Add Python array reader op

benchmark/fluid/args.py

@@ -122,5 +122,15 @@ def parse_args():
         type=str,
         default="",
         help='Directory that contains all the training recordio files.')
+    parser.add_argument(
+        '--use_py_reader_op',
+        action='store_true',
+        help='Whether to use Python reader op, omitted when use_reader_op is true'
+    )
+    parser.add_argument(
+        '--feed_queue_capacity',
+        type=int,
+        default=64,
+        help='Capacity of feed queue when use_py_reader_op is true')
     args = parser.parse_args()
     return args

benchmark/fluid/fluid_benchmark.py

@@ -25,6 +25,9 @@ import paddle.fluid.profiler as profiler
 import paddle.fluid.transpiler.distribute_transpiler as distribute_transpiler
 
 from args import *
+import threading
+
+feed_queue = None
 
 
 def append_nccl2_prepare(trainer_id):
@@ -131,7 +134,7 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
     exe = fluid.Executor(place)
     exe.run(startup_prog)
 
-    if not args.use_reader_op:
+    if not args.use_reader_op and not args.use_py_reader_op:
         feed_var_list = [
             var for var in train_prog.global_block().vars.itervalues()
             if var.is_data
@@ -141,12 +144,12 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
     iters, num_samples, start_time = 0, 0, time.time()
     for pass_id in range(args.pass_num):
         train_losses = []
-        if not args.use_reader_op:
+        if not args.use_reader_op and not args.use_py_reader_op:
             reader_generator = train_reader()
         batch_id = 0
         data = None
         while True:
-            if not args.use_reader_op:
+            if not args.use_reader_op and not args.use_py_reader_op:
                 data = next(reader_generator, None)
                 if data == None:
                     break
@@ -156,7 +159,7 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
                 start_time = time.time()
                 num_samples = 0
 
-            if args.use_reader_op:
+            if args.use_reader_op or args.use_py_reader_op:
                 try:
                     loss = exe.run(train_prog, fetch_list=[avg_loss])
                 except fluid.core.EnforceNotMet as ex:
@@ -170,7 +173,7 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
             # FIXME(wuyi): For use_reader_op, if the current
             # pass is not the last, the last batch of this pass
             # is also equal to args.batch_size.
-            if args.use_reader_op:
+            if args.use_reader_op or args.use_py_reader_op:
                 num_samples += args.batch_size * args.gpus
             else:
                 num_samples += len(data)
@@ -180,12 +183,13 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
         print_train_time(start_time, time.time(), num_samples)
         print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))),
         # evaluation
-        if not args.no_test and batch_acc and not args.use_reader_op:
+        if not args.no_test and batch_acc and not args.use_reader_op and not args.use_py_reader_op:
             pass_test_acc = test(exe, infer_prog, test_reader, feeder,
                                  batch_acc)
             print(", Test Accuracy: %f" % pass_test_acc)
         print("\n")
         # TODO(wuyi): add warmup passes to get better perf data.
+        close_feed_queue()
         exit(0)
 
 
@@ -195,7 +199,7 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
                    batch_acc, args, train_prog, startup_prog, nccl_id_var,
                    num_trainers, trainer_id):
     place = core.CPUPlace() if args.device == 'CPU' else core.CUDAPlace(0)
-    if not args.use_reader_op:
+    if not args.use_reader_op and not args.use_py_reader_op:
         feed_var_list = [
             var for var in train_prog.global_block().vars.itervalues()
             if var.is_data
@@ -238,12 +242,12 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
         num_samples = 0
         iters = 0
         start_time = time.time()
-        if not args.use_reader_op:
+        if not args.use_reader_op and not args.use_py_reader_op:
             reader_generator = train_reader()
         batch_id = 0
         data = None
         while True:
-            if not args.use_reader_op:
+            if not args.use_reader_op and not args.use_py_reader_op:
                 data = next(reader_generator, None)
                 if data == None:
                     break
@@ -257,14 +261,14 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
             if iters == args.skip_batch_num:
                 start_time = time.time()
                 num_samples = 0
-            if args.use_fake_data or args.use_reader_op:
+            if args.use_fake_data or args.use_reader_op or args.use_py_reader_op:
                 try:
                     loss, = exe.run([avg_loss.name])
                 except fluid.core.EnforceNotMet as ex:
                     break
             else:
                 loss, = exe.run([avg_loss.name], feed=feeder.feed(data))
-            if args.use_reader_op:
+            if args.use_reader_op or args.use_py_reader_op:
                 num_samples += args.batch_size * args.gpus
             else:
                 num_samples += len(data)
@@ -275,7 +279,7 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
             batch_id += 1
 
         print_train_time(start_time, time.time(), num_samples)
-        if not args.no_test and batch_acc and not args.use_reader_op:
+        if not args.no_test and batch_acc and not args.use_reader_op and not args.use_py_reader_op:
             # we have not implement record io for test
             # skip test when use args.use_reader_op
             test_acc = test(startup_exe, infer_prog, test_reader, feeder,
@@ -307,7 +311,46 @@ def print_paddle_envs():
     print('------------------------------------------------')
 
 
+def feed_data(feed_queue, train_reader, test_reader, dshapes, args):
+    train_cnt = 0
+    test_cnt = 0
+    print_per_train_batch = 1
+    train_data_generator = train_reader()
+    start = time.time()
+    while True:
+        next_data = next(train_data_generator, None)
+        if next_data is None:
+            break
+
+        next_data = list(next_data)
+        for i in range(len(next_data)):
+            if not isinstance(next_data[i], np.ndarray):
+                next_data[i] = np.array(next_data[i])
+            next_data[i] = next_data[i].reshape([-1] + dshapes[i])
+
+        if not feed_queue.enqueue(next_data):
+            break
+
+        train_cnt += 1
+        '''
+        if train_cnt % print_per_train_batch == 0:
+            end = time.time()
+            print('Feed queue size: %d, capacity: %d, speed: %.5fsec/batch'
+                % (feed_queue.size(), feed_queue.capacity(), (end-start)/print_per_train_batch))
+            start = end
+        '''
+    feed_queue.close()
+
+
+def close_feed_queue():
+    global feed_queue
+    if feed_queue is not None:
+        feed_queue.close()
+
+
 def main():
+    global feed_queue
+
     args = parse_args()
     print_arguments(args)
     print_paddle_envs()
@@ -321,8 +364,23 @@ def main():
         pr = cProfile.Profile()
         pr.enable()
     model_def = __import__("models.%s" % args.model, fromlist=["models"])
-    train_args = list(model_def.get_model(args))
+    model = model_def.get_model(args)
+
+    if not args.use_reader_op and args.use_py_reader_op:
+        feed_queue = model[-4]
+        train_reader = model[-3]
+        test_reader = model[-2]
+        dshapes = model[-1]
+        feed_thread = threading.Thread(
+            target=feed_data,
+            args=(feed_queue, train_reader, test_reader, dshapes, args))
+        #feed_thread.setDaemon(True)
+        feed_thread.start()
+        model = model[:-4]
+
+    train_args = list(model)
     train_args.append(args)
+
     # Run optimizer.minimize(avg_loss)
     train_args[2].minimize(train_args[0])
     if args.memory_optimize:
@@ -338,6 +396,7 @@ def main():
             train_args.extend([nccl_id_var, num_trainers, trainer_id])
             train_parallel(*train_args)
         train(*train_args)
+        close_feed_queue()
         exit(0)
 
     # for other update methods, use default programs
@@ -362,3 +421,4 @@ def main():
 
 if __name__ == "__main__":
     main()
+    close_feed_queue()

benchmark/fluid/models/machine_translation.py

@@ -182,7 +182,7 @@ def lodtensor_to_ndarray(lod_tensor):
 
 
 def get_model(args):
-    if args.use_reader_op:
+    if args.use_reader_op or args.use_py_reader_op:
         raise Exception("machine_translation do not support reader op for now.")
     embedding_dim = 512
     encoder_size = 512

benchmark/fluid/models/mnist.py

@@ -66,13 +66,14 @@ def cnn_model(data):
 
 
 def get_model(args):
+    dshape = [1, 28, 28]
     if args.use_reader_op:
         filelist = [
             os.path.join(args.data_path, f) for f in os.listdir(args.data_path)
         ]
         data_file = fluid.layers.open_files(
             filenames=filelist,
-            shapes=[[-1, 1, 28, 28], (-1, 1)],
+            shapes=[[-1] + dshape, (-1, 1)],
             lod_levels=[0, 0],
             dtypes=["float32", "int64"],
             thread_num=args.gpus,
@@ -81,8 +82,18 @@ def get_model(args):
             fluid.layers.batch(
                 data_file, batch_size=args.batch_size))
         images, label = fluid.layers.read_file(data_file)
+    elif args.use_py_reader_op:
+        data_file, feed_queue = fluid.layers.py_array_reader(
+            capacity=args.feed_queue_capacity,
+            shapes=[[-1] + dshape, [-1, 1]],
+            lod_levels=[0, 0],
+            dtypes=['float32', 'int64'])
+        data_file = fluid.layers.double_buffer(
+            fluid.layers.batch(
+                data_file, batch_size=args.batch_size))
+        images, label = fluid.layers.read_file(data_file)
     else:
-        images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
+        images = fluid.layers.data(name='pixel', shape=dshape, dtype=DTYPE)
         label = fluid.layers.data(name='label', shape=[1], dtype='int64')
 
     if args.device == 'CPU' and args.cpus > 1:
@@ -118,8 +129,16 @@ def get_model(args):
         learning_rate=0.001, beta1=0.9, beta2=0.999)
 
     # Reader
+    underlying_train_reader = paddle.dataset.mnist.train()
+    underlying_test_reader = paddle.dataset.mnist.test()
     train_reader = paddle.batch(
-        paddle.dataset.mnist.train(), batch_size=args.batch_size * args.gpus)
+        underlying_train_reader, batch_size=args.batch_size * args.gpus)
     test_reader = paddle.batch(
-        paddle.dataset.mnist.test(), batch_size=args.batch_size)
+        underlying_test_reader, batch_size=args.batch_size)
+
+    if not args.use_reader_op and args.use_py_reader_op:
+        return avg_cost, inference_program, opt, train_reader, test_reader, batch_acc, \
+    feed_queue, underlying_train_reader, underlying_test_reader, \
+    (dshape, [1])
+    else:
         return avg_cost, inference_program, opt, train_reader, test_reader, batch_acc

benchmark/fluid/models/resnet.py

@@ -163,6 +163,16 @@ def get_model(args):
             fluid.layers.batch(
                 data_file, batch_size=args.batch_size))
         input, label = fluid.layers.read_file(data_file)
+    elif args.use_py_reader_op:
+        data_file, feed_queue = fluid.layers.py_array_reader(
+            capacity=args.feed_queue_capacity,
+            shapes=[[-1] + dshape, [-1, 1]],
+            lod_levels=[0, 0],
+            dtypes=['float32', 'int64'])
+        data_file = fluid.layers.double_buffer(
+            fluid.layers.batch(
+                data_file, batch_size=args.batch_size))
+        input, label = fluid.layers.read_file(data_file)
     else:
         input = fluid.layers.data(name='data', shape=dshape, dtype='float32')
         label = fluid.layers.data(name='label', shape=[1], dtype='int64')
@@ -204,5 +214,11 @@ def get_model(args):
     batched_test_reader = paddle.batch(
         train_reader, batch_size=args.batch_size, drop_last=True)
 
+    if not args.use_reader_op and args.use_py_reader_op:
+        return avg_cost, inference_program, optimizer, batched_train_reader,\
+                      batched_test_reader, batch_acc, \
+                      feed_queue, train_reader, test_reader, \
+                      (dshape, [1])
+    else:
         return avg_cost, inference_program, optimizer, batched_train_reader,\
                       batched_test_reader, batch_acc

benchmark/fluid/models/stacked_dynamic_lstm.py

@@ -44,7 +44,7 @@ def crop_sentence(reader, crop_size):
 
 
 def get_model(args):
-    if args.use_reader_op:
+    if args.use_reader_op or args.use_py_reader_op:
         raise Exception(
             "stacked_dynamic_lstm do not support reader op for now.")
     lstm_size = 512

benchmark/fluid/models/vgg.py

@@ -54,12 +54,16 @@ def vgg16_bn_drop(input):
 
 def get_model(args):
     if args.data_set == "cifar10":
+        underlying_train_reader = paddle.dataset.cifar.train10()
+        underlying_test_reader = paddle.dataset.cifar.test10()
         classdim = 10
         if args.data_format == 'NCHW':
             data_shape = [3, 32, 32]
         else:
             data_shape = [32, 32, 3]
     else:
+        underlying_train_reader = paddle.dataset.flowers.train()
+        underlying_test_reader = paddle.dataset.flowers.test()
         classdim = 102
         if args.data_format == 'NCHW':
             data_shape = [3, 224, 224]
@@ -81,6 +85,16 @@ def get_model(args):
             fluid.layers.batch(
                 data_file, batch_size=args.batch_size))
         images, label = fluid.layers.read_file(data_file)
+    elif args.use_py_reader_op:
+        data_file, feed_queue = fluid.layers.py_array_reader(
+            capacity=args.feed_queue_capacity,
+            shapes=[[-1] + data_shape, [-1, 1]],
+            lod_levels=[0, 0],
+            dtypes=["float32", "int64"])
+        data_file = fluid.layers.double_buffer(
+            fluid.layers.batch(
+                data_file, batch_size=args.batch_size))
+        images, label = fluid.layers.read_file(data_file)
     else:
         images = fluid.layers.data(
             name='data', shape=data_shape, dtype='float32')
@@ -109,13 +123,14 @@ def get_model(args):
     # data reader
     train_reader = paddle.batch(
         paddle.reader.shuffle(
-            paddle.dataset.cifar.train10()
-            if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
-            buf_size=5120),
+            underlying_train_reader, buf_size=5120),
         batch_size=args.batch_size * args.gpus)
     test_reader = paddle.batch(
-        paddle.dataset.cifar.test10()
-        if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
-        batch_size=args.batch_size)
+        underlying_test_reader, batch_size=args.batch_size)
 
+    if not args.use_reader_op and args.use_py_reader_op:
+        return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc, \
+                      feed_queue, underlying_train_reader, underlying_test_reader, \
+                      (data_shape, [1])
+    else:
         return avg_cost, inference_program, optimizer, train_reader, test_reader, batch_acc

paddle/fluid/operators/reader/CMakeLists.txt

@@ -24,6 +24,7 @@ reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_o
 reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc)
 reader_library(create_threaded_reader_op SRCS create_threaded_reader_op.cc)
 reader_library(create_custom_reader_op SRCS create_custom_reader_op.cc)
+reader_library(create_py_array_reader_op SRCS create_py_array_reader_op.cc)
 
 cc_test(reader_blocking_queue_test SRCS reader_blocking_queue_test.cc)
 # Export local libraries to parent

paddle/fluid/operators/reader/create_py_array_reader_op.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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 "paddle/fluid/operators/reader/py_array_feed_queue.h"
+
+namespace paddle {
+namespace operators {
+namespace reader {
+
+class PyArrayReader : public framework::ReaderBase {
+ public:
+  explicit PyArrayReader(const std::shared_ptr<PyArrayFeedQueue>& queue) {
+    PADDLE_ENFORCE(queue != nullptr, "PyArrayFeedQueue must not be null");
+    queue_ = queue;
+  }
+
+  void ReadNext(std::vector<framework::LoDTensor>* out) override {
+    *out = queue_->Dequeue();
+  }
+
+  void ReInit() override {
+    // PADDLE_THROW("PyArrayReader does not support ReInit()");
+  }
+
+ private:
+  std::shared_ptr<PyArrayFeedQueue> queue_;
+};
+
+class CreatePyArrayReaderOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+
+ private:
+  void RunImpl(const framework::Scope& scope,
+               const platform::Place& dev_place) const override {
+    const std::string& feeder_name = Attr<std::string>("feeder_name");
+    auto* feeder_holder_var = scope.FindVar(feeder_name);
+    PADDLE_ENFORCE(feeder_holder_var != nullptr,
+                   "No PyArrayFeedQueue variable with name %s found",
+                   feeder_name);
+    auto* feeder_holder =
+        feeder_holder_var->template GetMutable<PyArrayFeedQueueHolder>();
+    auto* out = scope.FindVar(Output("Out"))
+                    ->template GetMutable<framework::ReaderHolder>();
+    out->Reset(new PyArrayReader(feeder_holder->GetFeeder()));
+  }
+};
+
+class CreatePyArrayReaderOpMaker : public FileReaderMakerBase {
+ protected:
+  void Apply() override {
+    AddAttr<std::string>("feeder_name",
+                         "Name of the `PyArrayFeedQueueHolder` variable");
+
+    AddComment(R"DOC(
+			Create PyArrayReader to accept Python data feeding.
+      )DOC");
+  }
+};
+
+}  // namespace reader
+}  // namespace operators
+}  // namespace paddle
+
+namespace reader = ::paddle::operators::reader;
+
+REGISTER_FILE_READER_OPERATOR(create_py_array_reader,
+                              reader::CreatePyArrayReaderOp,
+                              reader::CreatePyArrayReaderOpMaker);

paddle/fluid/operators/reader/py_array_feed_queue.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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 <algorithm>
+#include <condition_variable>  //NOLINT
+#include <memory>
+#include <mutex>  // NOLINT
+#include <vector>
+#include "glog/logging.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/operators/reader/py_blocking_queue.h"
+#include "paddle/fluid/operators/reader/reader_op_registry.h"
+#include "paddle/fluid/pybind/tensor_py.h"
+
+namespace paddle {
+namespace operators {
+namespace reader {
+
+using PyTuple = ::pybind11::tuple;
+using PyArray = ::pybind11::array;
+
+template <typename T>
+using PyArrayT = ::pybind11::array_t<T, ::pybind11::array::c_style |
+                                            ::pybind11::array::forcecast>;
+
+class PyArrayToTensorVisitor : public boost::static_visitor<void> {
+ public:
+#define PY_ARRAY_TO_TENSOR_WITH_TYPE(dtype, func_name)                       \
+  pybind::func_name(tensor_, static_cast<const PyArrayT<dtype>&>(py_array_), \
+                    place)
+
+#define PY_ARRAY_TO_TENSOR(func_name)                  \
+  if (IsType<size_t>()) {                              \
+    PY_ARRAY_TO_TENSOR_WITH_TYPE(size_t, func_name);   \
+  } else if (IsType<int64_t>()) {                      \
+    PY_ARRAY_TO_TENSOR_WITH_TYPE(int64_t, func_name);  \
+  } else if (IsType<int32_t>()) {                      \
+    PY_ARRAY_TO_TENSOR_WITH_TYPE(int32_t, func_name);  \
+  } else if (IsType<int16_t>()) {                      \
+    PY_ARRAY_TO_TENSOR_WITH_TYPE(int16_t, func_name);  \
+  } else if (IsType<uint8_t>()) {                      \
+    PY_ARRAY_TO_TENSOR_WITH_TYPE(uint8_t, func_name);  \
+  } else if (IsType<float>()) {                        \
+    PY_ARRAY_TO_TENSOR_WITH_TYPE(float, func_name);    \
+  } else if (IsType<double>()) {                       \
+    PY_ARRAY_TO_TENSOR_WITH_TYPE(double, func_name);   \
+  } else {                                             \
+    PADDLE_THROW("unsupported dtype of python array"); \
+  }
+
+  PyArrayToTensorVisitor(const PyArray& py_array, framework::Tensor* tensor)
+      : py_array_(py_array), tensor_(tensor) {}
+
+  void operator()(const platform::CPUPlace& place) {
+    PY_ARRAY_TO_TENSOR(PyCPUTensorSetFromArray);
+  }
+
+  void operator()(const platform::CUDAPlace& place) {
+#ifdef PADDLE_WITH_CUDA
+    PY_ARRAY_TO_TENSOR(PyCUDATensorSetFromArray);
+#else
+    PADDLE_THROW("CUDAPlace is not supported in CPU only version");
+#endif
+  }
+
+  void operator()(const platform::CUDAPinnedPlace& place) {
+#ifdef PADDLE_WITH_CUDA
+    PY_ARRAY_TO_TENSOR(PyCUDAPinnedTensorSetFromArray);
+#else
+    PADDLE_THROW("CUDAPinnedPlace is not supported in CPU only version");
+#endif
+  }
+
+#undef PY_ARRAY_TO_TENSOR
+#undef PY_ARRAY_TO_TENSOR_WITH_TYPE
+
+ private:
+  template <typename T>
+  inline bool IsType() const {
+    return ::pybind11::isinstance<PyArrayT<T>>(py_array_);
+  }
+
+ private:
+  const PyArray& py_array_;
+  framework::Tensor* tensor_;
+};
+
+class PyArrayFeedQueueHolder;
+
+// PyArrayFeedQueue must be thread-safe
+class PyArrayFeedQueue {
+  friend class PyArrayFeedQueueHolder;
+
+ private:
+  PyArrayFeedQueue(size_t capacity, const std::vector<framework::DDim>& dims,
+                   const platform::Place& place)
+      : dims_(dims), place_(place) {
+    queue_.reset(
+        new PyBlockingQueue<std::vector<framework::LoDTensor>>(capacity));
+  }
+
+ public:
+  ~PyArrayFeedQueue() { Close(); }
+
+  bool Enqueue(const std::vector<PyArray>& py_array_vec) {
+    auto lod_tensor_vec = PyArrayVecToLoDTensorVec(py_array_vec);
+    VLOG(5) << "Enqueue at address " << reinterpret_cast<void*>(this);
+    return queue_->Send(std::move(lod_tensor_vec));
+  }
+
+  bool Enqueue(const std::vector<framework::LoDTensor>& tensor_vec) {
+    VLOG(5) << "Enqueue at address " << reinterpret_cast<void*>(this);
+    return queue_->Send(tensor_vec);
+  }
+
+  std::vector<framework::LoDTensor> Dequeue() {
+    VLOG(5) << "Dequeue at address " << reinterpret_cast<void*>(this);
+    std::vector<framework::LoDTensor> ret;
+    return queue_->Receive(&ret) ? ret : std::vector<framework::LoDTensor>();
+  }
+
+  inline size_t Size() const { return queue_->Size(); }
+
+  inline size_t Cap() const { return queue_->Cap(); }
+
+  inline bool IsClosed() const { return queue_->IsClosed(); }
+
+  inline void Close() { queue_->Close(); }
+
+ private:
+  std::vector<framework::LoDTensor> PyArrayVecToLoDTensorVec(
+      const std::vector<PyArray>& py_array_vec) {
+    PADDLE_ENFORCE(dims_.size() == py_array_vec.size(),
+                   "expected input tensor number %d but found %d", dims_.size(),
+                   py_array_vec.size());
+
+    size_t i = 0;
+    if (py_array_vec.size() > 1) {
+      size_t dim0 = py_array_vec[0].shape()[0];
+      for (size_t j = 1; j < py_array_vec.size(); ++j) {
+        PADDLE_ENFORCE(dim0 == py_array_vec[j].shape()[0],
+                       "0-dim of the %d-th input tensor is %d, but 0-dim of "
+                       "the 0-th input tensor is %d",
+                       j, py_array_vec[j].shape()[0], dim0);
+      }
+    }
+
+    std::vector<framework::LoDTensor> lod_tensor_vec;
+    lod_tensor_vec.reserve(py_array_vec.size());
+
+    std::for_each(
+        py_array_vec.begin(), py_array_vec.end(), [&](const PyArray& py_array) {
+          for (int64_t j = 1; j < dims_[i].size(); ++j) {
+            PADDLE_ENFORCE(
+                dims_[i][j] == static_cast<int64_t>(py_array.shape()[j]),
+                "expected %d-dim of %d-th input tensor is %d but found %d", j,
+                i, dims_[i][j], py_array.shape()[j]);
+          }
+
+          lod_tensor_vec.emplace_back(framework::LoDTensor());
+          PyArrayToTensorVisitor visitor(py_array, &(lod_tensor_vec.back()));
+          boost::apply_visitor(visitor, place_);
+          ++i;
+        });
+    return lod_tensor_vec;
+  }
+
+  std::unique_ptr<PyBlockingQueue<std::vector<framework::LoDTensor>>> queue_;
+  std::vector<framework::DDim> dims_;
+  platform::Place place_;
+};
+
+class PyArrayFeedQueueHolder {
+ public:
+  PyArrayFeedQueueHolder() {}
+
+  void InitOnce(size_t capacity, const std::vector<framework::DDim>& dims,
+                const platform::Place& place) {
+    PADDLE_ENFORCE(
+        feeder_ == nullptr,
+        "PyArrayFeedQueueHolder::InitOnce() can only be called once");
+    feeder_.reset(new PyArrayFeedQueue(capacity, dims, place));
+  }
+
+  std::shared_ptr<PyArrayFeedQueue> GetFeeder() { return feeder_; }
+  const std::shared_ptr<PyArrayFeedQueue>& GetFeeder() const { return feeder_; }
+
+ private:
+  std::shared_ptr<PyArrayFeedQueue> feeder_;
+};
+
+}  // namespace reader
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/reader/py_blocking_queue.h

+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// 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 <condition_variable>  // NOLINT
+#include <deque>
+
+#include "Python.h"
+#include "paddle/fluid/platform/enforce.h"
+#include "pybind11/pybind11.h"
+
+namespace paddle {
+namespace operators {
+namespace reader {
+
+// PyBlockingQueue is designed for PyArrayFeedQueue
+// PyBlockingQueue would release GIL of Python when
+// the queue is full to avoid deadlock.
+template <typename T>
+class PyBlockingQueue {
+ public:
+  explicit PyBlockingQueue(size_t capacity)
+      : capacity_(capacity), closed_(false) {
+    PADDLE_ENFORCE_GT(
+        capacity_, 0,
+        "The capacity of a reader::PyBlockingQueue must be greater than 0.");
+  }
+
+  ~PyBlockingQueue() { Close(); }
+
+  bool Send(const T& elem) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    receive_cv_.notify_one();
+    if (queue_.size() >= capacity_ && (!closed_)) {
+      pybind11::gil_scoped_release release;
+      send_cv_.wait(lock, [&] { return queue_.size() < capacity_ || closed_; });
+    }
+    if (closed_) {
+      VLOG(5)
+          << "WARNING: Sending an element to a closed reader::BlockingQueue.";
+      return false;
+    }
+    PADDLE_ENFORCE_LT(queue_.size(), capacity_);
+    queue_.push_back(elem);
+    return true;
+  }
+
+  bool Send(T&& elem) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    receive_cv_.notify_one();
+    if (queue_.size() >= capacity_ && (!closed_)) {
+      pybind11::gil_scoped_release release;
+      send_cv_.wait(lock, [&] { return queue_.size() < capacity_ || closed_; });
+    }
+    if (closed_) {
+      VLOG(5)
+          << "WARNING: Sending an element to a closed reader::BlokcingQueue.";
+      return false;
+    }
+    PADDLE_ENFORCE_LT(queue_.size(), capacity_);
+    queue_.emplace_back(std::move(elem));
+    return true;
+  }
+
+  bool Receive(T* elem) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    send_cv_.notify_one();
+    receive_cv_.wait(lock, [&] { return !queue_.empty() || closed_; });
+    if (!queue_.empty()) {
+      PADDLE_ENFORCE_NOT_NULL(elem);
+      *elem = queue_.front();
+      queue_.pop_front();
+      return true;
+    } else {
+      PADDLE_ENFORCE(closed_);
+      return false;
+    }
+  }
+
+  void Close() {
+    std::lock_guard<std::mutex> lock(mutex_);
+    closed_ = true;
+    send_cv_.notify_all();
+    receive_cv_.notify_all();
+  }
+
+  bool IsClosed() const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return closed_;
+  }
+
+  size_t Cap() const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return capacity_;
+  }
+
+  size_t Size() const {
+    std::lock_guard<std::mutex> lock(mutex_);
+    return queue_.size();
+  }
+
+ private:
+  size_t capacity_;
+  bool closed_;
+  std::deque<T> queue_;
+
+  mutable std::mutex mutex_;
+  mutable std::condition_variable receive_cv_;
+  mutable std::condition_variable send_cv_;
+};
+}  // namespace reader
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/pybind/pybind.cc

@@ -34,6 +34,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/operators/activation_op.h"
+#include "paddle/fluid/operators/reader/py_array_feed_queue.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -297,6 +298,42 @@ All parameter, weight, gradient are variables in Paddle.
   py::class_<framework::ReaderHolder>(m, "Reader", "")
       .def("reset", &framework::ReaderHolder::ReInit);
 
+  using PyArrayFeedQueue = ::paddle::operators::reader::PyArrayFeedQueue;
+  using PyArrayFeedQueueHolder =
+      ::paddle::operators::reader::PyArrayFeedQueueHolder;
+  using PyArray = ::paddle::operators::reader::PyArray;
+  py::class_<PyArrayFeedQueue>(m, "PyArrayFeedQueue", "")
+      .def(
+          "enqueue",
+          [](PyArrayFeedQueue &self, const std::vector<PyArray> &py_array_vec) {
+            return self.Enqueue(py_array_vec);
+          })
+      .def("enqueue",
+           [](PyArrayFeedQueue &self,
+              const std::vector<framework::LoDTensor> &lod_tensor_vec) {
+             return self.Enqueue(lod_tensor_vec);
+           })
+      .def("size", [](const PyArrayFeedQueue &self) { return self.Size(); })
+      .def("capacity", [](const PyArrayFeedQueue &self) { return self.Cap(); })
+      .def("close", [](PyArrayFeedQueue &self) { return self.Close(); })
+      .def("is_closed",
+           [](const PyArrayFeedQueue &self) { return self.IsClosed(); });
+
+  m.def("init_py_array_feed_queue",
+        [](Variable &var, size_t capacity,
+           const std::vector<std::vector<int64_t>> &shapes,
+           const ::paddle::platform::Place &place) -> PyArrayFeedQueue * {
+          std::vector<DDim> dims(shapes.size());
+          std::transform(shapes.begin(), shapes.end(), dims.begin(),
+                         [](const std::vector<int64_t> &shape) {
+                           return make_ddim(shape);
+                         });
+          auto *holder = var.GetMutable<PyArrayFeedQueueHolder>();
+          holder->InitOnce(capacity, dims, place);
+          return holder->GetFeeder().get();
+        },
+        py::return_value_policy::reference);
+
   py::class_<Scope>(m, "Scope", "")
       .def("var",
            [](Scope &self, const std::string &name) -> Variable * {
@@ -463,10 +500,11 @@ All parameter, weight, gradient are variables in Paddle.
 #ifdef PADDLE_WITH_DISTRIBUTE
       .def("complete", &Executor::Complete)
 #endif
-      .def("run",
-           (void (Executor::*)(const ProgramDesc &, Scope *, int, bool, bool)) &
-               Executor::Run);
-
+      .def("run", [](Executor &self, const ProgramDesc &prog, Scope *scope,
+                     int block_id, bool create_local_scope, bool create_vars) {
+        pybind11::gil_scoped_release release;
+        self.Run(prog, scope, block_id, create_local_scope, create_vars);
+      });
   m.def("init_gflags", framework::InitGflags);
   m.def("init_glog", framework::InitGLOG);
   m.def("init_devices",

paddle/fluid/pybind/tensor_py.h

@@ -146,7 +146,7 @@ void PyCPUTensorSetFromArray(
 template <>
 // This following specialization maps uint16_t in the parameter type to
 // platform::float16.
-void PyCPUTensorSetFromArray(
+inline void PyCPUTensorSetFromArray(
     framework::Tensor *self,
     pybind11::array_t<uint16_t,
                       pybind11::array::c_style | pybind11::array::forcecast>
@@ -185,7 +185,7 @@ void PyCUDATensorSetFromArray(
 template <>
 // This following specialization maps uint16_t in the parameter type to
 // platform::float16.
-void PyCUDATensorSetFromArray(
+inline void PyCUDATensorSetFromArray(
     framework::Tensor *self,
     pybind11::array_t<uint16_t,
                       pybind11::array::c_style | pybind11::array::forcecast>
@@ -224,7 +224,7 @@ void PyCUDAPinnedTensorSetFromArray(
 template <>
 // This following specialization maps uint16_t in the parameter type to
 // platform::float16.
-void PyCUDAPinnedTensorSetFromArray(
+inline void PyCUDAPinnedTensorSetFromArray(
     framework::Tensor *self,
     pybind11::array_t<uint16_t,
                       pybind11::array::c_style | pybind11::array::forcecast>

python/paddle/fluid/layers/io.py

@@ -24,7 +24,8 @@ from layer_function_generator import generate_layer_fn, templatedoc
 __all__ = [
     'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'Recv',
     'open_recordio_file', 'open_files', 'read_file', 'shuffle', 'batch',
-    'double_buffer', 'random_data_generator', 'Preprocessor', 'load'
+    'double_buffer', 'random_data_generator', 'py_array_reader', 'Preprocessor',
+    'load'
 ]
 
 
@@ -448,6 +449,61 @@ def random_data_generator(low, high, shapes, lod_levels, for_parallel=True):
     return monkey_patch_reader_methods(main_prog_var)
 
 
+# UNCHECK(zengjinle)
+def py_array_reader(capacity,
+                    shapes,
+                    lod_levels,
+                    dtypes,
+                    place=None,
+                    for_parallel=True):
+
+    if place is None:
+        place = core.CPUPlace()
+
+    if not isinstance(place, core.Place):
+        new_place = core.Place()
+        new_place.set_place(place)
+        place = new_place
+
+    dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes]
+    shape_concat = []
+    ranks = []
+
+    for shape in shapes:
+        shape_concat.extend(shape)
+        ranks.append(len(shape))
+
+    feeder_name = unique_name('py_array_feed_queue')
+    var = global_scope().var(feeder_name)
+
+    #feed_shapes = [shape[1:] for shape in shapes]
+    feed_queue = core.init_py_array_feed_queue(var, capacity, shapes, place)
+
+    startup_blk = default_startup_program().current_block()
+    startup_var = startup_blk.create_var(
+        name=unique_name('create_py_array_reader'))
+    startup_blk.append_op(
+        type='create_py_array_reader',
+        outputs={'Out': [startup_var]},
+        attrs={
+            'shape_concat': shape_concat,
+            'lod_levels': lod_levels,
+            'ranks': ranks,
+            'feeder_name': feeder_name
+        })
+
+    startup_var.desc.set_dtypes(dtypes)
+    startup_var.persistable = True
+
+    main_prog_var = _copy_reader_var_(default_main_program().current_block(),
+                                      startup_var)
+
+    if for_parallel:
+        main_prog_var = parallel(reader=main_prog_var)
+
+    return monkey_patch_reader_methods(main_prog_var), feed_queue
+
+
 def open_files(filenames,
                shapes,
                lod_levels,