qingqing01
7 years ago
committed by
sneaxiy
parent
7748963479
commit
79918a8442
@ -0,0 +1,26 @@
|
||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// Unless required by applicable law or agreed to in writing, software
|
||||
// distributed under the License is distributed on an "AS IS" BASIS,
|
||||
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
// See the License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
#include "paddle/fluid/operators/sequence_mask_op.h"
|
||||
|
||||
REGISTER_OPERATOR(sequence_mask, paddle::operators::SequenceMaskOp,
|
||||
paddle::operators::SequenceMaskOpMaker,
|
||||
paddle::framework::EmptyGradOpMaker);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
sequence_mask,
|
||||
paddle::operators::SequenceMaskKernel<paddle::platform::CPUDeviceContext,
|
||||
int>,
|
||||
paddle::operators::SequenceMaskKernel<paddle::platform::CPUDeviceContext,
|
||||
int64_t>);
|
||||
@ -0,0 +1,22 @@
|
||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// Unless required by applicable law or agreed to in writing, software
|
||||
// distributed under the License is distributed on an "AS IS" BASIS,
|
||||
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
// See the License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
#include "paddle/fluid/operators/sequence_mask_op.h"
|
||||
|
||||
REGISTER_OP_CUDA_KERNEL(
|
||||
sequence_mask,
|
||||
paddle::operators::SequenceMaskKernel<paddle::platform::CUDADeviceContext,
|
||||
int>,
|
||||
paddle::operators::SequenceMaskKernel<paddle::platform::CUDADeviceContext,
|
||||
int64_t>);
|
||||
@ -0,0 +1,117 @@
|
||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// Unless required by applicable law or agreed to in writing, software
|
||||
// distributed under the License is distributed on an "AS IS" BASIS,
|
||||
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
// See the License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "paddle/fluid/framework/op_registry.h"
|
||||
#include "paddle/fluid/platform/for_range.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class SequenceMaskOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
void InferShape(framework::InferShapeContext *ctx) const override {
|
||||
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must exist");
|
||||
auto max_len = ctx->Attrs().Get<int>("max_len");
|
||||
PADDLE_ENFORCE_GT(max_len, 1, "Attr(max_len) must be larger than 1");
|
||||
PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) must exist");
|
||||
auto dim = framework::vectorize2int(ctx->GetInputDim("X"));
|
||||
dim.push_back(max_len);
|
||||
ctx->SetOutputDim("Y", framework::make_ddim(dim));
|
||||
}
|
||||
};
|
||||
|
||||
class SequenceMaskOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
void Make() override {
|
||||
AddInput("X", "The input of sequence_mask op.");
|
||||
AddOutput("Y", "The output mask of sequence_mask op.");
|
||||
AddAttr<int>("max_len", "The maximum length of the sequence.")
|
||||
.GreaterThan(1);
|
||||
AddAttr<int>("out_dtype", "Output data type");
|
||||
AddComment(R"DOC(
|
||||
SequenceMask Operator
|
||||
|
||||
This operator outputs a Mask according to Input(X) and Attr(max_len).
|
||||
Supposing Input(X) is a Tensor with shape [d_1, d_2, ..., d_n], the
|
||||
Output(Y) is a mask with shape [d_1, d_2, ..., d_n, max_len], where:
|
||||
|
||||
Y(i_1, i_2, ..., i_n, j) = (j < X(i_1, i_2, ..., i_n))
|
||||
)DOC");
|
||||
}
|
||||
};
|
||||
|
||||
template <typename Tx, typename Ty>
|
||||
struct SequenceMaskForRangeFunctor {
|
||||
HOSTDEVICE SequenceMaskForRangeFunctor(const Tx *x, Ty *y, int max_len)
|
||||
: x_(x), y_(y), max_len_(max_len) {}
|
||||
|
||||
HOSTDEVICE void operator()(int y_idx) const {
|
||||
int x_idx = y_idx / max_len_;
|
||||
int j = y_idx % max_len_;
|
||||
y_[y_idx] = static_cast<Ty>(j < x_[x_idx] ? 1 : 0);
|
||||
}
|
||||
|
||||
private:
|
||||
const Tx *x_;
|
||||
Ty *y_;
|
||||
int max_len_;
|
||||
};
|
||||
|
||||
template <typename DeviceContext, typename Tx>
|
||||
struct SequenceMaskFunctor {
|
||||
using Tensor = framework::LoDTensor;
|
||||
|
||||
SequenceMaskFunctor(const DeviceContext &ctx, const Tx *x, Tensor *y,
|
||||
int limits, int max_len)
|
||||
: ctx_(ctx), x_(x), y_(y), limits_(limits), max_len_(max_len) {}
|
||||
|
||||
template <typename Ty>
|
||||
void operator()() const {
|
||||
auto *y_data = y_->mutable_data<Ty>(ctx_.GetPlace());
|
||||
platform::ForRange<DeviceContext> for_range(ctx_, limits_);
|
||||
for_range(SequenceMaskForRangeFunctor<Tx, Ty>(x_, y_data, max_len_));
|
||||
}
|
||||
|
||||
private:
|
||||
const DeviceContext &ctx_;
|
||||
const Tx *x_;
|
||||
Tensor *y_;
|
||||
int limits_;
|
||||
int max_len_;
|
||||
};
|
||||
|
||||
template <typename DeviceContext, typename Tx>
|
||||
class SequenceMaskKernel : public framework::OpKernel<Tx> {
|
||||
using Tensor = framework::LoDTensor;
|
||||
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext &ctx) const override {
|
||||
auto *x = ctx.Input<Tensor>("X");
|
||||
auto *y = ctx.Output<Tensor>("Y");
|
||||
auto max_len = ctx.Attr<int>("max_len");
|
||||
auto out_dtype = static_cast<framework::proto::VarType::Type>(
|
||||
ctx.Attr<int>("out_dtype"));
|
||||
auto &dev_ctx = ctx.template device_context<DeviceContext>();
|
||||
framework::VisitDataType(out_dtype, SequenceMaskFunctor<DeviceContext, Tx>(
|
||||
dev_ctx, x->data<Tx>(), y,
|
||||
x->numel() * max_len, max_len));
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,86 @@
|
||||
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
from op_test import OpTest
|
||||
from paddle.fluid.framework import convert_np_dtype_to_dtype_
|
||||
import numpy as np
|
||||
import copy
|
||||
import unittest
|
||||
|
||||
|
||||
class SequenceMaskTestBase(OpTest):
|
||||
def initDefaultParameters(self):
|
||||
self.op_type = 'sequence_mask'
|
||||
self.max_len = 10
|
||||
self.mask_dtype = 'int64'
|
||||
self.x = [[0, 3, 4], [5, 7, 9]]
|
||||
|
||||
def initParameters(self):
|
||||
pass
|
||||
|
||||
def setUp(self):
|
||||
self.initDefaultParameters()
|
||||
self.initParameters()
|
||||
if not isinstance(self.x, np.ndarray):
|
||||
self.x = np.array(self.x)
|
||||
|
||||
self.inputs = {'X': self.x}
|
||||
self.outputs = {'Y': self.calc_ground_truth_mask()}
|
||||
self.attrs = {
|
||||
'max_len': self.max_len,
|
||||
'out_dtype': convert_np_dtype_to_dtype_(self.mask_dtype)
|
||||
}
|
||||
|
||||
def calc_ground_truth_mask(self):
|
||||
shape = self.x.shape + (self.max_len, )
|
||||
index_broadcast = np.broadcast_to(
|
||||
np.reshape(
|
||||
range(self.max_len), newshape=[1] * self.x.ndim + [-1]),
|
||||
shape=shape)
|
||||
x_broadcast = np.broadcast_to(
|
||||
np.reshape(
|
||||
self.x, newshape=self.x.shape + (-1, )), shape=shape)
|
||||
return (index_broadcast < x_broadcast).astype(self.mask_dtype)
|
||||
|
||||
def test_check_output(self):
|
||||
self.check_output()
|
||||
|
||||
|
||||
class SequenceMaskTest1(SequenceMaskTestBase):
|
||||
def initParameters(self):
|
||||
self.mask_dtype = 'bool'
|
||||
|
||||
|
||||
class SequenceMaskTest2(SequenceMaskTestBase):
|
||||
def initParameters(self):
|
||||
self.mask_dtype = 'uint8'
|
||||
|
||||
|
||||
class SequenceMaskTest3(SequenceMaskTestBase):
|
||||
def initParameters(self):
|
||||
self.mask_dtype = 'int32'
|
||||
|
||||
|
||||
class SequenceMaskTest4(SequenceMaskTestBase):
|
||||
def initParameters(self):
|
||||
self.mask_dtype = 'float32'
|
||||
|
||||
|
||||
class SequenceMaskTest5(SequenceMaskTestBase):
|
||||
def initParameters(self):
|
||||
self.mask_dtype = 'float64'
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
Loading…
Reference in new issue