ public:
   void softreluDerivative(Matrix& output);
   void scaledTanh(Matrix& output, real p1, real p2);
 
-  void cosSim(Matrix& output1, Matrix& output2, real scale);
-  void cosSimDerivative(Matrix& output,
-                        Matrix& prevOut1,
-                        Matrix& prevOut2,
-                        Matrix& prevGrad1,
-                        Matrix& prevGrad2,
-                        real scale);
-
   virtual void print(std::ostream& os) const;
   virtual void print(std::ostream& os, size_t height, size_t width) const;
 
@@ -1752,14 +1724,6 @@ public:
   void softreluDerivative(Matrix& output);
   void scaledTanh(Matrix& output, real p1, real p2);
 
-  void cosSim(Matrix& output1, Matrix& output2, real scale);
-  void cosSimDerivative(Matrix& output,
-                        Matrix& prevOut1,
-                        Matrix& prevOut2,
-                        Matrix& prevGrad1,
-                        Matrix& prevGrad2,
-                        real scale);
-
   void print(std::ostream& os) const;
   void print(std::ostream& os, size_t height, size_t width) const;
   void printOneRow(std::ostream& os, size_t idx) const;

paddle/math/tests/test_Matrix.cpp

@@ -181,28 +181,6 @@ TEST(Matrix, copyByRowIndex) {
   }
 }
 
-void testCosSim(int heightX, int heightY, int width, real scale) {
-  AutoCompare test(heightX, 1);
-  CpuMatrix arg1(heightX, width);
-  CpuMatrix arg2(heightY, width);
-  arg1.randomizeUniform();
-  arg2.randomizeUniform();
-  arg2.add(-0.5);
-  test.cmpWithArg(&Matrix::cosSim, arg1, arg2, scale);
-}
-
-TEST(Matrix, cosSim) {
-  for (auto heightX : {10, 100, 1000}) {
-    for (auto heightY : {1, heightX}) {
-      for (auto width : {10, 100, 1000}) {
-        for (auto scale : {1.0, 2.0}) {
-          testCosSim(heightX, heightY, width, scale);
-        }
-      }
-    }
-  }
-}
-
 void testParamReluForward(int height, int width, int w_height, int w_width) {
   AutoCompare test(height, width);
   CpuMatrix arg1(height, width);

paddle/math/tests/test_matrixCompare.cpp

@@ -720,61 +720,6 @@ TEST(Matrix, sequenceAvgForward) {
   }
 }
 
-void testCosSimDerivate(int heightX, int heightY, int width, real scale) {
-  MatrixPtr prevOutX = CpuMatrix::create(heightX, width, false, false);
-  MatrixPtr prevOutY = CpuMatrix::create(heightY, width, false, false);
-  MatrixPtr grad = CpuMatrix::create(heightX, 1, false, false);
-  MatrixPtr output = CpuMatrix::create(heightX, 1, false, false);
-  MatrixPtr prevGradX = CpuMatrix::create(heightX, width, false, false);
-  MatrixPtr prevGradY = CpuMatrix::create(heightY, width, false, false);
-
-  prevOutX->randomizeUniform();
-  prevOutY->randomizeUniform();
-  grad->randomizeUniform();
-  output->randomizeUniform();
-  prevGradX->randomizeUniform();
-  prevGradY->randomizeUniform();
-
-  MatrixPtr prevOutXGpu = GpuMatrix::create(heightX, width, false, true);
-  MatrixPtr prevOutYGpu = GpuMatrix::create(heightY, width, false, true);
-  MatrixPtr gradGpu = GpuMatrix::create(heightX, 1, false, true);
-  MatrixPtr outputGpu = GpuMatrix::create(heightX, 1, false, true);
-  MatrixPtr prevGradXGpu = GpuMatrix::create(heightX, width, false, true);
-  MatrixPtr prevGradYGpu = GpuMatrix::create(heightY, width, false, true);
-
-  prevOutXGpu->copyFrom(*prevOutX);
-  prevOutYGpu->copyFrom(*prevOutY);
-  gradGpu->copyFrom(*grad);
-  outputGpu->copyFrom(*output);
-  prevGradXGpu->copyFrom(*prevGradX);
-  prevGradYGpu->copyFrom(*prevGradY);
-
-  grad->cosSimDerivative(
-      *output, *prevOutX, *prevOutY, *prevGradX, *prevGradY, scale);
-
-  gradGpu->cosSimDerivative(*outputGpu,
-                            *prevOutXGpu,
-                            *prevOutYGpu,
-                            *prevGradXGpu,
-                            *prevGradYGpu,
-                            scale);
-
-  TensorCheckErr(*prevGradX, *prevGradXGpu);
-  TensorCheckErr(*prevGradY, *prevGradYGpu);
-}
-
-TEST(Matrix, cosSimDerivate) {
-  for (auto heightX : {1, 10, 100}) {
-    for (auto heightY : {1, heightX}) {
-      for (auto width : {1, 10, 100}) {
-        for (auto scale : {1.0, 2.0}) {
-          testCosSimDerivate(heightX, heightY, width, scale);
-        }
-      }
-    }
-  }
-}
-
 void testParamReluBackwardDiff(int height,
                                int width,
                                int w_height,

paddle/utils/Util.cpp

@@ -289,6 +289,7 @@ void mkDir(const char* filename) {
 void mkDirRecursively(const char* dir) {
   struct stat sb;
 
+  if (*dir == 0) return;  // empty string
   if (!stat(dir, &sb)) return;
 
   mkDirRecursively(path::dirname(dir).c_str());

python/paddle/trainer/config_parser.py

@@ -893,11 +893,11 @@ class MaxOut(Cfg):
         self.add_keys(locals())
 
 
-def DataBase(async_load_data=False,
-             constant_slots=None,
-             data_ratio=1,
-             is_main_data=True,
-             usage_ratio=None):
+def create_data_config_proto(async_load_data=False,
+                             constant_slots=None,
+                             data_ratio=1,
+                             is_main_data=True,
+                             usage_ratio=None):
     # default: all sub dataproviders are treat as "main data".
     # see proto/DataConfig.proto for is_main_data
     data_config = DataConfig()
@@ -923,7 +923,7 @@ def SimpleData(files=None,
                context_len=None,
                buffer_capacity=None,
                **xargs):
-    data_config = DataBase(**xargs)
+    data_config = create_data_config_proto(**xargs)
     data_config.type = 'simple'
     data_config.files = files
     data_config.feat_dim = feat_dim
@@ -945,7 +945,7 @@ def PyData(files=None,
            constant_slots=None,
            load_thread_num=None,
            **xargs):
-    data_config = DataBase(**xargs)
+    data_config = create_data_config_proto(**xargs)
     data_config.type = 'py'
     if load_data_module in g_py_module_name_list:
 
@@ -996,7 +996,7 @@ def ProtoData(files=None,
               constant_slots=None,
               load_thread_num=None,
               **xargs):
-    data_config = DataBase(**xargs)
+    data_config = create_data_config_proto(**xargs)
     if type is None:
         data_config.type = 'proto'
     else:
@@ -1035,7 +1035,7 @@ def Data(type,
          buffer_capacity=None,
          **xargs):
 
-    data_config = DataBase(**xargs)
+    data_config = create_data_config_proto(**xargs)
     data_config.type = type
     data_config.files = files
     data_config.feat_dim = feat_dim

python/paddle/trainer_config_helpers/data_sources.py

@@ -58,8 +58,8 @@ def define_py_data_source(file_list,
     :param obj: python object name. May be a function name if using
                 PyDataProviderWrapper.
     :type obj: basestring
-    :param args: The best practice is using dict to pass arguments into 
-                 DataProvider, and use :code:`@init_hook_wrapper` to 
+    :param args: The best practice is using dict to pass arguments into
+                 DataProvider, and use :code:`@init_hook_wrapper` to
                  receive arguments.
     :type args: string or picklable object
     :param async: Load Data asynchronously or not.
@@ -98,7 +98,7 @@ def define_py_data_sources(train_list,
     The annotation is almost the same as define_py_data_sources2, except that
     it can specific train_async and data_cls.
 
-    :param data_cls: 
+    :param data_cls:
     :param train_list: Train list name.
     :type train_list: basestring
     :param test_list: Test list name.
@@ -111,8 +111,8 @@ def define_py_data_sources(train_list,
                 a tuple or list to this argument.
     :type obj: basestring or tuple or list
     :param args: The best practice is using dict() to pass arguments into
-                 DataProvider, and use :code:`@init_hook_wrapper` to receive 
-                 arguments. If train and test is different, then pass a tuple 
+                 DataProvider, and use :code:`@init_hook_wrapper` to receive
+                 arguments. If train and test is different, then pass a tuple
                  or list to this argument.
     :type args: string or picklable object or list or tuple.
     :param train_async: Is training data load asynchronously or not.
@@ -163,12 +163,12 @@ def define_py_data_sources2(train_list, test_list, module, obj, args=None):
 
     ..  code-block:: python
 
-        define_py_data_sources2(train_list="train.list", 
-                                test_list="test.list", 
+        define_py_data_sources2(train_list="train.list",
+                                test_list="test.list",
                                 module="data_provider"
                                 # if train/test use different configurations,
                                 # obj=["process_train", "process_test"]
-                                obj="process", 
+                                obj="process",
                                 args={"dictionary": dict_name})
 
     The related data provider can refer to :ref:`api_pydataprovider2_sequential_model` .
@@ -185,8 +185,8 @@ def define_py_data_sources2(train_list, test_list, module, obj, args=None):
                 a tuple or list to this argument.
     :type obj: basestring or tuple or list
     :param args: The best practice is using dict() to pass arguments into
-                 DataProvider, and use :code:`@init_hook_wrapper` to receive 
-                 arguments. If train and test is different, then pass a tuple 
+                 DataProvider, and use :code:`@init_hook_wrapper` to receive
+                 arguments. If train and test is different, then pass a tuple
                  or list to this argument.
     :type args: string or picklable object or list or tuple.
     :return: None
@@ -195,13 +195,13 @@ def define_py_data_sources2(train_list, test_list, module, obj, args=None):
 
     def py_data2(files, load_data_module, load_data_object, load_data_args,
                  **kwargs):
-        data = DataBase()
+        data = create_data_config_proto()
         data.type = 'py2'
         data.files = files
         data.load_data_module = load_data_module
         data.load_data_object = load_data_object
         data.load_data_args = load_data_args
-        data.async_load_data = True
+        data.async_load_data = False
         return data
 
     define_py_data_sources(

python/paddle/trainer_config_helpers/layers.py

@@ -3677,26 +3677,27 @@ def pad_layer(input,
 
     For example,
 
-    .. code-block::
-
-      input(2,2,2,3)  = [
-                          [ [[1,2,3], [3,4,5]],
-                            [[2,3,5], [1,6,7]] ],
-                          [ [[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]],
-                            [[1,2,3], [3,4,5]],
-                            [[2,3,5], [1,6,7]],
-                            [[0,0,0], [0,0,0]] ],
-                          [ [[0,0,0], [0,0,0]],
-                            [[4,3,1], [1,8,7]],
-                            [[3,8,9], [2,3,5]],
-                            [[0,0,0], [0,0,0]] ]
-                        ]
+    .. code-block:: python
+
+       input(2,2,2,3)  = [
+                           [ [[1,2,3], [3,4,5]],
+                             [[2,3,5], [1,6,7]] ],
+                           [ [[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]],
+                             [[1,2,3], [3,4,5]],
+                             [[2,3,5], [1,6,7]],
+                             [[0,0,0], [0,0,0]] ],
+                           [ [[0,0,0], [0,0,0]],
+                             [[4,3,1], [1,8,7]],
+                             [[3,8,9], [2,3,5]],
+                             [[0,0,0], [0,0,0]] ]
+                         ]
 
     The simply usage is:
 
@@ -4191,13 +4192,7 @@ def block_expand_layer(input,
 
 @wrap_name_default()
 @layer_support()
-def maxout_layer(input,
-                 groups,
-                 num_channels=None,
-                 size_x=None,
-                 size_y=None,
-                 name=None,
-                 layer_attr=None):
+def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
     """
     A layer to do max out on conv layer output.
       - Input: output of a conv layer.
@@ -4227,12 +4222,6 @@ def maxout_layer(input,
     :type num_channels: int|None
     :param groups: The group number of input layer.
     :type groups: int
-    :param size_x: conv output width. If None will be set
-                   automatically from previous output.
-    :type size_x: int|None
-    :param size_y: conv output height. If None will be set
-                   automatically from previous output.
-    :type size_y: int|None
     :param name: The name of this layer, which can not specify.
     :type name: None|basestring.
     :param layer_attr: Extra Layer attribute.

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

@@ -19,7 +19,7 @@ model_config {
 data_config {
   type: "py2"
   files: "train.list"
-  async_load_data: true
+  async_load_data: false
   for_test: false
   load_data_module: "a"
   load_data_object: "c"
@@ -58,7 +58,7 @@ opt_config {
 test_data_config {
   type: "py2"
   files: "test.list"
-  async_load_data: true
+  async_load_data: false
   for_test: true
   load_data_module: "b"
   load_data_object: "d"