Refine DataLoader support multi-processing (#23107)

* add DataLoader, Dataset, BatchSampler

paddle/fluid/imperative/data_loader.cc

@@ -22,21 +22,23 @@
 #include <atomic>
 #include <csignal>
 #include <map>
+#include <set>
 
+#include "paddle/fluid/memory/allocation/mmap_allocator.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
 namespace imperative {
 
-static std::map<int64_t, pid_t> load_process_pids;
+static std::map<int64_t, std::set<pid_t>> load_process_pids;
 
-void SetLoadProcessPID(int64_t key, pid_t pid) {
-  VLOG(3) << "Dygraph Data Loader: set loader child process PID (" << key
-          << ", " << pid << ")";
-  load_process_pids[key] = pid;
+void SetLoadProcessPIDs(int64_t key, std::set<pid_t> pids) {
+  VLOG(3) << "DataLoader: set loader child process PID (" << key
+          << ", pid number: " << pids.size() << ")";
+  load_process_pids[key] = pids;
 }
 
-void EraseLoadProcessPID(int64_t key) {
+void EraseLoadProcessPIDs(int64_t key) {
   auto it = load_process_pids.find(key);
   // Note: Can not find key also possible
   if (it != load_process_pids.end()) {
@@ -54,8 +56,12 @@ void EraseLoadProcessPID(int64_t key) {
 // siginfo_t doc: https://www.mkssoftware.com/docs/man5/siginfo_t.5.asp
 // waitid doc: https://linux.die.net/man/2/waitid
 
+// clear mmap fds on signal handler, make sure mmap clear will be called
+// on signal handling and no need to register mmap clear up handler on
+// python side. If shared memory is not used Clear() will do nothing.
 #define SIGNAL_HANDLE(SIGNAL)                               \
   do {                                                      \
+    memory::allocation::MemoryMapFdSet::Instance().Clear(); \
     struct sigaction sa;                                    \
     sa.sa_handler = SIG_DFL;                                \
     sa.sa_flags = 0;                                        \
@@ -106,30 +112,40 @@ void SetLoadProcessSignalHandler() {
 
 void ThrowErrorIfLoadProcessFailed() {
   int error;
+  std::set<pid_t> *pids_set;
   pid_t process_pid;
   siginfo_t infop;
 
-  for (auto &w : load_process_pids) {
-    process_pid = w.second;
-    // Use waitid rather than waitpid so that we can set NOWAIT, and that Python
-    // and other handlers can get whatever info they want about the child.
+  for (auto &p : load_process_pids) {
+    pids_set = &(p.second);
+    for (auto pid_it = pids_set->begin(); pid_it != pids_set->end(); ++pid_it) {
+      process_pid = *pid_it;
+      // Use waitid rather than waitpid so that we can set NOWAIT, and that
+      // Python and other handlers can get whatever info they want about the
+      // child.
       infop.si_pid = 0;
-    VLOG(3) << "Dygraph Data Loader: monitor loader child process "
-            << process_pid;
+      VLOG(3) << "DataLoader: monitor loader child process " << process_pid;
       error = waitid(P_PID, process_pid, &infop, WEXITED | WNOHANG | WNOWAIT);
       // ignore errors and case with no waitable child
       if (error < 0 || infop.si_pid == 0) continue;
       if (infop.si_code == CLD_EXITED &&
           infop.si_status != EXIT_SUCCESS) {  // exit with error
+        pids_set->clear();
         PADDLE_THROW(platform::errors::Fatal(
             "DataLoader process (pid %ld) exited unexpectedly with code %d. "
-          "Error detailed are lost due to multiprocessing. Rerunning with "
+            "Error detailed are lost due to multiprocessing. Rerunning with:\n"
+            "  1. If run DataLoader by DataLoader.from_generator(...), run "
+            "with "
             "DataLoader.from_generator(..., use_multiprocess=False) may give "
-          "better error trace.",
+            "better error trace.\n"
+            "  2. If run DataLoader by DataLoader(dataset, ...), run with "
+            "DataLoader(dataset, ..., num_workers=0) may give better error "
+            "trace",
             process_pid, infop.si_status));
       } else if (infop.si_code == CLD_KILLED ||
                  infop.si_code == CLD_DUMPED) {  // killed by signal
         if (infop.si_status == SIGBUS) {
+          pids_set->clear();
           PADDLE_THROW(platform::errors::Fatal(
               "DataLoader process (pid %ld) exited is killed by signal: %s.\n"
               "  It may be caused by insufficient shared storage space. This "
@@ -138,7 +154,14 @@ void ThrowErrorIfLoadProcessFailed() {
               "space of `/dev/shm`. Shared storage space needs to be greater "
               "than (DataLoader Num * DataLoader queue capacity * 1 batch data "
               "size).\n  You can solve this problem by increasing the shared "
-            "storage space or reducing the queue capacity appropriately.",
+              "storage space or reducing the queue capacity appropriately.\n",
+              "  1. If run DataLoader by DataLoader.from_generator(...), queue "
+              "capacity is set by from_generator(..., capacity=xx, ...).\n"
+              "  2. If run DataLoader by DataLoader(dataset, ...), queue "
+              "capacity is set as 2 times of the max value of num_workers and "
+              "len(places).\n"
+              "  3. If run by DataLoader(dataset, ..., use_shared_memory=True),"
+              " set use_shared_memory=False for not using shared memory.",
               process_pid, strsignal(infop.si_status)));
         } else {
           PADDLE_THROW(platform::errors::Fatal(
@@ -147,6 +170,7 @@ void ThrowErrorIfLoadProcessFailed() {
         }
       }
     }
+  }
 }
 
 }  // namespace imperative

paddle/fluid/imperative/data_loader.h

@@ -18,12 +18,13 @@
 
 #include <unistd.h>
 #include <cstdint>
+#include <set>
 
 namespace paddle {
 namespace imperative {
 
-extern void SetLoadProcessPID(int64_t key, pid_t pid);
-extern void EraseLoadProcessPID(int64_t key);
+extern void SetLoadProcessPIDs(int64_t key, std::set<pid_t> pids);
+extern void EraseLoadProcessPIDs(int64_t key);
 extern void SetLoadProcessSignalHandler();
 extern void ThrowErrorIfLoadProcessFailed();
 

paddle/fluid/pybind/imperative.cc

@@ -20,6 +20,7 @@ limitations under the License. */
 #include <pybind11/functional.h>
 #include <pybind11/stl.h>
 #include <memory>
+#include <set>
 #include <string>
 #include <unordered_map>
 #include <utility>
@@ -290,11 +291,22 @@ void BindImperative(py::module *m_ptr) {
 
 #ifndef _WIN32
   // Dygraph DataLoader signal handler
-  m.def("_set_process_pid", [](int64_t key, pid_t pid) {
-    imperative::SetLoadProcessPID(key, pid);
+  m.def("_set_process_pids", [](int64_t key, py::object &obj) {
+    PADDLE_ENFORCE_EQ(
+        py::isinstance<py::tuple>(obj) || py::isinstance<py::list>(obj), true,
+        platform::errors::InvalidArgument(
+            "The subprocess ids set in DataLoader is illegal."
+            "Expected data type is tuple or list, but received %s",
+            obj.get_type()));
+    py::list pids = py::cast<py::list>(obj);
+    std::set<pid_t> pids_set = {};
+    for (size_t i = 0; i < pids.size(); i++) {
+      pids_set.insert(pids[i].cast<pid_t>());
+    }
+    imperative::SetLoadProcessPIDs(key, pids_set);
   });
-  m.def("_erase_process_pid",
-        [](int64_t key) { imperative::EraseLoadProcessPID(key); });
+  m.def("_erase_process_pids",
+        [](int64_t key) { imperative::EraseLoadProcessPIDs(key); });
   m.def("_set_process_signal_handler",
         []() { imperative::SetLoadProcessSignalHandler(); });
   m.def("_throw_error_if_process_failed",

python/paddle/distributed/utils.py

@@ -252,7 +252,9 @@ def get_cluster(node_ips, node_ip, paddle_ports, selected_gpus):
 def terminate_local_procs(procs):
     for p in procs:
         if p.proc.poll() is None:
-            p.proc.terminate()
+            # subprocess need to release resource(e.g. shared memory)
+            # use join to wait subprocess releasing
+            p.proc.join(timeout=1)
             p.log_fn.close()
             logger.debug("terminate process id:{}".format(p.proc.pid))
 

python/paddle/fluid/core.py

@@ -185,8 +185,8 @@ if avx_supported():
         from .core_avx import _load_dygraph_dict
         from .core_avx import _create_loaded_parameter
         if sys.platform != 'win32':
-            from .core_avx import _set_process_pid
-            from .core_avx import _erase_process_pid
+            from .core_avx import _set_process_pids
+            from .core_avx import _erase_process_pids
             from .core_avx import _set_process_signal_handler
             from .core_avx import _throw_error_if_process_failed
             from .core_avx import _convert_to_tensor_list
@@ -229,8 +229,8 @@ if load_noavx:
         from .core_noavx import _load_dygraph_dict
         from .core_noavx import _create_loaded_parameter
         if sys.platform != 'win32':
-            from .core_noavx import _set_process_pid
-            from .core_noavx import _erase_process_pid
+            from .core_noavx import _set_process_pids
+            from .core_noavx import _erase_process_pids
             from .core_noavx import _set_process_signal_handler
             from .core_noavx import _throw_error_if_process_failed
             from .core_noavx import _convert_to_tensor_list

python/paddle/fluid/dataloader/__init__.py

+#   Copyright (c) 2020 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.
+
+from __future__ import print_function
+
+from . import dataset
+from .dataset import *
+
+from . import batch_sampler
+from .batch_sampler import *
+
+__all__ = dataset.__all__ \
+        + batch_sampler.__all__

python/paddle/fluid/dataloader/batch_sampler.py

+#   Copyright (c) 2020 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.
+
+from __future__ import print_function
+from __future__ import division
+
+import numpy as np
+from .dataset import Dataset
+
+__all__ = ["BatchSampler"]
+
+
+class BatchSampler(object):
+    """
+    A base implement of batch sampler used by `paddle.io.DataLoader`
+    which yield mini-batch indices(a list/tuple with length as
+    mini-batch size and holds sample indices) iterably.
+
+    Batch sampler used by :code:`paddle.io.DataLoader` should be a subclass
+    of :code:`paddle.io.BatchSampler`, BatchSampler subclasses should
+    implement following methods:
+
+    :code:`__iter__`: return mini-batch indices iterably.
+
+    :code:`__len__`: get mini-batch number in an epoch.
+
+
+    Args:
+        dataset(Dataset): this could be a :code:`paddle.io.Dataset` 
+                implement or other python object which implemented
+                :code:`__len__` for BatchSampler to get indices as the
+                range of :attr:`dataset` length. Default None.
+        indices (list|tuple): a substitution parameter for
+                :attr:`dataset` either :attr:`dataset` or
+                :attr:`indices` should be set, give the whole
+                indices to sampler from directly. Default None.
+        shuffle(bool): whether to shuffle indices order before genrating
+                batch indices. Default False.
+        batch_size(int): sample indice number in a mini-batch indices.
+        drop_last(bool): whether drop the last incomplete batch dataset size
+            is not divisible by the batch size. Default False
+
+    Returns:
+        BatchSampler: an iterable object for indices iterating
+
+    Examples:
+        
+        .. code-block:: python
+            
+            from paddle.io import BatchSampler, Dataset
+
+            # init with indices
+            bs = BatchSampler(indices=list(range(100)),
+                              shuffle=True,
+                              batch_size=8,
+                              drop_last=True)
+
+            for batch_indices in bs:
+                print(batch_indices)
+
+            # init with dataset
+            class RandomDataset(Dataset):
+                def __init__(self, num_samples):
+                    self.num_samples = num_samples
+            
+                def __getitem__(self, idx):
+                    image = np.random.random([784]).astype('float32')
+                    label = np.random.randint(0, 9, (1, )).astype('int64')
+                    return image, label
+                
+                def __len__(self):
+                    return self.num_samples
+            
+            bs = BatchSampler(dataset=RandomDataset(100),
+                              shuffle=False,
+                              batch_size=16,
+                              drop_last=False)
+
+            for batch_indices in bs:
+                print(batch_indices)
+
+    see `paddle.io.DataLoader`
+
+    """
+
+    def __init__(self,
+                 dataset=None,
+                 indices=None,
+                 shuffle=False,
+                 batch_size=1,
+                 drop_last=False):
+        if dataset is None:
+            assert indices is not None, \
+                "either dataset or indices should be set"
+            assert isinstance(indices, list) or isinstance(indices, tuple), \
+                "indices should be a list or tuple, but got {}".format(type(indices))
+            self.indices = indices
+        else:
+            assert isinstance(dataset, Dataset), \
+                "dataset should be an instance of paddle.io.Dataset"
+            assert indices is None, \
+                "should not set both dataset and indices"
+            self.indices = list(range(len(dataset)))
+
+        assert isinstance(batch_size, int) and batch_size > 0, \
+            "batch_size should be a positive integer, but got {}".format(batch_size)
+        self.batch_size = batch_size
+        assert isinstance(shuffle, bool), \
+            "shuffle should be a boolean value, but got {}".format(type(shuffle))
+        self.shuffle = shuffle
+        assert isinstance(drop_last, bool), \
+            "drop_last should be a boolean value, but got {}".format(type(drop_last))
+        self.drop_last = drop_last
+
+    def __iter__(self):
+        if self.shuffle:
+            np.random.shuffle(self.indices)
+        _iter = iter(self.indices)
+
+        batch_indices = []
+        for idx in _iter:
+            batch_indices.append(idx)
+            if len(batch_indices) == self.batch_size:
+                yield batch_indices
+                batch_indices = []
+        if not self.drop_last and len(batch_indices) > 0:
+            yield batch_indices
+
+    def __len__(self):
+        num_samples = len(self.indices)
+        num_samples += int(not self.drop_last) * (self.batch_size - 1)
+        return num_samples // self.batch_size

python/paddle/fluid/dataloader/dataset.py

+#   Copyright (c) 2020 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.
+
+from __future__ import print_function
+
+import paddle.dataset.common
+
+__all__ = ["Dataset"]
+
+
+class Dataset(object):
+    """
+    An abstract class to encapsulates methods and behaviors of datasets.
+
+    All datasets in map-style(dataset samples can be get by a given key)
+    should be a subclass of `paddle.io.Dataset`. All subclasses should
+    implement following methods:
+
+    :code:`__getitem__`: get sample from dataset with a given index. This
+    method is required by reading dataset sample in :code:`paddle.io.DataLoader`.
+
+    :code:`__len__`: return dataset sample number. This method is required
+    by some implements of :code:`paddle.io.BatchSampler`
+
+    see :code:`paddle.io.DataLoader`.
+
+    Examples:
+        
+        .. code-block:: python
+
+            import numpy as np
+            from paddle.io import Dataset
+            
+            # define a random dataset
+            class RandomDataset(Dataset):
+                def __init__(self, num_samples):
+                    self.num_samples = num_samples
+            
+                def __getitem__(self, idx):
+                    image = np.random.random([784]).astype('float32')
+                    label = np.random.randint(0, 9, (1, )).astype('int64')
+                    return image, label
+                
+                def __len__(self):
+                    return self.num_samples
+            
+            dataset = RandomDataset(10)
+            for i in range(len(dataset)):
+                print(dataset[i])
+
+    """
+
+    def __init__(self):
+        pass
+
+    def __getitem__(self, idx):
+        raise NotImplementedError("'{}' not implement in class "\
+                "{}".format('__getitem__', self.__class__.__name__))
+
+    def __len__(self):
+        raise NotImplementedError("'{}' not implement in class "\
+                "{}".format('__len__', self.__class__.__name__))

python/paddle/fluid/io.py

@@ -37,6 +37,8 @@ from paddle.fluid.compiler import CompiledProgram
 from paddle.fluid.log_helper import get_logger
 from . import reader
 from .reader import *
+from . import dataloader
+from .dataloader import *
 from . import core
 from .. import compat as cpt
 

python/paddle/fluid/multiprocess_utils.py

+# Copyright (c) 2020 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.
+
+import six
+import sys
+import signal
+import atexit
+
+from . import core
+
+# NOTE: queue has a different name in python2 and python3
+if six.PY2:
+    import Queue as queue
+else:
+    import queue
+
+# NOTE: [ mmap files clear ] If there is still data in the multiprocess queue when the main process finishes reading,
+# the data in the queue needs to be popped. Then the LoDTensor read by the main process
+# from the child process will automatically clear the memory-mapped file.
+multiprocess_queue_set = set()
+
+
+def _clear_multiprocess_queue_set():
+    global multiprocess_queue_set
+    for data_queue in multiprocess_queue_set:
+        while True:
+            try:
+                data_queue.get_nowait()
+            except queue.Empty:
+                break
+
+
+# NOTE: main process clear function at exit
+def _cleanup():
+    # NOTE: inter-process Queue shared memory objects clear function
+    _clear_multiprocess_queue_set()
+    # NOTE: main process memory map files clear funciton
+    core._cleanup_mmap_fds()
+
+
+# NOTE: for child process clear function at exit
+def _cleanup_mmap():
+    # clear memory map files in child process
+    core._cleanup_mmap_fds()
+
+
+# NOTE used for register a function to be executed at interpreter exit.
+class CleanupFuncRegistrar():
+    # Record the cleanup functions that have been executed
+    _executed_func_set = set()
+    # Record the cleanup functions that have been registered
+    _registered_func_set = set()
+
+    @classmethod
+    def register(cls, function, signals=[]):
+        def _func_exectuor():
+            if function not in cls._executed_func_set:
+                try:
+                    function()
+                finally:
+                    cls._executed_func_set.add(function)
+
+        def _func_register(function):
+            if not callable(function):
+                raise TypeError("%s is not callable object." % (function))
+            # check function object whether hash-able
+            set([function])
+            if function not in cls._registered_func_set:
+                atexit.register(_func_exectuor)
+                cls._registered_func_set.add(function)
+
+        def _signal_handler(signum=None, frame=None):
+            _func_exectuor()
+            if signum is not None:
+                if signum == signal.SIGINT:
+                    raise KeyboardInterrupt
+                sys.exit(signum)
+
+        def _signal_register(signals):
+            signals = set(signals)
+            for sig in signals:
+                orig_handler = signal.signal(sig, _signal_handler)
+                if orig_handler not in (signal.SIG_DFL, signal.SIG_IGN):
+                    if (sig == signal.SIGINT and
+                            orig_handler is signal.default_int_handler):
+                        continue
+                    if orig_handler not in cls._registered_func_set:
+                        atexit.register(orig_handler)
+                        cls._registered_func_set.add(orig_handler)
+
+        # deal with signals
+        _signal_register(signals)
+        # deal with function
+        _func_register(function)
+
+
+# NOTE: [ mmap files clear ] When the main process exits unexpectedly, the remaining
+# shared memory objects in the inter-process Queue and the main process (mostly in the
+# BlockingQueue) may not be completely released, resulting in the corresponding
+# memory-mapped file remaining on the disk (/dev/shm), so register this function
+# to clean up shared memory objects in these two queues before the python interpreter exits.
+# NOTE: Currently multi-process DataLoader only supports Linux platform
+if not (sys.platform == 'darwin' or sys.platform == 'win32'):
+    CleanupFuncRegistrar.register(_cleanup)
+
+# ------------ SIGCHLD handler setting --------------
+_SIGCHLD_handler_set = False
+
+
+def _set_SIGCHLD_handler():
+    global _SIGCHLD_handler_set
+    if _SIGCHLD_handler_set:
+        return
+
+    current_handler = signal.getsignal(signal.SIGCHLD)
+    if not callable(current_handler):
+        current_handler = None
+
+    def __handler__(signum, frame):
+        # NOTE: Here the signum is SIGCHLD, when the child process exits,
+        # this handler will be called whenever the child process exits
+        # normally or abnormally.
+        core._throw_error_if_process_failed()
+        if current_handler is not None:
+            current_handler(signum, frame)
+
+    signal.signal(signal.SIGCHLD, __handler__)
+    _SIGCHLD_handler_set = True

python/paddle/fluid/reader.py

@@ -21,29 +21,28 @@ import paddle
 from .framework import Program, Variable, program_guard, default_main_program, default_startup_program, in_dygraph_mode, cpu_places
 from .executor import global_scope
 from .data_feeder import DataFeeder, BatchedTensorProvider
+from .multiprocess_utils import multiprocess_queue_set, CleanupFuncRegistrar, _cleanup_mmap, _cleanup, _set_SIGCHLD_handler
+from .dataloader import BatchSampler, Dataset
+from .dataloader.dataloader_iter import _DataLoaderIterSingleProcess, _DataLoaderIterMultiProcess
 from .layers.io import monkey_patch_reader_methods, _copy_reader_var_, double_buffer
 from .unique_name import UniqueNameGenerator
 import logging
 from .dataset import DatasetBase, InMemoryDataset
 
 ### Dygraph DataLoader configs ###
-import atexit
 import os
 import multiprocessing
 import signal
+
 # NOTE: queue has a different name in python2 and python3
-if sys.version_info[0] == 2:
+if six.PY2:
     import Queue as queue
 else:
     import queue
+
 # NOTE: [ avoid hanging & failed quickly ] These value is used in getting data from another process
 QUEUE_GET_TIMEOUT = 60
 
-# NOTE: [ mmap files clear ] If there is still data in the multiprocess queue when the main process finishes reading,
-# the data in the queue needs to be popped. Then the LoDTensor read by the main process
-# from the child process will automatically clear the memory-mapped file.
-multiprocess_queue_set = set()
-
 __all__ = ['PyReader', 'DataLoader']
 
 data_loader_unique_name_generator = UniqueNameGenerator()
@@ -75,84 +74,6 @@ def _convert_places(places):
     return ret
 
 
-def _clear_multiprocess_queue_set():
-    global multiprocess_queue_set
-    for data_queue in multiprocess_queue_set:
-        while True:
-            try:
-                data_queue.get_nowait()
-            except queue.Empty:
-                break
-
-
-# NOTE: main process clear function at exit
-def _cleanup():
-    # NOTE: inter-process Queue shared memory objects clear function
-    _clear_multiprocess_queue_set()
-    # NOTE: main process memory map files clear funciton
-    core._cleanup_mmap_fds()
-
-
-# NOTE used for register a function to be executed at interpreter exit.
-class CleanupFuncRegistrar():
-    # Record the cleanup functions that have been executed
-    _executed_func_set = set()
-    # Record the cleanup functions that have been registered
-    _registered_func_set = set()
-
-    @classmethod
-    def register(cls, function, signals=[signal.SIGTERM]):
-        def _func_exectuor():
-            if function not in cls._executed_func_set:
-                try:
-                    function()
-                finally:
-                    cls._executed_func_set.add(function)
-
-        def _func_register(function):
-            if not callable(function):
-                raise TypeError("%s is not callable object." % (function))
-            # check function object whether hash-able
-            set([function])
-            if function not in cls._registered_func_set:
-                atexit.register(_func_exectuor)
-                cls._registered_func_set.add(function)
-
-        def _signal_handler(signum=None, frame=None):
-            _func_exectuor()
-            if signum is not None:
-                if signum == signal.SIGINT:
-                    raise KeyboardInterrupt
-                sys.exit(signum)
-
-        def _signal_register(signals):
-            signals = set(signals)
-            for sig in signals:
-                orig_handler = signal.signal(sig, _signal_handler)
-                if orig_handler not in (signal.SIG_DFL, signal.SIG_IGN):
-                    if (sig == signal.SIGINT and
-                            orig_handler is signal.default_int_handler):
-                        continue
-                    if orig_handler not in cls._registered_func_set:
-                        atexit.register(orig_handler)
-                        cls._registered_func_set.add(orig_handler)
-
-        # deal with signals
-        _signal_register(signals)
-        # deal with function
-        _func_register(function)
-
-
-# NOTE: [ mmap files clear ] When the main process exits unexpectedly, the remaining
-# shared memory objects in the inter-process Queue and the main process (mostly in the
-# BlockingQueue) may not be completely released, resulting in the corresponding
-# memory-mapped file remaining on the disk (/dev/shm), so register this function
-# to clean up shared memory objects in these two queues before the python interpreter exits.
-# NOTE: Currently multi-process DataLoader only supports Linux platform
-if not (sys.platform == 'darwin' or sys.platform == 'win32'):
-    CleanupFuncRegistrar.register(_cleanup)
-
-
 class DataLoaderBase(object):
     def __init__(self):
         self._places = None
@@ -177,6 +98,264 @@ class DataLoaderBase(object):
 
 
 class DataLoader(object):
+    """
+    DataLoader prodives an iterator which iterates given dataset
+    once by the batch_sampler.
+
+    DataLoader supports single-process and multi-prcess data loading,
+    multi-process workers will be used to load data asynchronously if
+    :attr:`num_workers` is set as a positive number.
+
+    DataLoader only supports map-style dataset(can get a sample from
+    dataset with a given index) currently, for a map-style dataset,
+    please see :code:`paddle.io.Dataset`.
+
+    batch_sampler please see :code:`paddle.io.BatchSampler`
+
+    Args:  
+        dataset(Dataset): the dataset to load data from, should be an
+            instance of subclass of :code:`paddle.io.Dataset`.
+        feed_list (list(Variable)|tuple(Variable)): feed variable list.
+            The variables should be created by :code:`fluid.data()`.
+            :attr:`feed_list` must be set if :attr:`return_list` is
+            False. Default None.
+        places(list(Place)|tuple(Place)): a list of Place, to put data
+            onto, :attr:`places` must be set in both static graph and 
+            dynamic graph mode, in dynamic graph mode, place number must
+            be 1. Default None.
+        return_list (bool): whether the return value on each device is 
+            presented as a list. If :attr:`return_list=False`, the return
+            value on each device would be a dict of str -> LoDTensor, where
+            the key of the dict is the name of each fed variables. If 
+            :attr:`return_list=True`, the return value on each device would
+            be a list(LoDTensor). :attr:`return_list` can only be True
+            in dynamic graph mode. Default False.
+        batch_sampler(BatchSampler): an instance of `paddle.io.BatchSampler`
+            to generate batch indices to draw samples from :attr:`dataset`
+            and combine a batch. Default None.
+        batch_size(int): sample number in a mini-batch, a substitution
+            parameter for :attr:`batch_sampler`, if :attr:`batch_sampler`
+            is not set, a default `paddle.io.BatchSampler` will be used
+            and initialize by :attr:`batch_size`, :attr:`shuffle` and
+            :attr:`drop_last`. Default 1.
+        shuffle(bool): whther to shuffle indices order before genrate
+            batch indices, a substitution parameter for :attr:`batch_sampler`
+            see :attr:`batch_size`. Default False.
+        drop_last(bool): whether drop the last incomplete batch dataset size
+            is not divisible by the batch size, a substitution parameter
+            for :attr:`batch_sampler`, see :attr:`batch_size`. Default False
+        collate_fn(callable): function to generate mini-batch data by merging
+            the sample list, None for only stack each fields of sample in axis
+            0(same as :attr::`np.stack(..., axis=0)`). Default None
+        num_workers(int): the number of subprocess to load data, 0 for no
+            subprocess used and loading data in main process. Default 0
+        use_buffer_reader (bool): whether to use bufferred reader. 
+            If use_buffer_reader=True, the DataLoader would prefetch next 
+            batch data asynchronously, so it would speed up data feeding 
+            and occupies a little more CPU or GPU memory, i.e., the memory
+            of one batch input data. Default True.
+        use_shared_memory (bool): whether to use shared memory to speed up
+            putting data into inter-process queue, set :attr:`use_shared_memory`
+            as True only when the shared memory space on your machine(e.g.
+            space of '/dev/shm' on Linux operating sysytem) is large enough.
+            Shared memory will only be enabled in multi-process mode(num_workers
+            > 0). Default True.
+        timeout(int): the timeout value for getting data form output queue
+            of subprocesses. Default 0.
+        worker_init_fn(callable): init function which will be called with
+            worker id on each subproces starting if not set as None. Default
+            None.
+
+    Returns:
+        DataLoader: an iterable object for data iterating
+
+    Examples:
+        
+        .. code-block:: python
+
+            import numpy as np
+            import paddle.fluid as fluid
+            from paddle.io import Dataset, BatchSampler, DataLoader
+
+            BATCH_NUM = 20
+            BATCH_SIZE = 16
+            EPOCH_NUM = 4
+
+            IMAGE_SIZE = 784
+            CLASS_NUM = 10
+
+            USE_GPU = False # whether use GPU to run model
+
+            # define a random dataset
+            class RandomDataset(Dataset):
+                def __init__(self, num_samples):
+                    self.num_samples = num_samples
+
+                def __getitem__(self, idx):
+                    image = np.random.random([IMAGE_SIZE]).astype('float32')
+                    label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
+                    return image, label
+
+                def __len__(self):
+                    return self.num_samples
+
+            # get places
+            places = fluid.cuda_places() if USE_GPU else fluid.cpu_places()
+
+            # -------------------- static graph ---------------------
+
+            def simple_net(image, label):
+                fc_tmp = fluid.layers.fc(image, size=CLASS_NUM, act='softmax')
+                cross_entropy = fluid.layers.softmax_with_cross_entropy(image, label)
+                loss = fluid.layers.reduce_mean(cross_entropy)
+                sgd = fluid.optimizer.SGD(learning_rate=1e-3)
+                sgd.minimize(loss)
+                return loss
+
+            image = fluid.data(name='image', shape=[None, IMAGE_SIZE], dtype='float32')
+            label = fluid.data(name='label', shape=[None, 1], dtype='int64')
+
+            loss = simple_net(image, label)
+
+            exe = fluid.Executor(places[0])
+            exe.run(fluid.default_startup_program())
+
+            prog = fluid.CompiledProgram(fluid.default_main_program()).with_data_parallel(loss_name=loss.name)
+
+            dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
+
+            loader = DataLoader(dataset,
+                                feed_list=[image, label],
+                                places=places,
+                                batch_size=BATCH_SIZE, 
+                                shuffle=True,
+                                drop_last=True,
+                                num_workers=2)
+
+            for e in range(EPOCH_NUM):
+                for i, data in enumerate(loader()):
+                    l = exe.run(prog, feed=data, fetch_list=[loss], return_numpy=True)
+                    print("Epoch {} batch {}: loss = {}".format(e, i, l[0][0]))
+
+            # -------------------------------------------------------
+                
+            # --------------------- dygraph mode --------------------
+
+            class SimpleNet(fluid.dygraph.Layer):
+                def __init__(self):
+                    super(SimpleNet, self).__init__()
+                    self.fc = fluid.dygraph.nn.Linear(IMAGE_SIZE, CLASS_NUM, act='softmax')
+
+                def forward(self, image, label=None):
+                    return self.fc(image)
+
+            with fluid.dygraph.guard(places[0]):
+                simple_net = SimpleNet()
+                opt = fluid.optimizer.SGD(learning_rate=1e-3,
+                                          parameter_list=simple_net.parameters())
+
+                loader = DataLoader(dataset,
+                                    places=places[0],
+                                    batch_size=BATCH_SIZE,
+                                    shuffle=True,
+                                    drop_last=True,
+                                    num_workers=2)
+
+                for e in range(EPOCH_NUM):
+                    for i, (image, label) in enumerate(loader()):
+                        out = simple_net(image)
+                        loss = fluid.layers.cross_entropy(out, label)
+                        avg_loss = fluid.layers.reduce_mean(loss)
+                        avg_loss.backward()
+                        opt.minimize(avg_loss)
+                        simple_net.clear_gradients()
+                        print("Epoch {} batch {}: loss = {}".format(e, i, np.mean(loss.numpy())))
+
+            # -------------------------------------------------------
+
+    """
+
+    def __init__(self,
+                 dataset,
+                 feed_list=None,
+                 places=None,
+                 return_list=False,
+                 batch_sampler=None,
+                 batch_size=1,
+                 shuffle=False,
+                 drop_last=False,
+                 collate_fn=None,
+                 num_workers=0,
+                 use_buffer_reader=True,
+                 use_shared_memory=True,
+                 timeout=0,
+                 worker_init_fn=None):
+        self.return_list = return_list
+        self.collate_fn = collate_fn
+        self.use_buffer_reader = use_buffer_reader
+        self.worker_init_fn = worker_init_fn
+
+        assert isinstance(dataset, Dataset), \
+            "dataset should be subclass instance of paddle.io.Dataset"
+        self.dataset = dataset
+
+        if not return_list and not in_dygraph_mode():
+            assert feed_list is not None, \
+                    "feed_list should be set when return_list=False"
+        self.feed_list = feed_list
+
+        assert places is not None, "places cannot be None"
+        self.places = _convert_places(places)
+        if in_dygraph_mode():
+            assert len(self.places) == 1, \
+                    "Number of places must be 1 in dygraph mode"
+
+        assert num_workers >= 0, "num_workers should be a non-negative value"
+        if num_workers > 0 and (sys.platform == 'darwin' or
+                                sys.platform == 'win32'):
+            logging.warning(
+                "multi-process mode not support MacOs and Windows currently." \
+                " use signle-process with num_workers = 0 instead")
+            num_workers = 0
+        self.num_workers = num_workers
+
+        self.use_shared_memory = use_shared_memory
+        if use_shared_memory and num_workers == 0:
+            self.use_shared_memory = False
+
+        assert timeout >= 0, "timeout should be a non-negative value"
+        self.timeout = timeout
+
+        if batch_sampler is not None:
+            assert isinstance(batch_sampler, BatchSampler), \
+                "batch_sampler should be None or subclass instance " \
+                "of paddle.io.BatchSampler"
+            assert batch_size == 1 and not shuffle and not drop_last, \
+                "batch_size/shuffle/drop_last should not be set when " \
+                "batch_sampler is given"
+            self.batch_sampler = batch_sampler
+        else:
+            assert batch_size is not None and batch_size > 0, \
+                "batch_size should be a positive value when " \
+                "batch_sampler is not given"
+            self.batch_sampler = BatchSampler(
+                dataset=dataset,
+                batch_size=batch_size,
+                shuffle=shuffle,
+                drop_last=drop_last)
+
+    def __len__(self):
+        return len(self.batch_sampler)
+
+    def __iter__(self):
+        if self.num_workers == 0:
+            return _DataLoaderIterSingleProcess(self)
+        else:
+            return _DataLoaderIterMultiProcess(self)
+
+    def __call__(self):
+        return self.__iter__()
+
     @staticmethod
     def from_generator(feed_list=None,
                        capacity=None,
@@ -553,22 +732,7 @@ class DygraphGeneratorLoader(DataLoaderBase):
         if process is not None:
             process.join()
             # erase process id
-            core._erase_process_pid(id(self))
-
-    def _set_child_signal_handler(self):
-        core._set_process_pid(id(self), self._process.pid)
-        current_handler = signal.getsignal(signal.SIGCHLD)
-        if not callable(current_handler):
-            current_handler = None
-
-        def __handler__(signum, frame):
-            # NOTE: Here the signum is SIGDHLD, when the child process exits, this handler
-            # will be called whenever the child process exits normally or abnormally.
-            core._throw_error_if_process_failed()
-            if current_handler is not None:
-                current_handler(signum, frame)
-
-        signal.signal(signal.SIGCHLD, __handler__)
+            core._erase_process_pids(id(self))
 
     def _init_iterable(self):
         self._wait_thread_ends()
@@ -605,7 +769,8 @@ class DygraphGeneratorLoader(DataLoaderBase):
             # with SIGSEGV and SIGBUS of child process; 2. if the main process end before child
             # process, it shuts the all its daemonic children down with a SIGTERM (instead of 
             # joining them without a timeout), so here nedd to deal with SIGTERM.
-            self._set_child_signal_handler()
+            core._set_process_pids(id(self), [self._process.pid])
+            _set_SIGCHLD_handler()
 
             # Set reader_thread
             self._thread_done_event = threading.Event()
@@ -666,16 +831,11 @@ class DygraphGeneratorLoader(DataLoaderBase):
             # set signal handler
             core._set_process_signal_handler()
 
-            # child process clear function at exit
-            def _cleanup():
-                # clear memory map files in child process
-                core._cleanup_mmap_fds()
-
             # NOTE: [ mmap files clear ] When the child process exits unexpectedly,
             # some shared memory objects may have been applied for but have not yet
             # been put into the inter-process Queue. This part of the object needs
             # to be cleaned up when the process ends.
-            CleanupFuncRegistrar.register(_cleanup)
+            CleanupFuncRegistrar.register(_cleanup_mmap)
 
             for batch in self._batch_reader():
                 tensor_list = core._convert_to_tensor_list(batch)

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -211,6 +211,8 @@ if (APPLE OR WIN32)
   list(REMOVE_ITEM TEST_OPS test_imperative_data_loader_fds_clear)
   list(REMOVE_ITEM TEST_OPS test_imperative_data_loader_exit_func)
   list(REMOVE_ITEM TEST_OPS test_imperative_signal_handler)
+  list(REMOVE_ITEM TEST_OPS test_multiprocess_dataloader_base)
+  list(REMOVE_ITEM TEST_OPS test_multiprocess_dataloader_exception)
 endif()
 
 if(NOT WITH_GPU OR WIN32 OR APPLE)
@@ -381,4 +383,6 @@ if(NOT WIN32 AND NOT APPLE)
     set_tests_properties(test_imperative_data_loader_base PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" RUN_SERIAL TRUE)
     set_tests_properties(test_imperative_data_loader_fds_clear PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" RUN_SERIAL TRUE)
     # set_tests_properties(test_imperative_data_loader_exception PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" RUN_SERIAL TRUE)
+    set_tests_properties(test_multiprocess_dataloader_base PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" RUN_SERIAL TRUE)
+    set_tests_properties(test_multiprocess_dataloader_exception PROPERTIES LABELS "RUN_TYPE=EXCLUSIVE" RUN_SERIAL TRUE)
 endif()

python/paddle/fluid/tests/unittests/test_batch_sampler.py

+# Copyright (c) 2020 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.
+
+from __future__ import division
+
+import unittest
+
+import paddle.fluid as fluid
+from paddle.io import BatchSampler, Dataset
+
+
+class RandomDataset(Dataset):
+    def __init__(self, sample_num, class_num):
+        self.sample_num = sample_num
+        self.class_num = class_num
+
+    def __getitem__(self, idx):
+        np.random.seed(idx)
+        image = np.random.random([IMAGE_SIZE]).astype('float32')
+        label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
+        return image, label
+
+    def __len__(self):
+        return self.sample_num
+
+
+class TestBatchSampler(unittest.TestCase):
+    def setUp(self):
+        self.num_samples = 1000
+        self.num_classes = 10
+        self.batch_size = 32
+        self.shuffle = False
+        self.drop_last = False
+
+    def init_batch_sampler(self):
+        dataset = RandomDataset(self.num_samples, self.num_classes)
+        bs = BatchSampler(
+            dataset=dataset,
+            batch_size=self.batch_size,
+            shuffle=self.shuffle,
+            drop_last=self.drop_last)
+        return bs
+
+    def test_main(self):
+        bs = self.init_batch_sampler()
+        # length check
+        bs_len = (self.num_samples + int(not self.drop_last) \
+                * (self.batch_size - 1)) // self.batch_size
+        self.assertTrue(bs_len == len(bs))
+
+        # output indices check
+        if not self.shuffle:
+            index = 0
+            for indices in bs:
+                for idx in indices:
+                    self.assertTrue(index == idx)
+                    index += 1
+
+
+class TestBatchSamplerDropLast(TestBatchSampler):
+    def setUp(self):
+        self.num_samples = 1000
+        self.num_classes = 10
+        self.batch_size = 32
+        self.shuffle = False
+        self.drop_last = True
+
+
+class TestBatchSamplerShuffle(TestBatchSampler):
+    def setUp(self):
+        self.num_samples = 1000
+        self.num_classes = 10
+        self.batch_size = 32
+        self.shuffle = True
+        self.drop_last = True
+
+
+class TestBatchSamplerWithIndices(TestBatchSampler):
+    def init_batch_sampler(self):
+        bs = BatchSampler(
+            indices=list(range(self.num_samples)),
+            batch_size=self.batch_size,
+            drop_last=self.drop_last)
+        return bs
+
+
+class TestBatchSamplerWithIndicesAndDataSource(unittest.TestCase):
+    def setUp(self):
+        self.num_samples = 1000
+        self.num_classes = 10
+        self.batch_size = 32
+        self.shuffle = False
+        self.drop_last = True
+
+    def test_main(self):
+        try:
+            dataset = RandomDataset(self.num_samples, self.num_classes)
+            bs = BatchSampler(
+                dataset=dataset,
+                indices=list(range(self.num_samples)),
+                batch_size=self.batch_size,
+                drop_last=self.drop_last)
+            self.assertTrue(False)
+        except AssertionError:
+            pass
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_dataloader_dataset.py

+# Copyright (c) 2020 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.
+
+from __future__ import division
+
+import unittest
+import numpy as np
+
+import paddle.fluid as fluid
+from paddle.io import *
+
+
+class TestDatasetAbstract(unittest.TestCase):
+    def test_main(self):
+        dataset = Dataset()
+        try:
+            d = dataset[0]
+            self.assertTrue(False)
+        except NotImplementedError:
+            pass
+
+        try:
+            l = len(dataset)
+            self.assertTrue(False)
+        except NotImplementedError:
+            pass
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_imperative_data_loader_fds_clear.py

@@ -17,6 +17,7 @@ import unittest
 import numpy as np
 import paddle.fluid as fluid
 from paddle.fluid import core
+from paddle.io import Dataset, DataLoader
 
 
 def get_random_images_and_labels(image_shape, label_shape):
@@ -35,6 +36,20 @@ def batch_generator_creator(batch_size, batch_num):
     return __reader__
 
 
+class RandomDataset(Dataset):
+    def __init__(self, sample_num):
+        self.sample_num = sample_num
+
+    def __getitem__(self, idx):
+        np.random.seed(idx)
+        image = np.random.random([784]).astype('float32')
+        label = np.random.randint(0, 9, (1, )).astype('int64')
+        return image, label
+
+    def __len__(self):
+        return self.sample_num
+
+
 class TestDygraphDataLoaderMmapFdsClear(unittest.TestCase):
     def setUp(self):
         self.batch_size = 8
@@ -74,5 +89,19 @@ class TestDygraphDataLoaderMmapFdsClear(unittest.TestCase):
                 self.run_one_epoch_with_break(loader)
 
 
+class TestMultiProcessDataLoaderMmapFdsClear(TestDygraphDataLoaderMmapFdsClear):
+    def prepare_data_loader(self):
+        place = fluid.CPUPlace()
+        with fluid.dygraph.guard(place):
+            dataset = RandomDataset(self.batch_size * self.batch_num)
+            loader = DataLoader(
+                dataset,
+                places=place,
+                batch_size=self.batch_size,
+                drop_last=True,
+                num_workers=2)
+            return loader
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_imperative_signal_handler.py

@@ -24,7 +24,7 @@ from paddle.fluid import core
 
 
 def set_child_signal_handler(self, child_pid):
-    core._set_process_pid(id(self), child_pid)
+    core._set_process_pids(id(self), tuple([child_pid]))
     current_handler = signal.getsignal(signal.SIGCHLD)
     if not callable(current_handler):
         current_handler = None

python/paddle/fluid/tests/unittests/test_multiprocess_dataloader_base.py

+# Copyright (c) 2020 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.
+
+from __future__ import division
+
+import os
+import sys
+import six
+import time
+import unittest
+import multiprocessing
+import numpy as np
+
+import paddle.fluid as fluid
+from paddle.io import Dataset, BatchSampler, DataLoader
+from paddle.fluid.dygraph.nn import Linear
+from paddle.fluid.dygraph.base import to_variable
+
+EPOCH_NUM = 5
+BATCH_SIZE = 16
+IMAGE_SIZE = 784
+SAMPLE_NUM = 400
+CLASS_NUM = 10
+
+
+class RandomDataset(Dataset):
+    def __init__(self, sample_num, class_num):
+        self.sample_num = sample_num
+        self.class_num = class_num
+
+    def __getitem__(self, idx):
+        np.random.seed(idx)
+        image = np.random.random([IMAGE_SIZE]).astype('float32')
+        label = np.random.randint(0, self.class_num - 1, (1, )).astype('int64')
+        return image, label
+
+    def __len__(self):
+        return self.sample_num
+
+
+def simple_fc_net_static():
+    startup_prog = fluid.Program()
+    main_prog = fluid.Program()
+    startup_prog.random_seed = 1
+    main_prog.random_seed = 1
+
+    with fluid.unique_name.guard():
+        with fluid.program_guard(main_prog, startup_prog):
+            image = fluid.data(
+                name='image', shape=[None, IMAGE_SIZE], dtype='float32')
+            label = fluid.data(name='label', shape=[None, 1], dtype='int64')
+            hidden = image
+            param_attr = fluid.ParamAttr(initializer=fluid.initializer.Constant(
+                value=0.8))
+            bias_attr = fluid.ParamAttr(initializer=fluid.initializer.Constant(
+                value=0.5))
+            for hidden_size in [10, 20, 30]:
+                hidden = fluid.layers.fc(hidden,
+                                         size=hidden_size,
+                                         act='tanh',
+                                         param_attr=param_attr,
+                                         bias_attr=bias_attr)
+
+            predict_label = fluid.layers.fc(hidden,
+                                            size=CLASS_NUM,
+                                            act='softmax',
+                                            param_attr=param_attr,
+                                            bias_attr=bias_attr)
+            loss = fluid.layers.reduce_mean(
+                fluid.layers.cross_entropy(
+                    input=predict_label, label=label))
+
+            optimizer = fluid.optimizer.Adam()
+            optimizer.minimize(loss)
+    return startup_prog, main_prog, image, label, loss
+
+
+class SimpleFCNet(fluid.dygraph.Layer):
+    def __init__(self):
+        super(SimpleFCNet, self).__init__()
+
+        param_attr = fluid.ParamAttr(initializer=fluid.initializer.Constant(
+            value=0.8))
+        bias_attr = fluid.ParamAttr(initializer=fluid.initializer.Constant(
+            value=0.5))
+        self._fcs = []
+        in_channel = IMAGE_SIZE
+        for hidden_size in [10, 20, 30]:
+            self._fcs.append(
+                Linear(
+                    in_channel,
+                    hidden_size,
+                    act='tanh',
+                    param_attr=param_attr,
+                    bias_attr=bias_attr))
+            in_channel = hidden_size
+        self._fcs.append(
+            Linear(
+                in_channel,
+                CLASS_NUM,
+                act='softmax',
+                param_attr=param_attr,
+                bias_attr=bias_attr))
+
+    def forward(self, image):
+        out = image
+        for fc in self._fcs:
+            out = fc(out)
+        return out
+
+
+class TestStaticDataLoader(unittest.TestCase):
+    def run_main(self, num_workers, places, with_data_parallel):
+        scope = fluid.Scope()
+        with fluid.scope_guard(scope):
+            startup_prog, main_prog, image, label, loss = simple_fc_net_static()
+
+            dataset = RandomDataset(SAMPLE_NUM, CLASS_NUM)
+            dataloader = DataLoader(
+                dataset,
+                feed_list=[image, label],
+                places=places,
+                num_workers=num_workers,
+                batch_size=BATCH_SIZE,
+                drop_last=True)
+            assert len(dataloader) == int(SAMPLE_NUM / BATCH_SIZE)
+
+            exe = fluid.Executor(place=places[0])
+            exe.run(startup_prog)
+
+            prog = fluid.CompiledProgram(main_prog)
+            if with_data_parallel:
+                prog = prog.with_data_parallel(
+                    loss_name=loss.name, places=places)
+
+            step_list = []
+            loss_list = []
+            start_t = time.time()
+            for _ in six.moves.range(EPOCH_NUM):
+                step = 0
+                for d in dataloader:
+                    assert len(d) == len(places), "{} != {}".format(
+                        len(d), len(places))
+                    for i, item in enumerate(d):
+                        image = item['image']
+                        label = item['label']
+                        assert image.shape() == [BATCH_SIZE, IMAGE_SIZE]
+                        assert label.shape() == [BATCH_SIZE, 1]
+                        assert image._place()._equals(places[i])
+                        assert label._place()._equals(places[i])
+                    L, = exe.run(program=prog,
+                                 feed=d,
+                                 fetch_list=[loss],
+                                 use_program_cache=True)
+                    loss_list.append(np.mean(L))
+                    step += 1
+                step_list.append(step)
+
+        end_t = time.time()
+        ret = {
+            "time": end_t - start_t,
+            "step": step_list,
+            "loss": np.array(loss_list)
+        }
+        print("time cost", ret['time'], 'step_list', ret['step'])
+        return ret
+
+    def prepare_places(self, with_data_parallel, with_cpu=True, with_gpu=True):
+        places = []
+        # FIXME: PR_CI_Py35 may hang on Multi-CPUs with multiprocess, but it
+        #        works fine locally, this should be fixed. OTOH, multiprocessing
+        #        is not recommended when running on CPU generally
+        if with_cpu and not sys.version.startswith('3.5'):
+            places.append([fluid.CPUPlace()])
+            if with_data_parallel:
+                places.append([fluid.CPUPlace()] * 2)
+
+        if with_gpu and fluid.core.is_compiled_with_cuda():
+            tmp = fluid.cuda_places()[:2]
+            assert len(tmp) > 0, "no gpu detected"
+            if with_data_parallel:
+                places.append(tmp)
+            places.append([tmp[0]])
+        return places
+
+    def test_main(self):
+        for with_data_parallel in [False] if self.__class__.__name__ \
+                == "TestDygraphDataLoader" else [True, False]:
+            for p in self.prepare_places(with_data_parallel):
+                results = []
+                for num_workers in [0, 2]:
+                    print(self.__class__.__name__, p, num_workers)
+                    ret = self.run_main(
+                        num_workers=num_workers,
+                        places=p,
+                        with_data_parallel=with_data_parallel)
+                    results.append(ret)
+                diff = np.max(
+                    np.abs(results[0]['loss'] - results[1]['loss']) /
+                    np.abs(results[0]['loss']))
+                self.assertLess(diff, 1e-2)
+
+
+class TestDygraphDataLoader(TestStaticDataLoader):
+    def run_main(self, num_workers, places, with_data_parallel):
+        fluid.default_startup_program().random_seed = 1
+        fluid.default_main_program().random_seed = 1
+        with fluid.dygraph.guard(places[0]):
+            fc_net = SimpleFCNet()
+            optimizer = fluid.optimizer.Adam(parameter_list=fc_net.parameters())
+
+            dataset = RandomDataset(SAMPLE_NUM, CLASS_NUM)
+            dataloader = DataLoader(
+                dataset,
+                places=places,
+                num_workers=num_workers,
+                batch_size=BATCH_SIZE,
+                drop_last=True)
+            assert len(dataloader) == int(SAMPLE_NUM / BATCH_SIZE)
+
+            step_list = []
+            loss_list = []
+            start_t = time.time()
+            for _ in six.moves.range(EPOCH_NUM):
+                step = 0
+                for image, label in dataloader():
+                    out = fc_net(image)
+                    loss = fluid.layers.cross_entropy(out, label)
+                    avg_loss = fluid.layers.reduce_mean(loss)
+                    avg_loss.backward()
+                    optimizer.minimize(avg_loss)
+                    fc_net.clear_gradients()
+
+                    loss_list.append(np.mean(avg_loss.numpy()))
+                    step += 1
+                step_list.append(step)
+
+        end_t = time.time()
+        ret = {
+            "time": end_t - start_t,
+            "step": step_list,
+            "loss": np.array(loss_list)
+        }
+        print("time cost", ret['time'], 'step_list', ret['step'])
+        return ret
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_multiprocess_dataloader_exception.py

+# Copyright (c) 2020 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.
+
+from __future__ import division
+
+import os
+import sys
+import six
+import time
+import unittest
+import multiprocessing
+import numpy as np
+
+import paddle.fluid as fluid
+from paddle.io import Dataset, BatchSampler, DataLoader
+from paddle.fluid.dygraph.nn import Linear
+from paddle.fluid.dygraph.base import to_variable
+
+
+class RandomDataset(Dataset):
+    def __init__(self, sample_num):
+        self.sample_num = sample_num
+
+    def __getitem__(self, idx):
+        np.random.seed(idx)
+        image = np.random.random([784]).astype('float32')
+        label = np.random.randint(0, 9, (1, )).astype('int64')
+        return image, label
+
+    def __len__(self):
+        return self.sample_num
+
+
+class TestDataLoaderAssert(unittest.TestCase):
+    def test_main(self):
+        place = fluid.cpu_places()[0]
+        with fluid.dygraph.guard(place):
+            dataset = RandomDataset(100)
+            batch_sampler = BatchSampler(dataset=dataset, batch_size=4)
+
+            # dataset is not instance of Dataset
+            try:
+                loader = DataLoader(dataset=batch_sampler, places=place)
+                self.assertTrue(False)
+            except AssertionError:
+                pass
+
+            # places is None
+            try:
+                loader = DataLoader(dataset=dataset, places=None)
+                self.assertTrue(False)
+            except AssertionError:
+                pass
+
+            # num_workers < 0
+            try:
+                loader = DataLoader(
+                    dataset=dataset, places=place, num_workers=-1)
+                self.assertTrue(False)
+            except AssertionError:
+                pass
+
+            # timeout < 0
+            try:
+                loader = DataLoader(dataset=dataset, places=place, timeout=-1)
+                self.assertTrue(False)
+            except AssertionError:
+                pass
+
+            # batch_sampler is not instance of BatchSampler
+            try:
+                loader = DataLoader(
+                    dataset=dataset, places=place, batch_sampler=dataset)
+                self.assertTrue(False)
+            except AssertionError:
+                pass
+
+            # set batch_sampler and shuffle/batch_size/drop_last
+            try:
+                loader = DataLoader(
+                    dataset=dataset,
+                    places=place,
+                    batch_sampler=batch_sampler,
+                    shuffle=True,
+                    drop_last=True)
+                self.assertTrue(False)
+            except AssertionError:
+                pass
+
+            # set batch_sampler correctly
+            try:
+                loader = DataLoader(
+                    dataset=dataset, places=place, batch_sampler=batch_sampler)
+                self.assertTrue(True)
+            except AssertionError:
+                self.assertTrue(False)
+
+
+# CI Converage cannot record stub in subprocess,
+# HACK a _worker_loop in main process call here
+class TestDataLoaderWorkerLoop(unittest.TestCase):
+    def run_without_worker_done(self, use_shared_memory=True):
+        try:
+            place = fluid.cpu_places()[0]
+            with fluid.dygraph.guard(place):
+                dataset = RandomDataset(800)
+
+                # test init_fn
+                def _init_fn(worker_id):
+                    pass
+
+                # test collate_fn
+                def _collate_fn(sample_list):
+                    return [
+                        np.stack(
+                            s, axis=0) for s in list(zip(*sample_list))
+                    ]
+
+                loader = DataLoader(
+                    dataset,
+                    num_workers=1,
+                    places=place,
+                    use_shared_memory=use_shared_memory)
+                assert loader.num_workers > 0, \
+                    "go to AssertionError and pass in Mac and Windows"
+                loader = iter(loader)
+                print("loader length", len(loader))
+                indices_queue = multiprocessing.Queue()
+                for i in range(10):
+                    indices_queue.put([i, i + 10])
+                indices_queue.put(None)
+                loader._worker_loop(
+                    loader._dataset, indices_queue, loader._data_queue,
+                    loader._workers_done_event, _collate_fn, _init_fn, 0)
+                self.assertTrue(False)
+        except AssertionError:
+            pass
+        except Exception:
+            self.assertTrue(False)
+
+    def run_with_worker_done(self, use_shared_memory=True):
+        try:
+            place = fluid.cpu_places()[0]
+            with fluid.dygraph.guard(place):
+                dataset = RandomDataset(800)
+
+                # test init_fn
+                def _init_fn(worker_id):
+                    pass
+
+                # test collate_fn
+                def _collate_fn(sample_list):
+                    return [
+                        np.stack(
+                            s, axis=0) for s in list(zip(*sample_list))
+                    ]
+
+                loader = DataLoader(
+                    dataset,
+                    num_workers=1,
+                    places=place,
+                    use_shared_memory=use_shared_memory)
+                assert loader.num_workers > 0, \
+                    "go to AssertionError and pass in Mac and Windows"
+                loader = iter(loader)
+                print("loader length", len(loader))
+                indices_queue = multiprocessing.Queue()
+                for i in range(10):
+                    indices_queue.put([i, i + 10])
+                indices_queue.put(None)
+                loader._workers_done_event.set()
+                loader._worker_loop(
+                    loader._dataset, indices_queue, loader._data_queue,
+                    loader._workers_done_event, _collate_fn, _init_fn, 0)
+                self.assertTrue(True)
+        except AssertionError:
+            pass
+        except Exception:
+            self.assertTrue(False)
+
+    def test_main(self):
+        for use_shared_memory in [True, False]:
+            self.run_without_worker_done(use_shared_memory)
+            self.run_with_worker_done(use_shared_memory)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/io/__init__.py

@@ -13,22 +13,27 @@
 # limitations under the License.
 
 # TODO: define all functions about input & output in this directory 
-# __all__ = ['Dataset',
-#            'Sampler',
-#            'Transform',
-#            'DataLoader',
-#            'load',
-#            'save',
-#            'load_program_state',
-#            'set_program_state',
-#            'load_inference_model',
-#            'save_inference_model',
-#            'batch',
-#            'shuffle',
-#            'buffered',
-#            'cache',
-#            'chain',
-#            'firstn',
-#            'compose',
-#            'map_readers',
-#            'xmap_readers']
+__all__ = [
+    'Dataset',
+    'BatchSampler',
+    #            'Transform',
+    'DataLoader',
+    #            'load',
+    #            'save',
+    #            'load_program_state',
+    #            'set_program_state',
+    #            'load_inference_model',
+    #            'save_inference_model',
+    #            'batch',
+    #            'shuffle',
+    #            'buffered',
+    #            'cache',
+    #            'chain',
+    #            'firstn',
+    #            'compose',
+    #            'map_readers',
+    #            'xmap_readers'
+]
+
+from ..fluid.io import DataLoader
+from ..fluid.dataloader import Dataset, BatchSampler

python/setup.py.in

@@ -149,6 +149,7 @@ packages=['paddle',
           'paddle.fluid.proto.profiler',
           'paddle.fluid.distributed',
           'paddle.fluid.layers',
+          'paddle.fluid.dataloader',
           'paddle.fluid.contrib',
           'paddle.fluid.contrib.decoder',
           'paddle.fluid.contrib.quantize',
@@ -176,6 +177,7 @@ packages=['paddle',
           'paddle.fluid.incubate.fleet.parameter_server.pslib',
           'paddle.fluid.incubate.fleet.collective',
           'paddle.fluid.incubate.fleet.utils',
+          'paddle.io',
           'paddle.nn',
           'paddle.nn.functional',
           'paddle.nn.layer',