Merge remote-tracking branch 'baidu/develop' into feature/add_v2_api_doc

demo/seqToseq/api_train_v2.py

@@ -126,51 +126,57 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
 
 def main():
     paddle.init(use_gpu=False, trainer_count=1)
+    is_generating = True
 
     # source and target dict dim.
     dict_size = 30000
     source_dict_dim = target_dict_dim = dict_size
 
-    # define network topology
-    cost = seqToseq_net(source_dict_dim, target_dict_dim)
-    parameters = paddle.parameters.create(cost)
-
-    # define optimize method and trainer
-    optimizer = paddle.optimizer.Adam(
-        learning_rate=5e-5,
-        regularization=paddle.optimizer.L2Regularization(rate=1e-3))
-    trainer = paddle.trainer.SGD(cost=cost,
-                                 parameters=parameters,
-                                 update_equation=optimizer)
-
-    # define data reader
-    feeding = {
-        'source_language_word': 0,
-        'target_language_word': 1,
-        'target_language_next_word': 2
-    }
-
-    wmt14_reader = paddle.batch(
-        paddle.reader.shuffle(
-            paddle.dataset.wmt14.train(dict_size=dict_size), buf_size=8192),
-        batch_size=5)
-
-    # define event_handler callback
-    def event_handler(event):
-        if isinstance(event, paddle.event.EndIteration):
-            if event.batch_id % 10 == 0:
-                print "\nPass %d, Batch %d, Cost %f, %s" % (
-                    event.pass_id, event.batch_id, event.cost, event.metrics)
-            else:
-                sys.stdout.write('.')
-                sys.stdout.flush()
-
-    # start to train
-    trainer.train(
-        reader=wmt14_reader,
-        event_handler=event_handler,
-        num_passes=10000,
-        feeding=feeding)
+    # train the network
+    if not is_generating:
+        cost = seqToseq_net(source_dict_dim, target_dict_dim)
+        parameters = paddle.parameters.create(cost)
+
+        # define optimize method and trainer
+        optimizer = paddle.optimizer.Adam(
+            learning_rate=5e-5,
+            regularization=paddle.optimizer.L2Regularization(rate=8e-4))
+        trainer = paddle.trainer.SGD(cost=cost,
+                                     parameters=parameters,
+                                     update_equation=optimizer)
+        # define data reader
+        wmt14_reader = paddle.batch(
+            paddle.reader.shuffle(
+                paddle.dataset.wmt14.train(dict_size), buf_size=8192),
+            batch_size=5)
+
+        # define event_handler callback
+        def event_handler(event):
+            if isinstance(event, paddle.event.EndIteration):
+                if event.batch_id % 10 == 0:
+                    print "\nPass %d, Batch %d, Cost %f, %s" % (
+                        event.pass_id, event.batch_id, event.cost,
+                        event.metrics)
+                else:
+                    sys.stdout.write('.')
+                    sys.stdout.flush()
+
+        # start to train
+        trainer.train(
+            reader=wmt14_reader, event_handler=event_handler, num_passes=2)
+
+    # generate a english sequence to french
+    else:
+        gen_creator = paddle.dataset.wmt14.test(dict_size)
+        gen_data = []
+        for item in gen_creator():
+            gen_data.append((item[0], ))
+            if len(gen_data) == 3:
+                break
+
+        beam_gen = seqToseq_net(source_dict_dim, target_dict_dim, is_generating)
+        parameters = paddle.dataset.wmt14.model()
+        trg_dict = paddle.dataset.wmt14.trg_dict(dict_size)
 
 
 if __name__ == '__main__':

paddle/gserver/layers/SequenceLastInstanceLayer.cpp

@@ -25,6 +25,11 @@ namespace paddle {
  * Input: a sequence
  * If SequenceLevel = kNonseq:
  *   Output: a sequence containing only the last instance of the input sequence
+ *   If stride_ > 0:
+ *      Output: a shorten sequence. The operation of getting last instance of a
+ *              sequence is independently performed on every slice of the input
+ *              sequence, which is obtained by sliding a window with the window
+ *              size set to stride_.
  * If SequenceLevel = kSeq:
  *   Check input sequence must has sub-sequence
  *   Output: a sequence containing only the last instance of each sub-sequence
@@ -37,6 +42,7 @@ class SequenceLastInstanceLayer : public SequencePoolLayer {
 protected:
   MatrixPtr tmpSrc_;
   MatrixPtr tmpDest_;
+  std::vector<int> instanceIds_;
 
 public:
   explicit SequenceLastInstanceLayer(const LayerConfig& config)
@@ -54,6 +60,7 @@ REGISTER_LAYER(seqlastins, SequenceLastInstanceLayer);
 bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
                                      const ParameterMap& parameterMap) {
   SequencePoolLayer::init(layerMap, parameterMap);
+  reversed_ = config_.select_first();
 
   tmpSrc_ =
       Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_);
@@ -66,7 +73,8 @@ bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
 void SequenceLastInstanceLayer::forward(PassType passType) {
   SequencePoolLayer::forward(passType);
 
-  const int* starts = startPositions_->getData(false);
+  auto starts = (stride_ > 0) ? stridePositions_->getData()
+                              : startPositions_->getData(false);
   MatrixPtr inputValue = getInputValue(0);
   MatrixPtr outputValue = getOutputValue();
 
@@ -74,9 +82,10 @@ void SequenceLastInstanceLayer::forward(PassType passType) {
     AsyncGpuBlock asyncGpuBlock;
     REGISTER_TIMER_INFO("SequenceLastInstanceLayerForward", getName().c_str());
 
+    instanceIds_.clear();
     for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) {
-      int insId =
-          config_.select_first() ? starts[seqId] : starts[seqId + 1] - 1;
+      int insId = reversed_ ? starts[seqId] : starts[seqId + 1] - 1;
+      instanceIds_.push_back(insId);
 
       outputValue->subMatrix(seqId, 1, tmpDest_)
           ->assign(*(inputValue->subMatrix(insId, 1, tmpSrc_)));
@@ -96,18 +105,13 @@ void SequenceLastInstanceLayer::backward(const UpdateCallback& callback) {
 
   MatrixPtr inputGrad = getInputGrad(0);
   MatrixPtr outputGrad = getOutputGrad();
-  const int* starts = startPositions_->getData(false);
-  size_t numSequences = startPositions_->getSize() - 1;
 
   if (inputGrad) {
     AsyncGpuBlock asyncGpuBlock;
     REGISTER_TIMER_INFO("SequenceLastInstanceLayerBackward", getName().c_str());
 
-    for (size_t seqId = 0; seqId < numSequences; ++seqId) {
-      int insId =
-          config_.select_first() ? starts[seqId] : starts[seqId + 1] - 1;
-
-      inputGrad->subMatrix(insId, 1, tmpDest_)
+    for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) {
+      inputGrad->subMatrix(instanceIds_[seqId], 1, tmpDest_)
           ->add(*(outputGrad->subMatrix(seqId, 1, tmpSrc_)));
     }
   }

paddle/gserver/layers/SequencePoolLayer.cpp

@@ -37,6 +37,7 @@ bool SequencePoolLayer::init(const LayerMap& layerMap,
   } else {
     LOG(FATAL) << "Unknown trans_type: " << config_.trans_type();
   }
+  stride_ = config_.seq_pool_stride();
   setNeedSequenceInfo(false);
   return true;
 }
@@ -55,8 +56,6 @@ void SequencePoolLayer::forward(PassType passType) {
   CHECK_EQ(starts->getData()[newBatchSize_], input.getBatchSize());
   CHECK_EQ(newBatchSize_, starts->getSize() - 1);
 
-  resetOutput(newBatchSize_, dim);
-
   /* If type_ = kNonSeq, both seq has or not has sub-seq degrade to a non-seq,
    * thus, in this case, output_ has no sequenceStartPositions.
    * If type_ = kSeq, seq has sub-seq degrades to a seq, thus, only in this
@@ -67,6 +66,15 @@ void SequencePoolLayer::forward(PassType passType) {
         << "when trans_type = seq, input must hasSubseq";
     output_.degradeSequence(input);
   }
+  if (stride_ > 0) {
+    CHECK_EQ(input.hasSubseq(), 0UL)
+        << "sequence stride pooling is invalid for hasSubseq now";
+    output_.poolSequenceWithStride(
+        input, stride_, &stridePositions_, reversed_);
+    newBatchSize_ = stridePositions_->getSize() - 1;
+  }
+
+  resetOutput(newBatchSize_, dim);
 }
 
 void SequencePoolLayer::backward(const UpdateCallback& callback) {

paddle/gserver/layers/SequencePoolLayer.h

@@ -26,6 +26,10 @@ namespace paddle {
  *    Output: output size is the number of input sequences (NOT input instances)
  *    output[i] = seqlastin/average/max_{for each instance in this
  * sequence}{input[i]}
+ *    If stride_ > 0:
+ *        Check input sequence must not have sub-sequence
+ *        Output: a shorten sequence, pooling is performed upon a small local
+ *                area
  * If SequenceLevel = kSeq:
  *    Check input sequence must has sub-sequence
  *    Output: output size is the number of input sub-sequences
@@ -42,6 +46,11 @@ protected:
   enum SequenceLevel { kNonSeq = 0, kSeq = 1 };
   size_t newBatchSize_;
   ICpuGpuVectorPtr startPositions_;
+  int stride_;
+  // Store the start position of each window.
+  IVectorPtr stridePositions_;
+  // Whether the input sequence is reversed or not.
+  bool reversed_ = false;
 
 public:
   explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {}

paddle/gserver/tests/test_LayerGrad.cpp

@@ -804,10 +804,14 @@ TEST(Layer, ExpandLayer) {
   testExpandLayer("seq", true);       // seq expand to hasSubseq
 }
 
-void testDegradeLayer(bool hasSubseq, string layer_type, string trans_type) {
+void testDegradeLayer(bool hasSubseq,
+                      string layer_type,
+                      string trans_type,
+                      int stride) {
   TestConfig config;
   config.layerConfig.set_type(layer_type);
   config.layerConfig.set_size(10);
+  config.layerConfig.set_seq_pool_stride(stride);
   config.biasSize = 0;
 
   config.inputDefs.push_back(
@@ -827,36 +831,46 @@ void testDegradeLayer(bool hasSubseq, string layer_type, string trans_type) {
   if (layer_type == "average") {
     for (auto strategy : {"average", "sum", "squarerootn"}) {
       LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type
-                << " average_strategy=" << strategy;
+                << " average_strategy=" << strategy
+                << " seq_pool_stride=" << stride;
       config.layerConfig.set_average_strategy(strategy);
       testDegradeLayerGrad(config, layer_type);
     }
   } else {
-    LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type;
+    LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type
+              << " seq_pool_stride=" << stride;
     testDegradeLayerGrad(config, layer_type);
   }
 }
 
 TEST(Layer, MaxLayer) {
-  testDegradeLayer(false, "max", "non-seq");  // seq max to non-seq
-  testDegradeLayer(true, "max", "non-seq");   // hasSubseq max to non-seq
-  testDegradeLayer(true, "max", "seq");       // hasSubseq max to seq
+  testDegradeLayer(false, "max", "non-seq", -1);  // seq max to non-seq
+  testDegradeLayer(true, "max", "non-seq", -1);   // hasSubseq max to non-seq
+  testDegradeLayer(true, "max", "seq", -1);       // hasSubseq max to seq
 }
 
 TEST(Layer, SequenceLastInstanceLayer) {
   testDegradeLayer(false,
                    "seqlastins",
-                   "non-seq");  // seq seqlastins to non-seq
+                   "non-seq",
+                   -1);  // seq seqlastins to non-seq
+  testDegradeLayer(false,
+                   "seqlastins",
+                   "non-seq",
+                   5);  // seq seqlastins to a shorten seq, stride window = 5
   testDegradeLayer(true,
                    "seqlastins",
-                   "non-seq");  // hasSubseq seqlastins to non-seq
-  testDegradeLayer(true, "seqlastins", "seq");  // hasSubseq seqlastins to seq
+                   "non-seq",
+                   -1);  // hasSubseq seqlastins to non-seq
+  testDegradeLayer(
+      true, "seqlastins", "seq", -1);  // hasSubseq seqlastins to seq
 }
 
 TEST(Layer, AverageLayer) {
-  testDegradeLayer(false, "average", "non-seq");  // seq average to non-seq
-  testDegradeLayer(true, "average", "non-seq");  // hasSubseq average to non-seq
-  testDegradeLayer(true, "average", "seq");      // hasSubseq average to seq
+  testDegradeLayer(false, "average", "non-seq", -1);  // seq average to non-seq
+  testDegradeLayer(
+      true, "average", "non-seq", -1);           // hasSubseq average to non-seq
+  testDegradeLayer(true, "average", "seq", -1);  // hasSubseq average to seq
 }
 
 TEST(Layer, SequenceConcatLayer) {

paddle/parameter/Argument.cpp

@@ -559,6 +559,49 @@ void Argument::degradeSequence(const Argument& input) {
   tgtBuf[numSequences] = numSubSequences;
 }
 
+void Argument::poolSequenceWithStride(const Argument& input,
+                                      size_t stride,
+                                      IVectorPtr* stridePostions,
+                                      bool reversed) {
+  // If input.sequenceStartPositions = [0, 9, 14, 17, 30] and stride = 5,
+  // then sequenceStartPositions = [0, 2, 3, 4, 7].
+  // If reversed = false, stridePostions = [0, 5, 9, 14, 17, 22, 27, 30];
+  // else reversed = true, stridePostions = [0, 4, 9, 14, 17, 20, 25, 30]
+
+  CHECK(input.sequenceStartPositions);
+  CHECK_EQ(input.hasSubseq(), 0UL);
+  CHECK_GT(stride, 0) << "stride must larger than 0";
+  size_t numSequences = input.getNumSequences();
+  ICpuGpuVector::resizeOrCreate(
+      sequenceStartPositions, numSequences + 1, false);
+  const int* starts = input.sequenceStartPositions->getData(false);
+  int* tgtBuf = sequenceStartPositions->getMutableData(false);
+  // first index of target sequence and stride positions are both 0
+  tgtBuf[0] = 0;
+  std::vector<int> stridePos;
+  for (size_t seqId = 0; seqId < numSequences; ++seqId) {
+    size_t seqLength = starts[seqId + 1] - starts[seqId];
+    stridePos.emplace_back(starts[seqId]);
+    if (seqLength == 0) {
+      // empty sequence
+      tgtBuf[seqId + 1] = tgtBuf[seqId];
+    } else {
+      int size = ceil((float)seqLength / stride);
+      tgtBuf[seqId + 1] = tgtBuf[seqId] + size;
+      for (int i = 0; i < size - 1; ++i) {
+        int cur = reversed ? starts[seqId + 1] - (size - 1 - i) * stride
+                           : stridePos.back() + stride;
+        stridePos.emplace_back(cur);
+      }
+    }
+  }
+  stridePos.emplace_back(starts[numSequences]);
+  int size = stridePos.size();
+  CHECK_EQ(size - 1, tgtBuf[numSequences]);
+  IVector::resizeOrCreate(*stridePostions, size, false);
+  (*stridePostions)->copyFrom(stridePos.data(), size);
+}
+
 void Argument::getValueString(
     std::unordered_map<std::string, std::string>* out) const {
   if (value) {

paddle/parameter/Argument.h

@@ -291,6 +291,15 @@ struct Argument {
    */
   void degradeSequence(const Argument& input);
 
+  /*
+   After pooling with stride n (n is smaller than sequence length),
+   a long sequence will be shorten.
+   This function is invalid for sequence having sub-sequence.
+   */
+  void poolSequenceWithStride(const Argument& input,
+                              size_t stride,
+                              IVectorPtr* stridePositions,
+                              bool reversed = false);
   /**
    * @brief getValueString will return the argument's output in string. There
    * are several kinds of output. The keys of output dictionary are 'value',

paddle/parameter/tests/CMakeLists.txt

 add_simple_unittest(test_common)
+add_simple_unittest(test_argument)

paddle/parameter/tests/test_argument.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 <gtest/gtest.h>
+#include <paddle/parameter/Argument.h>
+
+using namespace paddle;  // NOLINT
+
+TEST(Argument, poolSequenceWithStride) {
+  Argument input, output;
+  ICpuGpuVector::resizeOrCreate(input.sequenceStartPositions, 5, false);
+  int* inStart = input.sequenceStartPositions->getMutableData(false);
+  inStart[0] = 0;
+  inStart[1] = 9;
+  inStart[2] = 14;
+  inStart[3] = 17;
+  inStart[4] = 30;
+
+  int strideResult[] = {0, 5, 9, 14, 17, 22, 27, 30};
+  int strideResultReversed[] = {0, 4, 9, 14, 17, 20, 25, 30};
+
+  for (auto reversed : {false, true}) {
+    IVectorPtr stridePositions;
+    output.poolSequenceWithStride(
+        input, 5 /* stride */, &stridePositions, reversed);
+
+    const int* outStart = output.sequenceStartPositions->getData(false);
+    CHECK_EQ(outStart[0], 0);
+    CHECK_EQ(outStart[1], 2);
+    CHECK_EQ(outStart[2], 3);
+    CHECK_EQ(outStart[3], 4);
+    CHECK_EQ(outStart[4], 7);
+
+    CHECK_EQ(stridePositions->getSize(), 8);
+    auto result = reversed ? strideResultReversed : strideResult;
+    for (int i = 0; i < 8; i++) {
+      CHECK_EQ(stridePositions->getData()[i], result[i]);
+    }
+  }
+}
+
+int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
+  initMain(argc, argv);
+  return RUN_ALL_TESTS();
+}

paddle/py_paddle/dataprovider_converter.py

@@ -187,6 +187,13 @@ class SequenceScanner(IScanner):
         self.__inner_scanner__ = inner_scanner
         self.__setter__ = setter
 
+    def pre_scan(self, dat):
+        for each in dat:
+            self.__inner_scanner__.pre_scan(each)
+
+    def finish_pre_scan(self, argument):
+        self.__inner_scanner__.finish_pre_scan(argument)
+
     def scan(self, dat):
         self.__seq__.append(self.__seq__[-1] + self.get_size(dat))
         for each in dat:

paddle/py_paddle/util.py

@@ -83,13 +83,17 @@ def __arguments_to_numpy__(i, arg):
     assert isinstance(arg, swig_paddle.Arguments)
     value = arg.getSlotValue(i)
     ids = arg.getSlotIds(i)
+    prob = arg.getSlotIn(i)
     if value is not None:
         assert isinstance(value, swig_paddle.Matrix)
         value = value.copyToNumpyMat()
     if ids is not None:
         assert isinstance(ids, swig_paddle.IVector)
         ids = ids.copyToNumpyArray()
-    return {"value": value, "id": ids}
+    if prob is not None:
+        assert isinstance(prob, swig_paddle.Matrix)
+        prob = prob.copyToNumpyMat()
+    return {"value": value, "id": ids, "prob": prob}
 
 
 def __monkeypatch_gradient_machine__():

proto/ModelConfig.proto

@@ -441,6 +441,11 @@ message LayerConfig {
 
   // blank label used in ctc loss
   optional uint32 blank = 52 [default = 0];
+
+  // stride parameter for seqlastins layer, AverageLayer, MaxLayer, which 
+  // controls the scope of pooling operation. can be set > 0.
+  // leave empty or set to -1 to disable this stride pooling.
+  optional int32 seq_pool_stride = 53 [default = -1];
 }
 
 message EvaluatorConfig {

python/paddle/trainer/config_parser.py

@@ -2485,6 +2485,7 @@ class SequenceLastInstanceLayer(LayerBase):
                  active_type='linear',
                  trans_type='non-seq',
                  bias=False,
+                 stride=-1,
                  **xargs):
         super(SequenceLastInstanceLayer, self).__init__(
             name,
@@ -2495,10 +2496,11 @@ class SequenceLastInstanceLayer(LayerBase):
             **xargs)
         config_assert(
             len(inputs) == 1, 'SequenceLastInstanceLayer must have 1 input')
+        if trans_type == 'seq':
+            config_assert(stride == -1, 'subseq does not support stride window')
         self.config.trans_type = trans_type
-        for input_index in xrange(len(self.inputs)):
-            input_layer = self.get_input_layer(input_index)
-            self.set_layer_size(input_layer.size)
+        self.config.seq_pool_stride = stride
+        self.set_layer_size(self.get_input_layer(0).size)
         self.create_bias_parameter(bias, self.config.size)
 
 
@@ -2510,10 +2512,16 @@ class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
                  active_type='linear',
                  trans_type='non-seq',
                  bias=False,
+                 stride=-1,
                  **xargs):
         super(SequenceFirstInstanceLayer, self).__init__(
-            name, inputs=inputs, active_type=active_type, bias=bias, **xargs)
-        self.config.trans_type = trans_type
+            name,
+            inputs=inputs,
+            active_type=active_type,
+            trans_type=trans_type,
+            bias=bias,
+            stride=stride,
+            **xargs)
         self.config.select_first = True
 
 

python/paddle/trainer_config_helpers/layers.py

@@ -1342,10 +1342,16 @@ def grumemory(input,
 def last_seq(input,
              name=None,
              agg_level=AggregateLevel.EACH_TIMESTEP,
+             stride=-1,
              layer_attr=None):
     """
     Get Last Timestamp Activation of a sequence.
 
+    If stride > 0, this layer slides a window whose size is determined by stride, 
+    and return the last value of the window as the output. Thus, a long sequence 
+    will be shorten. Note that for sequence with sub-sequence, the default value 
+    of stride is -1.
+
     The simple usage is:
 
     .. code-block:: python
@@ -1357,6 +1363,8 @@ def last_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
+    :param stride: window size.  
+    :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.
@@ -1368,11 +1376,15 @@ def last_seq(input,
                        " series information at all. Maybe you want to use"
                        " first_seq instead.")
 
+    if agg_level == AggregateLevel.EACH_SEQUENCE:
+        assert stride == -1
+
     Layer(
         name=name,
         type=LayerType.SEQUENCE_LAST_INSTANCE,
         inputs=[input.name],
         trans_type=agg_level,
+        stride=stride,
         **ExtraLayerAttribute.to_kwargs(layer_attr))
     return LayerOutput(
         name,
@@ -1386,10 +1398,16 @@ def last_seq(input,
 def first_seq(input,
               name=None,
               agg_level=AggregateLevel.EACH_TIMESTEP,
+              stride=-1,
               layer_attr=None):
     """
     Get First Timestamp Activation of a sequence.
 
+    If stride > 0, this layer slides a window whose size is determined by stride, 
+    and return the first value of the window as the output. Thus, a long sequence 
+    will be shorten. Note that for sequence with sub-sequence, the default value 
+    of stride is -1.
+
     The simple usage is:
 
     .. code-block:: python
@@ -1401,6 +1419,8 @@ def first_seq(input,
     :type name: basestring
     :param input: Input layer name.
     :type input: LayerOutput
+    :param stride: window size.  
+    :type stride: Int
     :param layer_attr: extra layer attributes.
     :type layer_attr: ExtraLayerAttribute.
     :return: LayerOutput object.
@@ -1413,11 +1433,15 @@ def first_seq(input,
                        ' time series information at all. Maybe you want to use'
                        ' last_seq instead.')
 
+    if agg_level == AggregateLevel.EACH_SEQUENCE:
+        assert stride == -1
+
     Layer(
         name=name,
         type=LayerType.SEQUENCE_FIRST_INSTANCE,
         inputs=[input.name],
         trans_type=agg_level,
+        stride=stride,
         **ExtraLayerAttribute.to_kwargs(layer_attr))
     return LayerOutput(
         name,
@@ -4873,7 +4897,7 @@ def nce_layer(input,
     if neg_distribution is not None:
         assert isinstance(neg_distribution, collections.Sequence)
         assert len(neg_distribution) == num_classes
-        assert sum(neg_distribution) == 1
+        assert abs(sum(neg_distribution) - 1.0) < 1e-5
     if not isinstance(act, BaseActivation):
         raise TypeError()
 

python/paddle/trainer_config_helpers/tests/configs/last_first_seq.py

@@ -14,4 +14,7 @@ for op in seq_op:
     for al in agg_level:
         opts.append(op(input=din, agg_level=al))
 
+for op in seq_op:
+    opts.append(op(input=din, agg_level=AggregateLevel.EACH_TIMESTEP, stride=5))
+
 outputs(opts)

python/paddle/trainer_config_helpers/tests/configs/protostr/last_first_seq.protostr

@@ -15,6 +15,7 @@ layers {
   }
   select_first: true
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__first_seq_1__"
@@ -26,6 +27,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_0__"
@@ -36,6 +38,7 @@ layers {
     input_layer_name: "data"
   }
   trans_type: "seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_1__"
@@ -46,12 +49,38 @@ layers {
     input_layer_name: "data"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
+}
+layers {
+  name: "__first_seq_2__"
+  type: "seqlastins"
+  size: 30
+  active_type: "linear"
+  inputs {
+    input_layer_name: "data"
+  }
+  select_first: true
+  trans_type: "non-seq"
+  seq_pool_stride: 5
+}
+layers {
+  name: "__last_seq_2__"
+  type: "seqlastins"
+  size: 30
+  active_type: "linear"
+  inputs {
+    input_layer_name: "data"
+  }
+  trans_type: "non-seq"
+  seq_pool_stride: 5
 }
 input_layer_names: "data"
 output_layer_names: "__first_seq_0__"
 output_layer_names: "__first_seq_1__"
 output_layer_names: "__last_seq_0__"
 output_layer_names: "__last_seq_1__"
+output_layer_names: "__first_seq_2__"
+output_layer_names: "__last_seq_2__"
 sub_models {
   name: "root"
   layer_names: "data"
@@ -59,11 +88,15 @@ sub_models {
   layer_names: "__first_seq_1__"
   layer_names: "__last_seq_0__"
   layer_names: "__last_seq_1__"
+  layer_names: "__first_seq_2__"
+  layer_names: "__last_seq_2__"
   input_layer_names: "data"
   output_layer_names: "__first_seq_0__"
   output_layer_names: "__first_seq_1__"
   output_layer_names: "__last_seq_0__"
   output_layer_names: "__last_seq_1__"
+  output_layer_names: "__first_seq_2__"
+  output_layer_names: "__last_seq_2__"
   is_recurrent_layer_group: false
 }
 

python/paddle/trainer_config_helpers/tests/configs/protostr/shared_gru.protostr

@@ -128,6 +128,7 @@ layers {
     input_layer_name: "__simple_gru_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_1__"
@@ -138,6 +139,7 @@ layers {
     input_layer_name: "__simple_gru_1__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__fc_layer_0__"

python/paddle/trainer_config_helpers/tests/configs/protostr/shared_lstm.protostr

@@ -210,6 +210,7 @@ layers {
     input_layer_name: "__lstm_group_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_1__"
@@ -220,6 +221,7 @@ layers {
     input_layer_name: "__lstm_group_1__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__fc_layer_0__"

python/paddle/trainer_config_helpers/tests/configs/protostr/simple_rnn_layers.protostr

@@ -143,6 +143,7 @@ layers {
     input_layer_name: "__recurrent_layer_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__first_seq_0__"
@@ -154,6 +155,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_1__"
@@ -164,6 +166,7 @@ layers {
     input_layer_name: "__lstmemory_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__first_seq_1__"
@@ -175,6 +178,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__last_seq_2__"
@@ -185,6 +189,7 @@ layers {
     input_layer_name: "__gru_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__first_seq_2__"
@@ -196,6 +201,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 parameters {
   name: "___fc_layer_0__.w0"

python/paddle/trainer_config_helpers/tests/configs/protostr/test_rnn_group.protostr

@@ -96,6 +96,7 @@ layers {
     input_layer_name: "rnn_forward"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__recurrent_group_1__"
@@ -145,6 +146,7 @@ layers {
   }
   select_first: true
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__recurrent_group_2__"
@@ -193,6 +195,7 @@ layers {
     input_layer_name: "rnn_subseq_forward"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__lstm_group_0___recurrent_group"
@@ -282,6 +285,7 @@ layers {
     input_layer_name: "__lstm_group_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__gru_group_0___recurrent_group"
@@ -330,6 +334,7 @@ layers {
     input_layer_name: "__gru_group_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 layers {
   name: "__recurrent_group_3__"
@@ -378,6 +383,7 @@ layers {
     input_layer_name: "__fc_layer_0__"
   }
   trans_type: "non-seq"
+  seq_pool_stride: -1
 }
 parameters {
   name: "___mixed_0__.w0"

python/paddle/v2/data_feeder.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from py_paddle import DataProviderConverter
-
+import collections
 import paddle.trainer.PyDataProvider2 as pydp2
 
 __all__ = ['DataFeeder']
@@ -35,15 +35,30 @@ class DataFeeder(DataProviderConverter):
     DataFeeder converts this mini-batch data entries into Arguments in order
     to feed it to C++ interface.
     
-    The example usage:
+    The simple usage shows below
+
+    ..  code-block:: python
+
+        feeding = ['image', 'label']
+        data_types = enumerate_data_types_of_data_layers(topology)
+        feeder = DataFeeder(data_types=data_types, feeding=feeding)
+
+        minibatch_data = [([1.0, 2.0, 3.0, ...], 5)]
+
+        arg = feeder(minibatch_data)
+
+
+    If mini-batch data and data layers are not one to one mapping, we
+    could pass a dictionary to feeding parameter to represent the mapping
+    relationship.
 
 
     ..  code-block:: python
 
         data_types = [('image', paddle.data_type.dense_vector(784)),
                       ('label', paddle.data_type.integer_value(10))]
-        reader_dict = {'image':0, 'label':1}
-        feeder = DataFeeder(data_types=data_types, reader_dict=reader_dict)
+        feeding = {'image':0, 'label':1}
+        feeder = DataFeeder(data_types=data_types, feeding=feeding)
         minibatch_data = [
                            ( [1.0,2.0,3.0,4.0], 5, [6,7,8] ),  # first sample
                            ( [1.0,2.0,3.0,4.0], 5, [6,7,8] )   # second sample
@@ -65,9 +80,9 @@ class DataFeeder(DataProviderConverter):
                        a tuple of (data_name, data_type).
 
     :type data_types: list
-    :param reader_dict: A dictionary to specify the position of each data
-                        in the input data.
-    :type feeding: dict
+    :param feeding: A dictionary or a sequence to specify the position of each
+                    data in the input data.
+    :type feeding: dict|collections.Sequence|None
     """
 
     def __init__(self, data_types, feeding=None):
@@ -75,6 +90,13 @@ class DataFeeder(DataProviderConverter):
         input_types = []
         if feeding is None:
             feeding = default_feeding_map(data_types)
+        elif isinstance(feeding, collections.Sequence):
+            feed_list = feeding
+            feeding = dict()
+            for i, name in enumerate(feed_list):
+                feeding[name] = i
+        elif not isinstance(feeding, dict):
+            raise TypeError("Feeding should be dict or sequence or None.")
 
         self.feeding = feeding
         for each in data_types:

python/paddle/v2/dataset/wmt14.py

@@ -34,7 +34,7 @@ URL_TRAIN = 'http://paddlepaddle.cdn.bcebos.com/demo/wmt_shrinked_data/wmt14.tgz
 MD5_TRAIN = 'a755315dd01c2c35bde29a744ede23a6'
 # this is the pretrained model, whose bleu = 26.92
 URL_MODEL = 'http://paddlepaddle.bj.bcebos.com/demo/wmt_14/wmt14_model.tar.gz'
-MD5_MODEL = '6b097d23e15654608c6f74923e975535'
+MD5_MODEL = '4ce14a26607fb8a1cc23bcdedb1895e4'
 
 START = "<s>"
 END = "<e>"
@@ -140,6 +140,12 @@ def model():
     return parameters
 
 
+def trg_dict(dict_size):
+    tar_file = download(URL_TRAIN, 'wmt14', MD5_TRAIN)
+    src_dict, trg_dict = __read_to_dict__(tar_file, dict_size)
+    return trg_dict
+
+
 def fetch():
     download(URL_TRAIN, 'wmt14', MD5_TRAIN)
     download(URL_MODEL, 'wmt14', MD5_MODEL)

python/paddle/v2/inference.py

@@ -48,8 +48,13 @@ class Inference(object):
         self.__gradient_machine__.finish()
 
     def iter_infer_field(self, field, **kwargs):
+        if not isinstance(field, list) and not isinstance(field, tuple):
+            field = [field]
+
         for result in self.iter_infer(**kwargs):
-            yield [each_result[field] for each_result in result]
+            for each_result in result:
+                item = [each_result[each_field] for each_field in field]
+                yield item
 
     def infer(self, field='value', **kwargs):
         retv = None
@@ -87,9 +92,11 @@ def infer(output_layer, parameters, input, feeding=None, field='value'):
     :type input: collections.Iterable
     :param feeding: Reader dictionary. Default could generate from input
                         value.
-    :param field: The prediction field. It should in [`value`, `ids`]. `value`
-                  means return the prediction probabilities, `ids` means return
-                  the prediction labels. Default is `value`
+    :param field: The prediction field. It should in [`value`, `id`, `prob`]. 
+                  `value` and `prob` mean return the prediction probabilities, 
+                  `id` means return the prediction labels. Default is `value`.
+                  Note that `prob` only used when output_layer is beam_search 
+                  or max_id.
     :type field: str
     :return: a numpy array
     :rtype: numpy.ndarray

python/paddle/v2/trainer.py

@@ -83,7 +83,7 @@ class SGD(object):
         :type event_handler: (BaseEvent) => None
         :param feeding: Feeding is a map of neural network input name and array
                         index that reader returns.
-        :type feeding: dict
+        :type feeding: dict|list
         :return:
         """
         if event_handler is None: