Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix-7555

Author: pkuyym

doc/faq/local/index_cn.rst

@@ -211,3 +211,49 @@ decoder_inputs = paddle.layer.fc(
 * list 中元素的个数等于网络中输出层的个数；
 * list 中每个元素是一个layer的输出结果矩阵，类型是numpy的ndarray；
 * 每一个layer输出矩阵的高度，在非序列输入时：等于样本数；序列输入时等于：输入序列中元素的总数；宽度等于配置中layer的size；
+
+6.  如何在训练过程中获得某一个layer的output
+-----------------------------------------------
+
+可以在event_handler中，通过 :code:`event.gm.getLayerOutputs("layer_name")` 获得在模型配置中某一层的name :code:`layer_name` 在当前
+mini-batch forward的output的值。获得的值类型均为 :code:`numpy.ndarray` ，可以通过这个输出来完成自定义的评估指标计算等功能。例如下面代码：
+
+..      code-block:: python
+
+        def score_diff(right_score, left_score):
+            return np.average(np.abs(right_score - left_score))
+
+        def event_handler(event):
+            if isinstance(event, paddle.event.EndIteration):
+                if event.batch_id % 25 == 0:
+                    diff = score_diff(
+                        event.gm.getLayerOutputs("right_score")["right_score"][
+                            "value"],
+                        event.gm.getLayerOutputs("left_score")["left_score"][
+                            "value"])
+                    logger.info(("Pass %d Batch %d : Cost %.6f, "
+                                "average absolute diff scores: %.6f") %
+                                (event.pass_id, event.batch_id, event.cost, diff))
+
+注意：此方法不能获取 :code:`paddle.layer.recurrent_group` 里step的内容，但可以获取 :code:`paddle.layer.recurrent_group` 的输出。
+
+7.  如何在训练过程中获得参数的权重和梯度
+-----------------------------------------------
+
+在某些情况下，获得当前mini-batch的权重（或称作weights, parameters）有助于在训练时观察具体数值，方便排查以及快速定位问题。
+可以通过在 :code:`event_handler` 中打印其值（注意，需要使用 :code:`paddle.event.EndForwardBackward` 保证使用GPU训练时也可以获得），
+示例代码如下：
+
+..      code-block:: python
+
+        ...
+        parameters = paddle.parameters.create(cost)
+        ...
+        def event_handler(event):
+            if isinstance(event, paddle.event.EndForwardBackward):
+                if event.batch_id % 25 == 0:
+                    for p in parameters.keys():
+                        logger.info("Param %s, Grad %s",
+                            parameters.get(p), parameters.get_grad(p))
+
+注意：“在训练过程中获得某一个layer的output”和“在训练过程中获得参数的权重和梯度”都会造成训练中的数据从C++拷贝到numpy，会对训练性能造成影响。不要在注重性能的训练场景下使用。

paddle/framework/executor.cc

@@ -23,6 +23,7 @@ limitations under the License. */
 #include "paddle/framework/op_registry.h"
 #include "paddle/platform/place.h"
 
+DECLARE_bool(do_memory_benchmark);
 DEFINE_bool(check_nan_inf, false,
             "Checking whether operator produce NAN/INF or not. It will be "
             "extremely slow so please use this flag wisely.");
@@ -117,6 +118,10 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
     VLOG(3) << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);
+    if (FLAGS_do_memory_benchmark) {
+      VLOG(2) << "Memory used after operator " + op->Type() + " running: "
+              << memory::memory_usage(place_);
+    }
     if (FLAGS_check_nan_inf) {
       for (auto& vname : op->OutputVars(true)) {
         auto* var = local_scope->FindVar(vname);
@@ -130,6 +135,12 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
   if (create_vars && create_local_scope) {
     scope->DeleteScope(local_scope);
   }
+  if (FLAGS_do_memory_benchmark) {
+    VLOG(2) << "-------------------------------------------------------";
+    VLOG(2) << "Memory used after deleting local scope: "
+            << memory::memory_usage(place_);
+    VLOG(2) << "-------------------------------------------------------";
+  }
 }
 
 }  // namespace framework

paddle/framework/scope.cc

@@ -20,6 +20,10 @@ limitations under the License. */
 #include "paddle/framework/threadpool.h"
 #include "paddle/string/printf.h"
 
+DEFINE_bool(do_memory_benchmark, false,
+            "Doing memory benchmark. It will make deleting scope synchronized, "
+            "and add some memory usage logs");
+
 namespace paddle {
 namespace framework {
 
@@ -88,8 +92,12 @@ void Scope::DeleteScope(Scope* scope) {
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
   this->kids_.erase(it);
-  // Make delete async.
-  Async([scope] { delete scope; });
+  // When making memory benchmark on Fluid, we have to delete scope sync.
+  if (FLAGS_do_memory_benchmark) {
+    delete scope;
+  } else {
+    Async([scope] { delete scope; });
+  }
 }
 
 void Scope::Rename(const std::string& origin_name,

paddle/operators/compare_op.cc

@@ -39,6 +39,11 @@ N-dim tensor. X and Y could be any type.  The each element of the Out tensor is
 calculated by %s
 )DOC",
                                comment.type, comment.equation));
+    AddAttr<int>("axis",
+                 "(int, default -1). The start dimension index "
+                 "for broadcasting Y onto X.")
+        .SetDefault(-1)
+        .EqualGreaterThan(-1);
   }
 };
 
@@ -95,11 +100,5 @@ REGISTER_LOGICAL_OP(less_than, "Out = X < Y");
 REGISTER_LOGICAL_KERNEL(less_than, CPU, paddle::operators::LessThanFunctor);
 REGISTER_LOGICAL_OP(less_equal, "Out = X <= Y");
 REGISTER_LOGICAL_KERNEL(less_equal, CPU, paddle::operators::LessEqualFunctor);
-REGISTER_LOGICAL_OP(greater_than, "Out = X > Y");
-REGISTER_LOGICAL_KERNEL(greater_than, CPU,
-                        paddle::operators::GreaterThanFunctor);
-REGISTER_LOGICAL_OP(greater_equal, "Out = X >= Y");
-REGISTER_LOGICAL_KERNEL(greater_equal, CPU,
-                        paddle::operators::GreaterEqualFunctor);
 REGISTER_LOGICAL_OP(equal, "Out = X == Y");
 REGISTER_LOGICAL_KERNEL(equal, CPU, paddle::operators::EqualFunctor);

paddle/operators/compare_op.cu

@@ -16,8 +16,4 @@ limitations under the License. */
 
 REGISTER_LOGICAL_KERNEL(less_than, CUDA, paddle::operators::LessThanFunctor);
 REGISTER_LOGICAL_KERNEL(less_equal, CUDA, paddle::operators::LessEqualFunctor);
-REGISTER_LOGICAL_KERNEL(greater_than, CUDA,
-                        paddle::operators::GreaterThanFunctor);
-REGISTER_LOGICAL_KERNEL(greater_equal, CUDA,
-                        paddle::operators::GreaterEqualFunctor);
 REGISTER_LOGICAL_KERNEL(equal, CUDA, paddle::operators::EqualFunctor);

paddle/operators/compare_op.h

@@ -16,6 +16,7 @@ limitations under the License. */
 #include <math.h>
 #include <type_traits>
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/elementwise_op_function.h"
 #include "paddle/platform/transform.h"
 
 namespace paddle {
@@ -33,18 +34,6 @@ struct LessEqualFunctor {
   HOSTDEVICE bool operator()(const T& a, const T& b) const { return a <= b; }
 };
 
-template <typename T>
-struct GreaterThanFunctor {
-  using ELEM_TYPE = T;
-  HOSTDEVICE bool operator()(const T& a, const T& b) const { return a > b; }
-};
-
-template <typename T>
-struct GreaterEqualFunctor {
-  using ELEM_TYPE = T;
-  HOSTDEVICE bool operator()(const T& a, const T& b) const { return a >= b; }
-};
-
 template <typename T>
 struct EqualFunctor {
   using ELEM_TYPE = T;
@@ -65,14 +54,7 @@ class CompareOpKernel
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     using T = typename Functor::ELEM_TYPE;
-    auto* x = context.Input<framework::Tensor>("X");
-    auto* y = context.Input<framework::Tensor>("Y");
-    auto* out = context.Output<framework::Tensor>("Out");
-    Functor binary_func;
-    platform::Transform<DeviceContext> trans;
-    trans(context.template device_context<DeviceContext>(), x->data<T>(),
-          x->data<T>() + x->numel(), y->data<T>(),
-          out->mutable_data<bool>(context.GetPlace()), binary_func);
+    ElementwiseComputeEx<Functor, DeviceContext, T, bool>(context);
   }
 };
 

paddle/operators/elementwise_op_function.h

@@ -176,14 +176,15 @@ class MidWiseTransformIterator<T, platform::CUDADeviceContext>
 };
 #endif
 
-template <typename Functor, typename T, typename DeviceContext>
+template <typename Functor, typename T, typename DeviceContext,
+          typename OutType = T>
 class TransformFunctor {
  public:
   TransformFunctor(const framework::Tensor* x, const framework::Tensor* y,
                    framework::Tensor* z, const DeviceContext& ctx, Functor func)
       : x_(x->data<T>()),
         y_(y->data<T>()),
-        z_(z->mutable_data<T>(ctx.GetPlace())),
+        z_(z->mutable_data<OutType>(ctx.GetPlace())),
         nx_(x->numel()),
         ctx_(ctx),
         func_(func) {}
@@ -208,7 +209,7 @@ class TransformFunctor {
  private:
   const T* x_;
   const T* y_;
-  T* z_;
+  OutType* z_;
   int64_t nx_;
   const DeviceContext& ctx_;
   Functor func_;
@@ -364,15 +365,16 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
   }
 }
 
-template <typename Functor, typename DeviceContext, typename T>
+template <typename Functor, typename DeviceContext, typename T,
+          typename OutType = T>
 void ElementwiseComputeEx(const framework::ExecutionContext& ctx) {
   using Tensor = framework::Tensor;
 
   auto* x = ctx.Input<Tensor>("X");
   auto* y = ctx.Input<Tensor>("Y");
   auto* z = ctx.Output<Tensor>("Out");
-  z->mutable_data<T>(ctx.GetPlace());
-  TransformFunctor<Functor, T, DeviceContext> functor(
+  z->mutable_data<OutType>(ctx.GetPlace());
+  TransformFunctor<Functor, T, DeviceContext, OutType> functor(
       x, y, z, ctx.template device_context<DeviceContext>(), Functor());
 
   auto x_dims = x->dims();

python/paddle/v2/dataset/__init__.py

@@ -24,11 +24,23 @@
 import uci_housing
 import sentiment
 import wmt14
+import wmt16
 import mq2007
 import flowers
 import voc2012
 
 __all__ = [
-    'mnist', 'imikolov', 'imdb', 'cifar', 'movielens', 'conll05', 'sentiment'
-    'uci_housing', 'wmt14', 'mq2007', 'flowers', 'voc2012'
+    'mnist',
+    'imikolov',
+    'imdb',
+    'cifar',
+    'movielens',
+    'conll05',
+    'sentiment'
+    'uci_housing',
+    'wmt14',
+    'wmt16',
+    'mq2007',
+    'flowers',
+    'voc2012',
 ]

python/paddle/v2/dataset/common.py

@@ -25,8 +25,12 @@
 import cPickle as pickle
 
 __all__ = [
-    'DATA_HOME', 'download', 'md5file', 'split', 'cluster_files_reader',
-    'convert'
+    'DATA_HOME',
+    'download',
+    'md5file',
+    'split',
+    'cluster_files_reader',
+    'convert',
 ]
 
 DATA_HOME = os.path.expanduser('~/.cache/paddle/dataset')
@@ -58,12 +62,15 @@ def md5file(fname):
     return hash_md5.hexdigest()
 
 
-def download(url, module_name, md5sum):
+def download(url, module_name, md5sum, save_name=None):
     dirname = os.path.join(DATA_HOME, module_name)
     if not os.path.exists(dirname):
         os.makedirs(dirname)
 
-    filename = os.path.join(dirname, url.split('/')[-1])
+    filename = os.path.join(dirname,
+                            url.split('/')[-1]
+                            if save_name is None else save_name)
+
     retry = 0
     retry_limit = 3
     while not (os.path.exists(filename) and md5file(filename) == md5sum):
@@ -196,9 +203,11 @@ def convert(output_path, reader, line_count, name_prefix):
     Convert data from reader to recordio format files.
 
     :param output_path: directory in which output files will be saved.
-    :param reader: a data reader, from which the convert program will read data instances.
+    :param reader: a data reader, from which the convert program will read
+                   data instances.
     :param name_prefix: the name prefix of generated files.
-    :param max_lines_to_shuffle: the max lines numbers to shuffle before writing.
+    :param max_lines_to_shuffle: the max lines numbers to shuffle before
+                                 writing.
     """
 
     assert line_count >= 1

python/paddle/v2/dataset/tests/wmt16_test.py

@@ -0,0 +1,66 @@
+# Copyright (c) 2016 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 paddle.v2.dataset.wmt16
+import unittest
+
+
+class TestWMT16(unittest.TestCase):
+    def checkout_one_sample(self, sample):
+        # train data has 3 field: source language word indices,
+        # target language word indices, and target next word indices.
+        self.assertEqual(len(sample), 3)
+
+        # test start mark and end mark in source word indices.
+        self.assertEqual(sample[0][0], 0)
+        self.assertEqual(sample[0][-1], 1)
+
+        # test start mask in target word indices
+        self.assertEqual(sample[1][0], 0)
+
+        # test en mask in target next word indices
+        self.assertEqual(sample[2][-1], 1)
+
+    def test_train(self):
+        for idx, sample in enumerate(
+                paddle.v2.dataset.wmt16.train(
+                    src_dict_size=100000, trg_dict_size=100000)()):
+            if idx >= 10: break
+            self.checkout_one_sample(sample)
+
+    def test_test(self):
+        for idx, sample in enumerate(
+                paddle.v2.dataset.wmt16.test(
+                    src_dict_size=1000, trg_dict_size=1000)()):
+            if idx >= 10: break
+            self.checkout_one_sample(sample)
+
+    def test_val(self):
+        for idx, sample in enumerate(
+                paddle.v2.dataset.wmt16.validation(
+                    src_dict_size=1000, trg_dict_size=1000)()):
+            if idx >= 10: break
+            self.checkout_one_sample(sample)
+
+    def test_get_dict(self):
+        dict_size = 1000
+        word_dict = paddle.v2.dataset.wmt16.get_dict("en", dict_size, True)
+        self.assertEqual(len(word_dict), dict_size)
+        self.assertEqual(word_dict[0], "<s>")
+        self.assertEqual(word_dict[1], "<e>")
+        self.assertEqual(word_dict[2], "<unk>")
+
+
+if __name__ == "__main__":
+    unittest.main()