Merge pull request #12737 from tensor-tang/feature/op/fusion_lstm
add fusion lstmdataset_flowers_-_change_md5
commit
e955361267
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,42 @@
|
|||||||
|
/* 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. */
|
||||||
|
|
||||||
|
#pragma once
|
||||||
|
// #include <string>
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
using LoDTensor = framework::LoDTensor;
|
||||||
|
using Tensor = framework::Tensor;
|
||||||
|
|
||||||
|
class FusionLSTMOp : public framework::OperatorWithKernel {
|
||||||
|
public:
|
||||||
|
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||||
|
|
||||||
|
void InferShape(framework::InferShapeContext* ctx) const override;
|
||||||
|
|
||||||
|
protected:
|
||||||
|
framework::OpKernelType GetExpectedKernelType(
|
||||||
|
const framework::ExecutionContext& ctx) const override;
|
||||||
|
};
|
||||||
|
|
||||||
|
class FusionLSTMOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||||
|
public:
|
||||||
|
void Make() override;
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
@ -0,0 +1,43 @@
|
|||||||
|
/* 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. */
|
||||||
|
|
||||||
|
#pragma once
|
||||||
|
|
||||||
|
#include "paddle/fluid/operators/math/blas.h"
|
||||||
|
|
||||||
|
DECLARE_int32(paddle_num_threads);
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
namespace math {
|
||||||
|
|
||||||
|
template <typename DeviceContext, typename T>
|
||||||
|
inline void FCCompute(const BlasT<DeviceContext, T>& blas, const int M,
|
||||||
|
const int N, const int K, const T* X, const T* W, T* Y,
|
||||||
|
const T* B = NULL) {
|
||||||
|
blas.GEMM(CblasNoTrans, CblasNoTrans, M, N, K, static_cast<T>(1), X, W,
|
||||||
|
static_cast<T>(0), Y);
|
||||||
|
if (B) {
|
||||||
|
#ifdef PADDLE_WITH_MKLML
|
||||||
|
#pragma omp parallel for if (FLAGS_paddle_num_threads > 1)
|
||||||
|
#endif
|
||||||
|
for (int i = 0; i < M; i++) {
|
||||||
|
blas.AXPY(N, static_cast<T>(1), B, Y + i * N);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
} // namespace math
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
@ -0,0 +1,151 @@
|
|||||||
|
# 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 __future__ import print_function
|
||||||
|
|
||||||
|
import unittest
|
||||||
|
import numpy as np
|
||||||
|
from op_test import OpTest
|
||||||
|
from test_lstm_op import lstm, ACTIVATION
|
||||||
|
|
||||||
|
|
||||||
|
def fc(x, w, b):
|
||||||
|
return np.dot(x, w) + b
|
||||||
|
|
||||||
|
|
||||||
|
def fusion_lstm(
|
||||||
|
x, # T x M
|
||||||
|
lod, # 1 x N
|
||||||
|
wx=None, # M x 4D
|
||||||
|
bx=None, # 1 x 4D
|
||||||
|
h0=None, # N x D
|
||||||
|
c0=None, # N x D
|
||||||
|
w_h=None, # D x 4D
|
||||||
|
w_b=None, # 1 x 4D
|
||||||
|
w_c=None, # 1 x 3D
|
||||||
|
is_reverse=False,
|
||||||
|
act_gate=None,
|
||||||
|
act_cell=None,
|
||||||
|
act_cand=None):
|
||||||
|
return lstm(
|
||||||
|
fc(x, wx, bx), lod, h0, c0, w_h, w_b, w_c, is_reverse, act_gate,
|
||||||
|
act_cell, act_cand)
|
||||||
|
|
||||||
|
|
||||||
|
class TestLstmOp(OpTest):
|
||||||
|
def set_argument(self):
|
||||||
|
self.lod = [[2, 3, 2]]
|
||||||
|
|
||||||
|
def setUp(self):
|
||||||
|
self.op_type = 'fusion_lstm'
|
||||||
|
self.lod = [[2, 3, 2]]
|
||||||
|
self.M = 8
|
||||||
|
self.D = 16
|
||||||
|
self.has_initial_state = False
|
||||||
|
self.is_reverse = False
|
||||||
|
self.act_gate = 'sigmoid'
|
||||||
|
self.act_cell = 'tanh'
|
||||||
|
self.act_cand = 'tanh'
|
||||||
|
self.use_peepholes = False
|
||||||
|
self.set_argument()
|
||||||
|
|
||||||
|
T = sum(self.lod[0])
|
||||||
|
bs = len(self.lod[0])
|
||||||
|
|
||||||
|
x = np.random.normal(size=(T, self.M)).astype('float64')
|
||||||
|
if self.has_initial_state:
|
||||||
|
h0 = np.random.normal(size=(bs, self.D)).astype('float64')
|
||||||
|
c0 = np.random.normal(size=(bs, self.D)).astype('float64')
|
||||||
|
else:
|
||||||
|
h0 = np.zeros((bs, self.D)).astype('float64')
|
||||||
|
c0 = np.zeros((bs, self.D)).astype('float64')
|
||||||
|
|
||||||
|
wh = np.random.normal(size=(self.D, 4 * self.D)).astype('float64')
|
||||||
|
|
||||||
|
if self.use_peepholes:
|
||||||
|
b = np.random.normal(size=(1, 7 * self.D)).astype('float64')
|
||||||
|
else:
|
||||||
|
b = np.random.normal(size=(1, 4 * self.D)).astype('float64')
|
||||||
|
w_b = np.copy(b[:, 0:4 * self.D])
|
||||||
|
w_c = b[:, 4 * self.D:] if self.use_peepholes else None
|
||||||
|
|
||||||
|
# this is the weight of fc
|
||||||
|
wx = np.random.normal(size=(self.M, 4 * self.D)).astype('float64')
|
||||||
|
# this is the bias of fc
|
||||||
|
# and it should be manually added into the bias of this fusion LSTM
|
||||||
|
bx = np.random.normal(size=(1, 4 * self.D)).astype('float64')
|
||||||
|
b[0, 0:4 * self.D] += bx[0, :]
|
||||||
|
h, c = fusion_lstm(x, self.lod, wx, bx, h0, c0, wh, w_b, w_c,
|
||||||
|
self.is_reverse, ACTIVATION[self.act_gate],
|
||||||
|
ACTIVATION[self.act_cell], ACTIVATION[self.act_cand])
|
||||||
|
|
||||||
|
self.inputs = {
|
||||||
|
'X': (x, self.lod),
|
||||||
|
'WeightX': wx,
|
||||||
|
'WeightH': wh,
|
||||||
|
'Bias': b
|
||||||
|
}
|
||||||
|
|
||||||
|
if self.has_initial_state:
|
||||||
|
self.inputs['H0'] = h0
|
||||||
|
self.inputs['C0'] = c0
|
||||||
|
|
||||||
|
self.outputs = {
|
||||||
|
'Hidden': (h, self.lod),
|
||||||
|
'Cell': (c, self.lod),
|
||||||
|
}
|
||||||
|
self.attrs = {
|
||||||
|
'use_peepholes': self.use_peepholes,
|
||||||
|
'is_reverse': self.is_reverse,
|
||||||
|
'gate_activation': self.act_gate,
|
||||||
|
'cell_activation': self.act_cell,
|
||||||
|
'candidate_activation': self.act_cand
|
||||||
|
}
|
||||||
|
|
||||||
|
def test_check_output(self):
|
||||||
|
self.check_output(atol=1e-8)
|
||||||
|
|
||||||
|
|
||||||
|
class TestLstmOpInitReverse(TestLstmOp):
|
||||||
|
def set_argument(self):
|
||||||
|
self.has_initial_state = True
|
||||||
|
self.is_reverse = True
|
||||||
|
|
||||||
|
|
||||||
|
class TestLstmOpMD1(TestLstmOp):
|
||||||
|
def set_argument(self):
|
||||||
|
self.M = 36
|
||||||
|
self.D = 8
|
||||||
|
|
||||||
|
|
||||||
|
class TestLstmOpMD2(TestLstmOp):
|
||||||
|
def set_argument(self):
|
||||||
|
self.M = 8
|
||||||
|
self.D = 8
|
||||||
|
|
||||||
|
|
||||||
|
class TestLstmOpMD3(TestLstmOp):
|
||||||
|
def set_argument(self):
|
||||||
|
self.M = 15
|
||||||
|
self.D = 3
|
||||||
|
|
||||||
|
|
||||||
|
class TestLstmOpBS1(TestLstmOp):
|
||||||
|
def set_argument(self):
|
||||||
|
self.lod = [[3]]
|
||||||
|
self.D = 16
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == '__main__':
|
||||||
|
unittest.main()
|
Loading…
Reference in new issue