Merge branch 'develop' into concat

cmake/external/mkldnn.cmake

@@ -54,7 +54,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "a29d8487a63afca3d5b8c5bbdbb473cf8ccc6e51"
+    GIT_TAG             "64e03a1939e0d526aa8e9f2e3f7dc0ad8d372944"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}

paddle/fluid/framework/tensor.cc

@@ -31,7 +31,8 @@ size_t Tensor::memory_size() const {
   return holder_ == nullptr ? 0UL : holder_->size() - offset_;
 }
 
-void* Tensor::mutable_data(platform::Place place, std::type_index type) {
+void* Tensor::mutable_data(platform::Place place, std::type_index type,
+                           size_t requested_size) {
   if (holder_ != nullptr) {
     holder_->set_type(type);
   }
@@ -39,7 +40,7 @@ void* Tensor::mutable_data(platform::Place place, std::type_index type) {
                     "When calling this method, the Tensor's numel must be "
                     "equal or larger than zero. "
                     "Please check Tensor::Resize has been called first.");
-  int64_t size = numel() * SizeOfType(type);
+  size_t size = requested_size ? requested_size : numel() * SizeOfType(type);
   /* some versions of boost::variant don't have operator!= */
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
@@ -68,10 +69,10 @@ void* Tensor::mutable_data(platform::Place place, std::type_index type) {
                                  offset_);
 }
 
-void* Tensor::mutable_data(platform::Place place) {
+void* Tensor::mutable_data(platform::Place place, size_t requested_size) {
   PADDLE_ENFORCE(this->holder_ != nullptr,
                  "Cannot invoke mutable data if current hold nothing.");
-  return mutable_data(place, holder_->type());
+  return mutable_data(place, holder_->type(), requested_size);
 }
 
 Tensor& Tensor::ShareDataWith(const Tensor& src) {

paddle/fluid/framework/tensor.h

@@ -89,22 +89,24 @@ class Tensor {
    * @note    If not exist, then allocation.
    */
   template <typename T>
-  T* mutable_data(platform::Place place);
+  T* mutable_data(platform::Place place, size_t requested_size = 0);
 
-  void* mutable_data(platform::Place place, std::type_index type);
+  void* mutable_data(platform::Place place, std::type_index type,
+                     size_t requested_size = 0);
 
-  void* mutable_data(platform::Place place);
+  void* mutable_data(platform::Place place, size_t requested_size = 0);
 
   /**
    * @brief     Return a pointer to mutable memory block.
    *
    * @param[in] dims           The dimensions of the memory block.
    * @param[in] place          The place of the memory block.
+   * @param[in] requested_size The size of the block in bytes.
    *
    * @note      If not exist, then allocation.
    */
   template <typename T>
-  T* mutable_data(DDim dims, platform::Place place);
+  T* mutable_data(DDim dims, platform::Place place, size_t requested_size = 0);
 
   /*! Return the dimensions of the memory block. */
   const DDim& dims() const;

paddle/fluid/framework/tensor_impl.h

@@ -46,16 +46,17 @@ inline T* Tensor::data() {
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
+inline T* Tensor::mutable_data(DDim dims, platform::Place place,
+                               size_t requested_size) {
   static_assert(std::is_pod<T>::value, "T must be POD");
   Resize(dims);
-  return mutable_data<T>(place);
+  return mutable_data<T>(place, requested_size);
 }
 
 template <typename T>
-inline T* Tensor::mutable_data(platform::Place place) {
+inline T* Tensor::mutable_data(platform::Place place, size_t requested_size) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  return reinterpret_cast<T*>(mutable_data(place, typeid(T)));
+  return reinterpret_cast<T*>(mutable_data(place, typeid(T), requested_size));
 }
 
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {

paddle/fluid/operators/conv_mkldnn_op.cc

@@ -53,6 +53,18 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler {
     key_ += "-BWD";
   }
 
+  size_t GetDstMemorySize() const {
+    return conv_pd_->dst_primitive_desc().get_size();
+  }
+
+  size_t GetDiffWeightsMemorySize() const {
+    return conv_bwd_weights_pd_->diff_weights_primitive_desc().get_size();
+  }
+
+  size_t GetDiffSourceMemorySize() const {
+    return conv_bwd_data_pd_->diff_src_primitive_desc().get_size();
+  }
+
   std::shared_ptr<mkldnn::memory> AcquireSrcMemoryFromWeightsPrimitive(
       const std::shared_ptr<mkldnn::memory> user_memory_p,
       std::vector<mkldnn::primitive>& pipeline) {  // NOLINT
@@ -294,7 +306,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 
     const T* input_data = input->data<T>();
     const T* filter_data = filter->data<T>();
-    T* output_data = output->mutable_data<T>(ctx.GetPlace());
 
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> weights_tz =
@@ -354,6 +365,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto user_weights_memory_p = handler.AcquireWeightsMemory(
         user_weights_md, to_void_cast<T>(filter_data));
 
+    T* output_data =
+        output->mutable_data<T>(ctx.GetPlace(), handler.GetDstMemorySize());
     // create reorder primitive if the input format is not the preferred one
     auto src_memory_p =
         handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline);
@@ -476,13 +489,6 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     T* input_grad_data = nullptr;
     T* filter_grad_data = nullptr;
 
-    if (input_grad) {
-      input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
-    }
-    if (filter_grad) {
-      filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace());
-    }
-
     std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
     std::vector<int> weights_tz =
         paddle::framework::vectorize2int(filter->dims());
@@ -568,6 +574,9 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
           handler.AcquireDiffDstMemoryFromWeightsPrimitive(
               user_diff_dst_memory_p, pipeline);
 
+      const size_t size = handler.GetDiffWeightsMemorySize();
+      filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace(), size);
+
       auto diff_weights_memory_p =
           handler.AcquireDiffWeightsMemoryFromWeightsPrimitive(
               reinterpret_cast<void*>(filter_grad_data));
@@ -590,6 +599,9 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
           handler.AcquireDiffDstMemoryFromDataPrimitive(user_diff_dst_memory_p,
                                                         pipeline);
 
+      const size_t size = handler.GetDiffSourceMemorySize();
+      input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace(), size);
+
       auto diff_src_memory_p = handler.AcquireDiffSrcMemoryFromDataPrimitive(
           reinterpret_cast<void*>(input_grad_data));
 

python/paddle/dataset/common.py

@@ -19,6 +19,7 @@ import hashlib
 import os
 import errno
 import shutil
+import six
 import sys
 import importlib
 import paddle.dataset
@@ -94,6 +95,8 @@ def download(url, module_name, md5sum, save_name=None):
                 dl = 0
                 total_length = int(total_length)
                 for data in r.iter_content(chunk_size=4096):
+                    if six.PY2:
+                        data = six.b(data)
                     dl += len(data)
                     f.write(data)
                     done = int(50 * dl / total_length)

python/paddle/dataset/flowers.py

@@ -35,6 +35,7 @@ import itertools
 import functools
 from .common import download
 import tarfile
+import six
 import scipy.io as scio
 from paddle.dataset.image import *
 from paddle.reader import *
@@ -45,10 +46,10 @@ from six.moves import cPickle as pickle
 from six.moves import zip
 __all__ = ['train', 'test', 'valid']
 
-DATA_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/102flowers.tgz'
-LABEL_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/imagelabels.mat'
-SETID_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/setid.mat'
-DATA_MD5 = '33bfc11892f1e405ca193ae9a9f2a118'
+DATA_URL = 'http://paddlemodels.cdn.bcebos.com/flowers/102flowers.tgz'
+LABEL_URL = 'http://paddlemodels.cdn.bcebos.com/flowers/imagelabels.mat'
+SETID_URL = 'http://paddlemodels.cdn.bcebos.com/flowers/setid.mat'
+DATA_MD5 = '52808999861908f626f3c1f4e79d11fa'
 LABEL_MD5 = 'e0620be6f572b9609742df49c70aed4d'
 SETID_MD5 = 'a5357ecc9cb78c4bef273ce3793fc85c'
 # In official 'readme', tstid is the flag of test data
@@ -120,7 +121,10 @@ def reader_creator(data_file,
                 file = file.strip()
                 batch = None
                 with open(file, 'rb') as f:
+                    if six.PY2:
                         batch = pickle.load(f)
+                    else:
+                        batch = pickle.load(f, encoding='bytes')
                 data = batch['data']
                 labels = batch['label']
                 for sample, label in zip(data, batch['label']):