87 lines
3.3 KiB
C++
87 lines
3.3 KiB
C++
// Copyright (c) 2022 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 <vector>
|
|
|
|
#include "paddle/phi/backends/dynload/warpctc.h"
|
|
#include "paddle/phi/core/dense_tensor.h"
|
|
#include "paddle/phi/kernels/empty_kernel.h"
|
|
#include "paddle/phi/kernels/funcs/eigen/common.h"
|
|
#include "paddle/phi/kernels/funcs/math_function.h"
|
|
#include "paddle/phi/kernels/funcs/sequence_padding.h"
|
|
#include "paddle/phi/kernels/funcs/sequence_scale.h"
|
|
#include "paddle/utils/optional.h"
|
|
|
|
namespace phi {
|
|
|
|
template <typename T, typename Context>
|
|
void WarpctcGradKernel(const Context& dev_ctx,
|
|
const DenseTensor& logits UNUSED,
|
|
const optional<DenseTensor>& logits_length,
|
|
const DenseTensor& warpctcgrad,
|
|
const DenseTensor& loss_grad,
|
|
int blank UNUSED,
|
|
bool norm_by_times,
|
|
DenseTensor* logits_grad) {
|
|
dev_ctx.template Alloc<T>(logits_grad);
|
|
|
|
if (logits_length.is_initialized()) {
|
|
int64_t max_seq_length = warpctcgrad.dims()[0]; // Tmax
|
|
int64_t num_sequences = warpctcgrad.dims()[1]; // B
|
|
int64_t seq_width = warpctcgrad.dims()[2]; // D
|
|
|
|
// B
|
|
auto logits_len_e = EigenTensor<int64_t, 1>::From(*logits_length);
|
|
// (B, 1)
|
|
auto loss_grad_e = EigenTensor<T, 2>::From(loss_grad);
|
|
// (T, B, D)
|
|
auto warpctcgrad_e = EigenTensor<T, 3>::From(warpctcgrad);
|
|
|
|
auto logits_grad_e = EigenTensor<T, 3>::From(*logits_grad);
|
|
|
|
Eigen::DSizes<int64_t, 3> grad_shape(1, num_sequences, 1);
|
|
Eigen::DSizes<int64_t, 3> bcast(max_seq_length, 1, seq_width);
|
|
auto logits_g =
|
|
warpctcgrad_e * loss_grad_e.reshape(grad_shape).broadcast(bcast).eval();
|
|
|
|
auto* place = dev_ctx.eigen_device();
|
|
if (norm_by_times) {
|
|
auto scales = logits_len_e.cast<T>()
|
|
.inverse()
|
|
.reshape(grad_shape)
|
|
.broadcast(bcast)
|
|
.eval();
|
|
logits_grad_e.device(*place) = logits_g * scales;
|
|
} else {
|
|
logits_grad_e.device(*place) = logits_g;
|
|
}
|
|
} else {
|
|
funcs::UnpaddingDenseTensorFunctor<Context, T>()(dev_ctx,
|
|
warpctcgrad,
|
|
logits_grad,
|
|
-1,
|
|
0,
|
|
norm_by_times,
|
|
funcs::kLengthBatchWidth);
|
|
|
|
const T* loss_grad_data = loss_grad.data<T>();
|
|
funcs::ScaleDenseTensorFunctor<Context, T>()(
|
|
dev_ctx, loss_grad_data, logits_grad);
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|