update sequence_slice_op.h, change some code

Author: wanghaox

.gitignore

@@ -28,4 +28,3 @@ cmake_install.cmake
 paddle/.timestamp
 python/paddlepaddle.egg-info/
 paddle/pybind/pybind.h
-python/paddle/v2/framework/tests/tmp/*

cmake/external/openblas.cmake

@@ -98,7 +98,7 @@ IF(NOT ${CBLAS_FOUND})
         ENDIF()
         INSTALL(CODE "execute_process(
             COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
-                    destination ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
+                    ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
             )"
         )
         INSTALL(CODE "MESSAGE(STATUS \"Installing: \"

doc/design/evaluator.md

@@ -0,0 +1,58 @@
+## Evaluator Design
+
+### The Problem
+
+During training or serving, we provide the evaluation function to measure the model performance, e.g., accuracy, precision. In the operator based framework design, the data go through the network pipeline batch by batch. As a result, inside the operator, we only can calculate one minibatch metrics. We need to provide a mechanism to calculate the metrics for each N pass/batch the user wanted.
+
+### Evaluator Design
+Currently, every operation is expressed in the graph. we divide the evaluator process into three steps.
+
+1. Initialize the metric state and add it into the block.
+
+2. Calculate the statistic of the metric state in every mini-batch. The single operator is only responsible for calculating necessary statistics for one mini-batch. For example, accuracy operator only calculate a minibatch data if run once.
+
+
+3. Merge the mini-batch statistics to form the evaluation result for multiple mini-batches. When it comes to distributed training/Multi-GPU training, aggregate the value from different devices.
+
+### Implementation
+This design is shown in python API. 
+Each metric operator need to caculate the metric statistic and return the batch aware states, Python side responsible for accumulate the states for each pass. 
+
+    
+```python
+class Evaluator(object):
+    """
+    Evaluator Base class.
+    """
+    def __init__(self, name, **kwargs):
+       """
+       Different evaluator may has different metric states. E.g, Accuracy need two variables, total and right sample counts.
+       Auc need four variables, `true_positives`,
+         `true_negatives`, `false_positives` and `false_negatives`. So every evaluator should create its needed variables and append to main_program
+
+       The initialization of Evaluator should be responsible for:
+       create metric states and append to the main_program
+       """ 
+       pass
+
+    def _update_ops(self, input, label, **kwargs)
+       """
+       Add mini-batch evaluator caculate operators to the main_program.
+       Add increment operator to accumulate the metric states.
+       """
+    
+
+    def reset(self, executor, reset_program=None):
+      """
+      Reset metric states at the begin of each pass/user specified batch number.
+      Execute the reset_program to reset the states.
+      """
+      
+
+    def eval(self, executor, eval_program=None):
+      """
+      Merge the mini-batch statistics to form the evaluation result for multiple mini-batches.
+      Execute the eval_program and return the result.
+      """
+      return eval_result
+```

paddle/capi/Matrix.cpp

@@ -121,6 +121,7 @@ paddle_error paddle_matrix_get_shape(paddle_matrix mat,
 
 paddle_matrix paddle_matrix_create_sparse(
     uint64_t height, uint64_t width, uint64_t nnz, bool isBinary, bool useGpu) {
+#ifndef PADDLE_MOBILE_INFERENCE
   auto ptr = new paddle::capi::CMatrix();
   ptr->mat = paddle::Matrix::createSparseMatrix(
       height,
@@ -131,6 +132,9 @@ paddle_matrix paddle_matrix_create_sparse(
       false,
       useGpu);
   return ptr;
+#else
+  return nullptr;
+#endif
 }
 
 paddle_error paddle_matrix_sparse_copy_from(paddle_matrix mat,
@@ -140,6 +144,7 @@ paddle_error paddle_matrix_sparse_copy_from(paddle_matrix mat,
                                             uint64_t colSize,
                                             float* valueArray,
                                             uint64_t valueSize) {
+#ifndef PADDLE_MOBILE_INFERENCE
   if (mat == nullptr) return kPD_NULLPTR;
   auto ptr = cast(mat);
   if (rowArray == nullptr || colArray == nullptr ||
@@ -160,4 +165,7 @@ paddle_error paddle_matrix_sparse_copy_from(paddle_matrix mat,
   } else {
     return kPD_NOT_SUPPORTED;
   }
+#else
+  return kPD_NOT_SUPPORTED;
+#endif
 }

paddle/capi/matrix.h

@@ -48,6 +48,7 @@ PD_API paddle_matrix paddle_matrix_create(uint64_t height,
  * @param isBinary is binary (either 1 or 0 in matrix) or not.
  * @param useGpu is using GPU or not.
  * @return paddle_matrix.
+ * @note Mobile inference does not support this interface.
  */
 PD_API paddle_matrix paddle_matrix_create_sparse(
     uint64_t height, uint64_t width, uint64_t nnz, bool isBinary, bool useGpu);
@@ -129,6 +130,7 @@ PD_API paddle_error paddle_matrix_get_shape(paddle_matrix mat,
  * NULL if the matrix is binary.
  * @param [in] valueSize length of value array. Zero if the matrix is binary.
  * @return paddle_error
+ * @note Mobile inference does not support this interface.
  */
 PD_API paddle_error paddle_matrix_sparse_copy_from(paddle_matrix mat,
                                                    int* rowArray,

paddle/cuda/CMakeLists.txt

@@ -27,7 +27,9 @@ if(WITH_GPU)
     set_source_files_properties(${CUDA_CXX_SOURCES}
                                 PROPERTIES COMPILE_FLAGS "-D__NVCC__")
 else()
+    if (NOT MOBILE_INFERENCE)
     set(CUDA_CXX_SOURCES src/hl_warpctc_wrap.cc)
+    endif()
 endif()
 
 set(CUDA_CU_SOURCES

paddle/cuda/include/hl_cnn.h

@@ -18,7 +18,7 @@ limitations under the License. */
 #include "hl_base.h"
 
 /**
- * @brief   Maximum pool forward.
+ * @brief   Maximum pool forward with Mask output.
  *
  * @param[in]   frameCnt    batch size of input image.
  * @param[in]   inputData   input data.
@@ -35,7 +35,7 @@ limitations under the License. */
  * @param[in]   paddingW    padding width.
  * @param[out]  tgtData     output data.
  * @param[in]   tgtStride   stride between output data samples.
- *
+ * @param[out]  maskData    the location indices of select max data.
  */
 extern void hl_maxpool_forward(const int frameCnt,
                                const real* inputData,
@@ -51,7 +51,8 @@ extern void hl_maxpool_forward(const int frameCnt,
                                const int paddingH,
                                const int paddingW,
                                real* tgtData,
-                               const int tgtStride);
+                               const int tgtStride,
+                               real* maskData = NULL);
 
 /**
  * @brief   Maximum pool backward.

paddle/cuda/include/stub/hl_cnn_stub.h

@@ -31,7 +31,8 @@ inline void hl_maxpool_forward(const int frameCnt,
                                const int paddingH,
                                const int paddingW,
                                real* tgtData,
-                               const int tgtStride) {}
+                               const int tgtStride,
+                               real* MaskData) {}
 
 inline void hl_maxpool_backward(const int frameCnt,
                                 const real* inputData,

paddle/cuda/src/hl_cuda_cnn.cu

@@ -31,7 +31,8 @@ __global__ void KeMaxPoolForward(const int nthreads,
                                  const int offsetH,
                                  const int offsetW,
                                  real* tgtData,
-                                 const int tgtStride) {
+                                 const int tgtStride,
+                                 real* maskData) {
   int index = blockIdx.x * blockDim.x + threadIdx.x;
   if (index < nthreads) {
     int pw = index % pooledW;
@@ -45,16 +46,22 @@ __global__ void KeMaxPoolForward(const int nthreads,
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
     real maxval = -FLT_MAX;
+    int max_index = -1;
     inputData += (frameNum * channels + c) * height * width;
     for (int h = hstart; h < hend; ++h) {
       for (int w = wstart; w < wend; ++w) {
-        if (maxval < inputData[h * width + w])
-          maxval = inputData[h * width + w];
+        if (maxval < inputData[h * width + w]) {
+          max_index = h * width + w;
+          maxval = inputData[max_index];
+        }
       }
     }
     int tgtIndex =
         index % (pooledW * pooledH * channels) + frameNum * tgtStride;
     tgtData[tgtIndex] = maxval;
+    if (maskData != NULL) {
+      maskData[tgtIndex] = max_index;
+    }
   }
 }
 
@@ -72,7 +79,8 @@ void hl_maxpool_forward(const int frameCnt,
                         const int paddingH,
                         const int paddingW,
                         real* tgtData,
-                        const int tgtStride) {
+                        const int tgtStride,
+                        real* maskData) {
   int num_kernels = pooledH * pooledW * channels * frameCnt;
   int blocks = (num_kernels + 1024 - 1) / 1024;
   dim3 threads(1024, 1);
@@ -92,7 +100,8 @@ void hl_maxpool_forward(const int frameCnt,
                                                          paddingH,
                                                          paddingW,
                                                          tgtData,
-                                                         tgtStride);
+                                                         tgtStride,
+                                                         maskData);
   CHECK_SYNC("hl_maxpool_forward failed");
 }
 

paddle/framework/CMakeLists.txt

@@ -38,9 +38,9 @@ py_proto_compile(framework_py_proto SRCS framework.proto)
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
 add_custom_command(TARGET framework_py_proto POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto
-    COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto/
-    COMMENT "Copy generated python proto into directory paddle/v2/framework/proto."
+    COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/fluid/proto
+    COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/fluid/proto/
+    COMMENT "Copy generated python proto into directory paddle/v2/fluid/proto."
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 
 cc_library(backward SRCS backward.cc DEPS net_op)