提交 df5a95dc 编写于 作者: C Cao Ying 提交者: GitHub

Merge pull request #1701 from luotao1/stride

stride pooling for seqlastin and seqfirstin.
...@@ -25,6 +25,11 @@ namespace paddle { ...@@ -25,6 +25,11 @@ namespace paddle {
* Input: a sequence * Input: a sequence
* If SequenceLevel = kNonseq: * If SequenceLevel = kNonseq:
* Output: a sequence containing only the last instance of the input sequence * 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: * If SequenceLevel = kSeq:
* Check input sequence must has sub-sequence * Check input sequence must has sub-sequence
* Output: a sequence containing only the last instance of each sub-sequence * Output: a sequence containing only the last instance of each sub-sequence
...@@ -37,6 +42,7 @@ class SequenceLastInstanceLayer : public SequencePoolLayer { ...@@ -37,6 +42,7 @@ class SequenceLastInstanceLayer : public SequencePoolLayer {
protected: protected:
MatrixPtr tmpSrc_; MatrixPtr tmpSrc_;
MatrixPtr tmpDest_; MatrixPtr tmpDest_;
std::vector<int> instanceIds_;
public: public:
explicit SequenceLastInstanceLayer(const LayerConfig& config) explicit SequenceLastInstanceLayer(const LayerConfig& config)
...@@ -54,6 +60,7 @@ REGISTER_LAYER(seqlastins, SequenceLastInstanceLayer); ...@@ -54,6 +60,7 @@ REGISTER_LAYER(seqlastins, SequenceLastInstanceLayer);
bool SequenceLastInstanceLayer::init(const LayerMap& layerMap, bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) { const ParameterMap& parameterMap) {
SequencePoolLayer::init(layerMap, parameterMap); SequencePoolLayer::init(layerMap, parameterMap);
reversed_ = config_.select_first();
tmpSrc_ = tmpSrc_ =
Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_); Matrix::create(nullptr, /* height= */ 1, 1, /* trans= */ false, useGpu_);
...@@ -66,7 +73,8 @@ bool SequenceLastInstanceLayer::init(const LayerMap& layerMap, ...@@ -66,7 +73,8 @@ bool SequenceLastInstanceLayer::init(const LayerMap& layerMap,
void SequenceLastInstanceLayer::forward(PassType passType) { void SequenceLastInstanceLayer::forward(PassType passType) {
SequencePoolLayer::forward(passType); SequencePoolLayer::forward(passType);
const int* starts = startPositions_->getData(false); auto starts = (stride_ > 0) ? stridePositions_->getData()
: startPositions_->getData(false);
MatrixPtr inputValue = getInputValue(0); MatrixPtr inputValue = getInputValue(0);
MatrixPtr outputValue = getOutputValue(); MatrixPtr outputValue = getOutputValue();
...@@ -74,9 +82,10 @@ void SequenceLastInstanceLayer::forward(PassType passType) { ...@@ -74,9 +82,10 @@ void SequenceLastInstanceLayer::forward(PassType passType) {
AsyncGpuBlock asyncGpuBlock; AsyncGpuBlock asyncGpuBlock;
REGISTER_TIMER_INFO("SequenceLastInstanceLayerForward", getName().c_str()); REGISTER_TIMER_INFO("SequenceLastInstanceLayerForward", getName().c_str());
instanceIds_.clear();
for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) { for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) {
int insId = int insId = reversed_ ? starts[seqId] : starts[seqId + 1] - 1;
config_.select_first() ? starts[seqId] : starts[seqId + 1] - 1; instanceIds_.push_back(insId);
outputValue->subMatrix(seqId, 1, tmpDest_) outputValue->subMatrix(seqId, 1, tmpDest_)
->assign(*(inputValue->subMatrix(insId, 1, tmpSrc_))); ->assign(*(inputValue->subMatrix(insId, 1, tmpSrc_)));
...@@ -96,18 +105,13 @@ void SequenceLastInstanceLayer::backward(const UpdateCallback& callback) { ...@@ -96,18 +105,13 @@ void SequenceLastInstanceLayer::backward(const UpdateCallback& callback) {
MatrixPtr inputGrad = getInputGrad(0); MatrixPtr inputGrad = getInputGrad(0);
MatrixPtr outputGrad = getOutputGrad(); MatrixPtr outputGrad = getOutputGrad();
const int* starts = startPositions_->getData(false);
size_t numSequences = startPositions_->getSize() - 1;
if (inputGrad) { if (inputGrad) {
AsyncGpuBlock asyncGpuBlock; AsyncGpuBlock asyncGpuBlock;
REGISTER_TIMER_INFO("SequenceLastInstanceLayerBackward", getName().c_str()); REGISTER_TIMER_INFO("SequenceLastInstanceLayerBackward", getName().c_str());
for (size_t seqId = 0; seqId < numSequences; ++seqId) { for (size_t seqId = 0; seqId < newBatchSize_; ++seqId) {
int insId = inputGrad->subMatrix(instanceIds_[seqId], 1, tmpDest_)
config_.select_first() ? starts[seqId] : starts[seqId + 1] - 1;
inputGrad->subMatrix(insId, 1, tmpDest_)
->add(*(outputGrad->subMatrix(seqId, 1, tmpSrc_))); ->add(*(outputGrad->subMatrix(seqId, 1, tmpSrc_)));
} }
} }
......
...@@ -37,6 +37,7 @@ bool SequencePoolLayer::init(const LayerMap& layerMap, ...@@ -37,6 +37,7 @@ bool SequencePoolLayer::init(const LayerMap& layerMap,
} else { } else {
LOG(FATAL) << "Unknown trans_type: " << config_.trans_type(); LOG(FATAL) << "Unknown trans_type: " << config_.trans_type();
} }
stride_ = config_.seq_pool_stride();
setNeedSequenceInfo(false); setNeedSequenceInfo(false);
return true; return true;
} }
...@@ -55,8 +56,6 @@ void SequencePoolLayer::forward(PassType passType) { ...@@ -55,8 +56,6 @@ void SequencePoolLayer::forward(PassType passType) {
CHECK_EQ(starts->getData()[newBatchSize_], input.getBatchSize()); CHECK_EQ(starts->getData()[newBatchSize_], input.getBatchSize());
CHECK_EQ(newBatchSize_, starts->getSize() - 1); 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, /* If type_ = kNonSeq, both seq has or not has sub-seq degrade to a non-seq,
* thus, in this case, output_ has no sequenceStartPositions. * thus, in this case, output_ has no sequenceStartPositions.
* If type_ = kSeq, seq has sub-seq degrades to a seq, thus, only in this * If type_ = kSeq, seq has sub-seq degrades to a seq, thus, only in this
...@@ -67,6 +66,15 @@ void SequencePoolLayer::forward(PassType passType) { ...@@ -67,6 +66,15 @@ void SequencePoolLayer::forward(PassType passType) {
<< "when trans_type = seq, input must hasSubseq"; << "when trans_type = seq, input must hasSubseq";
output_.degradeSequence(input); 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) { void SequencePoolLayer::backward(const UpdateCallback& callback) {
......
...@@ -26,6 +26,10 @@ namespace paddle { ...@@ -26,6 +26,10 @@ namespace paddle {
* Output: output size is the number of input sequences (NOT input instances) * Output: output size is the number of input sequences (NOT input instances)
* output[i] = seqlastin/average/max_{for each instance in this * output[i] = seqlastin/average/max_{for each instance in this
* sequence}{input[i]} * 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: * If SequenceLevel = kSeq:
* Check input sequence must has sub-sequence * Check input sequence must has sub-sequence
* Output: output size is the number of input sub-sequences * Output: output size is the number of input sub-sequences
...@@ -42,6 +46,11 @@ protected: ...@@ -42,6 +46,11 @@ protected:
enum SequenceLevel { kNonSeq = 0, kSeq = 1 }; enum SequenceLevel { kNonSeq = 0, kSeq = 1 };
size_t newBatchSize_; size_t newBatchSize_;
ICpuGpuVectorPtr startPositions_; 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: public:
explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {} explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {}
......
...@@ -804,10 +804,14 @@ TEST(Layer, ExpandLayer) { ...@@ -804,10 +804,14 @@ TEST(Layer, ExpandLayer) {
testExpandLayer("seq", true); // seq expand to hasSubseq 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; TestConfig config;
config.layerConfig.set_type(layer_type); config.layerConfig.set_type(layer_type);
config.layerConfig.set_size(10); config.layerConfig.set_size(10);
config.layerConfig.set_seq_pool_stride(stride);
config.biasSize = 0; config.biasSize = 0;
config.inputDefs.push_back( config.inputDefs.push_back(
...@@ -827,36 +831,46 @@ void testDegradeLayer(bool hasSubseq, string layer_type, string trans_type) { ...@@ -827,36 +831,46 @@ void testDegradeLayer(bool hasSubseq, string layer_type, string trans_type) {
if (layer_type == "average") { if (layer_type == "average") {
for (auto strategy : {"average", "sum", "squarerootn"}) { for (auto strategy : {"average", "sum", "squarerootn"}) {
LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type LOG(INFO) << " hasSubseq=" << hasSubseq << " trans_type=" << trans_type
<< " average_strategy=" << strategy; << " average_strategy=" << strategy
<< " seq_pool_stride=" << stride;
config.layerConfig.set_average_strategy(strategy); config.layerConfig.set_average_strategy(strategy);
testDegradeLayerGrad(config, layer_type); testDegradeLayerGrad(config, layer_type);
} }
} else { } 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); testDegradeLayerGrad(config, layer_type);
} }
} }
TEST(Layer, MaxLayer) { TEST(Layer, MaxLayer) {
testDegradeLayer(false, "max", "non-seq"); // seq max to non-seq testDegradeLayer(false, "max", "non-seq", -1); // seq max to non-seq
testDegradeLayer(true, "max", "non-seq"); // hasSubseq max to non-seq testDegradeLayer(true, "max", "non-seq", -1); // hasSubseq max to non-seq
testDegradeLayer(true, "max", "seq"); // hasSubseq max to seq testDegradeLayer(true, "max", "seq", -1); // hasSubseq max to seq
} }
TEST(Layer, SequenceLastInstanceLayer) { TEST(Layer, SequenceLastInstanceLayer) {
testDegradeLayer(false, testDegradeLayer(false,
"seqlastins", "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, testDegradeLayer(true,
"seqlastins", "seqlastins",
"non-seq"); // hasSubseq seqlastins to non-seq "non-seq",
testDegradeLayer(true, "seqlastins", "seq"); // hasSubseq seqlastins to seq -1); // hasSubseq seqlastins to non-seq
testDegradeLayer(
true, "seqlastins", "seq", -1); // hasSubseq seqlastins to seq
} }
TEST(Layer, AverageLayer) { TEST(Layer, AverageLayer) {
testDegradeLayer(false, "average", "non-seq"); // seq average to non-seq testDegradeLayer(false, "average", "non-seq", -1); // seq average to non-seq
testDegradeLayer(true, "average", "non-seq"); // hasSubseq average to non-seq testDegradeLayer(
testDegradeLayer(true, "average", "seq"); // hasSubseq average to seq true, "average", "non-seq", -1); // hasSubseq average to non-seq
testDegradeLayer(true, "average", "seq", -1); // hasSubseq average to seq
} }
TEST(Layer, SequenceConcatLayer) { TEST(Layer, SequenceConcatLayer) {
......
...@@ -559,6 +559,49 @@ void Argument::degradeSequence(const Argument& input) { ...@@ -559,6 +559,49 @@ void Argument::degradeSequence(const Argument& input) {
tgtBuf[numSequences] = numSubSequences; 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( void Argument::getValueString(
std::unordered_map<std::string, std::string>* out) const { std::unordered_map<std::string, std::string>* out) const {
if (value) { if (value) {
......
...@@ -291,6 +291,15 @@ struct Argument { ...@@ -291,6 +291,15 @@ struct Argument {
*/ */
void degradeSequence(const Argument& input); 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 * @brief getValueString will return the argument's output in string. There
* are several kinds of output. The keys of output dictionary are 'value', * are several kinds of output. The keys of output dictionary are 'value',
......
add_simple_unittest(test_common) add_simple_unittest(test_common)
add_simple_unittest(test_argument)
/* 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();
}
...@@ -441,6 +441,11 @@ message LayerConfig { ...@@ -441,6 +441,11 @@ message LayerConfig {
// blank label used in ctc loss // blank label used in ctc loss
optional uint32 blank = 52 [default = 0]; 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 { message EvaluatorConfig {
......
...@@ -2485,6 +2485,7 @@ class SequenceLastInstanceLayer(LayerBase): ...@@ -2485,6 +2485,7 @@ class SequenceLastInstanceLayer(LayerBase):
active_type='linear', active_type='linear',
trans_type='non-seq', trans_type='non-seq',
bias=False, bias=False,
stride=-1,
**xargs): **xargs):
super(SequenceLastInstanceLayer, self).__init__( super(SequenceLastInstanceLayer, self).__init__(
name, name,
...@@ -2495,10 +2496,11 @@ class SequenceLastInstanceLayer(LayerBase): ...@@ -2495,10 +2496,11 @@ class SequenceLastInstanceLayer(LayerBase):
**xargs) **xargs)
config_assert( config_assert(
len(inputs) == 1, 'SequenceLastInstanceLayer must have 1 input') 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 self.config.trans_type = trans_type
for input_index in xrange(len(self.inputs)): self.config.seq_pool_stride = stride
input_layer = self.get_input_layer(input_index) self.set_layer_size(self.get_input_layer(0).size)
self.set_layer_size(input_layer.size)
self.create_bias_parameter(bias, self.config.size) self.create_bias_parameter(bias, self.config.size)
...@@ -2510,10 +2512,16 @@ class SequenceFirstInstanceLayer(SequenceLastInstanceLayer): ...@@ -2510,10 +2512,16 @@ class SequenceFirstInstanceLayer(SequenceLastInstanceLayer):
active_type='linear', active_type='linear',
trans_type='non-seq', trans_type='non-seq',
bias=False, bias=False,
stride=-1,
**xargs): **xargs):
super(SequenceFirstInstanceLayer, self).__init__( super(SequenceFirstInstanceLayer, self).__init__(
name, inputs=inputs, active_type=active_type, bias=bias, **xargs) name,
self.config.trans_type = trans_type inputs=inputs,
active_type=active_type,
trans_type=trans_type,
bias=bias,
stride=stride,
**xargs)
self.config.select_first = True self.config.select_first = True
......
...@@ -1342,10 +1342,16 @@ def grumemory(input, ...@@ -1342,10 +1342,16 @@ def grumemory(input,
def last_seq(input, def last_seq(input,
name=None, name=None,
agg_level=AggregateLevel.EACH_TIMESTEP, agg_level=AggregateLevel.EACH_TIMESTEP,
stride=-1,
layer_attr=None): layer_attr=None):
""" """
Get Last Timestamp Activation of a sequence. 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: The simple usage is:
.. code-block:: python .. code-block:: python
...@@ -1357,6 +1363,8 @@ def last_seq(input, ...@@ -1357,6 +1363,8 @@ def last_seq(input,
:type name: basestring :type name: basestring
:param input: Input layer name. :param input: Input layer name.
:type input: LayerOutput :type input: LayerOutput
:param stride: window size.
:type stride: Int
:param layer_attr: extra layer attributes. :param layer_attr: extra layer attributes.
:type layer_attr: ExtraLayerAttribute. :type layer_attr: ExtraLayerAttribute.
:return: LayerOutput object. :return: LayerOutput object.
...@@ -1368,11 +1376,15 @@ def last_seq(input, ...@@ -1368,11 +1376,15 @@ def last_seq(input,
" series information at all. Maybe you want to use" " series information at all. Maybe you want to use"
" first_seq instead.") " first_seq instead.")
if agg_level == AggregateLevel.EACH_SEQUENCE:
assert stride == -1
Layer( Layer(
name=name, name=name,
type=LayerType.SEQUENCE_LAST_INSTANCE, type=LayerType.SEQUENCE_LAST_INSTANCE,
inputs=[input.name], inputs=[input.name],
trans_type=agg_level, trans_type=agg_level,
stride=stride,
**ExtraLayerAttribute.to_kwargs(layer_attr)) **ExtraLayerAttribute.to_kwargs(layer_attr))
return LayerOutput( return LayerOutput(
name, name,
...@@ -1386,10 +1398,16 @@ def last_seq(input, ...@@ -1386,10 +1398,16 @@ def last_seq(input,
def first_seq(input, def first_seq(input,
name=None, name=None,
agg_level=AggregateLevel.EACH_TIMESTEP, agg_level=AggregateLevel.EACH_TIMESTEP,
stride=-1,
layer_attr=None): layer_attr=None):
""" """
Get First Timestamp Activation of a sequence. 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: The simple usage is:
.. code-block:: python .. code-block:: python
...@@ -1401,6 +1419,8 @@ def first_seq(input, ...@@ -1401,6 +1419,8 @@ def first_seq(input,
:type name: basestring :type name: basestring
:param input: Input layer name. :param input: Input layer name.
:type input: LayerOutput :type input: LayerOutput
:param stride: window size.
:type stride: Int
:param layer_attr: extra layer attributes. :param layer_attr: extra layer attributes.
:type layer_attr: ExtraLayerAttribute. :type layer_attr: ExtraLayerAttribute.
:return: LayerOutput object. :return: LayerOutput object.
...@@ -1413,11 +1433,15 @@ def first_seq(input, ...@@ -1413,11 +1433,15 @@ def first_seq(input,
' time series information at all. Maybe you want to use' ' time series information at all. Maybe you want to use'
' last_seq instead.') ' last_seq instead.')
if agg_level == AggregateLevel.EACH_SEQUENCE:
assert stride == -1
Layer( Layer(
name=name, name=name,
type=LayerType.SEQUENCE_FIRST_INSTANCE, type=LayerType.SEQUENCE_FIRST_INSTANCE,
inputs=[input.name], inputs=[input.name],
trans_type=agg_level, trans_type=agg_level,
stride=stride,
**ExtraLayerAttribute.to_kwargs(layer_attr)) **ExtraLayerAttribute.to_kwargs(layer_attr))
return LayerOutput( return LayerOutput(
name, name,
......
...@@ -14,4 +14,7 @@ for op in seq_op: ...@@ -14,4 +14,7 @@ for op in seq_op:
for al in agg_level: for al in agg_level:
opts.append(op(input=din, agg_level=al)) 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) outputs(opts)
...@@ -15,6 +15,7 @@ layers { ...@@ -15,6 +15,7 @@ layers {
} }
select_first: true select_first: true
trans_type: "seq" trans_type: "seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__first_seq_1__" name: "__first_seq_1__"
...@@ -26,6 +27,7 @@ layers { ...@@ -26,6 +27,7 @@ layers {
} }
select_first: true select_first: true
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__last_seq_0__" name: "__last_seq_0__"
...@@ -36,6 +38,7 @@ layers { ...@@ -36,6 +38,7 @@ layers {
input_layer_name: "data" input_layer_name: "data"
} }
trans_type: "seq" trans_type: "seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__last_seq_1__" name: "__last_seq_1__"
...@@ -46,12 +49,38 @@ layers { ...@@ -46,12 +49,38 @@ layers {
input_layer_name: "data" input_layer_name: "data"
} }
trans_type: "non-seq" 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" input_layer_names: "data"
output_layer_names: "__first_seq_0__" output_layer_names: "__first_seq_0__"
output_layer_names: "__first_seq_1__" output_layer_names: "__first_seq_1__"
output_layer_names: "__last_seq_0__" output_layer_names: "__last_seq_0__"
output_layer_names: "__last_seq_1__" output_layer_names: "__last_seq_1__"
output_layer_names: "__first_seq_2__"
output_layer_names: "__last_seq_2__"
sub_models { sub_models {
name: "root" name: "root"
layer_names: "data" layer_names: "data"
...@@ -59,11 +88,15 @@ sub_models { ...@@ -59,11 +88,15 @@ sub_models {
layer_names: "__first_seq_1__" layer_names: "__first_seq_1__"
layer_names: "__last_seq_0__" layer_names: "__last_seq_0__"
layer_names: "__last_seq_1__" layer_names: "__last_seq_1__"
layer_names: "__first_seq_2__"
layer_names: "__last_seq_2__"
input_layer_names: "data" input_layer_names: "data"
output_layer_names: "__first_seq_0__" output_layer_names: "__first_seq_0__"
output_layer_names: "__first_seq_1__" output_layer_names: "__first_seq_1__"
output_layer_names: "__last_seq_0__" output_layer_names: "__last_seq_0__"
output_layer_names: "__last_seq_1__" output_layer_names: "__last_seq_1__"
output_layer_names: "__first_seq_2__"
output_layer_names: "__last_seq_2__"
is_recurrent_layer_group: false is_recurrent_layer_group: false
} }
...@@ -128,6 +128,7 @@ layers { ...@@ -128,6 +128,7 @@ layers {
input_layer_name: "__simple_gru_0__" input_layer_name: "__simple_gru_0__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__last_seq_1__" name: "__last_seq_1__"
...@@ -138,6 +139,7 @@ layers { ...@@ -138,6 +139,7 @@ layers {
input_layer_name: "__simple_gru_1__" input_layer_name: "__simple_gru_1__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__fc_layer_0__" name: "__fc_layer_0__"
......
...@@ -210,6 +210,7 @@ layers { ...@@ -210,6 +210,7 @@ layers {
input_layer_name: "__lstm_group_0__" input_layer_name: "__lstm_group_0__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__last_seq_1__" name: "__last_seq_1__"
...@@ -220,6 +221,7 @@ layers { ...@@ -220,6 +221,7 @@ layers {
input_layer_name: "__lstm_group_1__" input_layer_name: "__lstm_group_1__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__fc_layer_0__" name: "__fc_layer_0__"
......
...@@ -143,6 +143,7 @@ layers { ...@@ -143,6 +143,7 @@ layers {
input_layer_name: "__recurrent_layer_0__" input_layer_name: "__recurrent_layer_0__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__first_seq_0__" name: "__first_seq_0__"
...@@ -154,6 +155,7 @@ layers { ...@@ -154,6 +155,7 @@ layers {
} }
select_first: true select_first: true
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__last_seq_1__" name: "__last_seq_1__"
...@@ -164,6 +166,7 @@ layers { ...@@ -164,6 +166,7 @@ layers {
input_layer_name: "__lstmemory_0__" input_layer_name: "__lstmemory_0__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__first_seq_1__" name: "__first_seq_1__"
...@@ -175,6 +178,7 @@ layers { ...@@ -175,6 +178,7 @@ layers {
} }
select_first: true select_first: true
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__last_seq_2__" name: "__last_seq_2__"
...@@ -185,6 +189,7 @@ layers { ...@@ -185,6 +189,7 @@ layers {
input_layer_name: "__gru_0__" input_layer_name: "__gru_0__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__first_seq_2__" name: "__first_seq_2__"
...@@ -196,6 +201,7 @@ layers { ...@@ -196,6 +201,7 @@ layers {
} }
select_first: true select_first: true
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
parameters { parameters {
name: "___fc_layer_0__.w0" name: "___fc_layer_0__.w0"
......
...@@ -96,6 +96,7 @@ layers { ...@@ -96,6 +96,7 @@ layers {
input_layer_name: "rnn_forward" input_layer_name: "rnn_forward"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__recurrent_group_1__" name: "__recurrent_group_1__"
...@@ -145,6 +146,7 @@ layers { ...@@ -145,6 +146,7 @@ layers {
} }
select_first: true select_first: true
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__recurrent_group_2__" name: "__recurrent_group_2__"
...@@ -193,6 +195,7 @@ layers { ...@@ -193,6 +195,7 @@ layers {
input_layer_name: "rnn_subseq_forward" input_layer_name: "rnn_subseq_forward"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__lstm_group_0___recurrent_group" name: "__lstm_group_0___recurrent_group"
...@@ -282,6 +285,7 @@ layers { ...@@ -282,6 +285,7 @@ layers {
input_layer_name: "__lstm_group_0__" input_layer_name: "__lstm_group_0__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__gru_group_0___recurrent_group" name: "__gru_group_0___recurrent_group"
...@@ -330,6 +334,7 @@ layers { ...@@ -330,6 +334,7 @@ layers {
input_layer_name: "__gru_group_0__" input_layer_name: "__gru_group_0__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
layers { layers {
name: "__recurrent_group_3__" name: "__recurrent_group_3__"
...@@ -378,6 +383,7 @@ layers { ...@@ -378,6 +383,7 @@ layers {
input_layer_name: "__fc_layer_0__" input_layer_name: "__fc_layer_0__"
} }
trans_type: "non-seq" trans_type: "non-seq"
seq_pool_stride: -1
} }
parameters { parameters {
name: "___mixed_0__.w0" name: "___mixed_0__.w0"
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册