warpctc_op.h 9.4 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Y
Yiqun Liu 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

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

17
#include <vector>
Y
Yi Wang 已提交
18 19 20 21 22
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/sequence_padding.h"
#include "paddle/fluid/operators/math/sequence_scale.h"
#include "paddle/fluid/platform/dynload/warpctc.h"
Y
Yiqun Liu 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;

template <typename DeviceContext>
class WarpCTCFunctor {
 public:
  /*
   * \brief Compute the connectionist temporal classification loss,
   *        and optionally compute the gradient with respect to the inputs.
   *
   * If gradient is nullptr, it only computes the ctc loss,
   * or computes both ctc loss and gradient.
   *
   * \param ctx               execution context of this functor
   * \param input             batch matrix of input probabilities, in
   *                          max_sequence_length x num_sequences x
   *                          sequence_width, (row-major) format
   * \param gradient          batch matrix of gradient, with the same shape as
   *                          input.
   * \param cpu_labels        labels always in CPU memory.
   * \param cpu_label_lengths length of all labels in CPU memory.
   * \param cpu_input_lengths length of all sequences in CPU memory.
   * \param sequence_width    number of possible output symbols.
   * \param num_sequences     number of sequence.
   * \param blank             blank label used in ctc loss function.
   * \param cpu_losss         cost of each sequence in CPU memory.
   */
  void operator()(const framework::ExecutionContext& ctx, const float* input,
                  float* gradient, const int* cpu_labels,
                  const int* cpu_label_lengths, const int* cpu_input_lengths,
                  const size_t sequence_width, const size_t num_sequences,
                  const size_t blank, float* cpu_loss) {
    // Init warp-ctc options
    init(ctx, blank);

    // Compute the required workspace size.
    // There is no memory allocated operations within warp-ctc.
    size_t workspace_bytes = 0;
    ctcStatus_t status = platform::dynload::get_workspace_size(
        cpu_label_lengths, cpu_input_lengths, static_cast<int>(sequence_width),
        static_cast<int>(num_sequences), options_, &workspace_bytes);
    PADDLE_ENFORCE_EQ(CTC_STATUS_SUCCESS, status,
                      "warp-ctc [version %d] Error in get_workspace_size: ",
                      warpctc_version_,
                      platform::dynload::ctcGetStatusString(status));
    PADDLE_ENFORCE_GT(workspace_bytes, 0UL,
                      "Bytes of workspace got by warp-ctc function, "
                      "get_workspace_size(), should be larger than 0.");

    Tensor workspace;
    size_t workspace_elements = workspace_bytes / sizeof(float) + 1UL;
    float* workspace_data = workspace.mutable_data<float>(
        framework::make_ddim({static_cast<int64_t>(workspace_elements)}),
        ctx.GetPlace());
    math::SetConstant<DeviceContext, float>()(
        ctx.template device_context<DeviceContext>(), &workspace,
        static_cast<float>(0));

    // compute loss and gradient
    status = platform::dynload::compute_ctc_loss(
        input, gradient, cpu_labels, cpu_label_lengths, cpu_input_lengths,
        static_cast<int>(sequence_width), static_cast<int>(num_sequences),
        cpu_loss, workspace_data, options_);
    PADDLE_ENFORCE_EQ(CTC_STATUS_SUCCESS, status,
                      "warp-ctc [version %d] Error in compute_ctc_loss: ",
                      warpctc_version_,
                      platform::dynload::ctcGetStatusString(status));
  }

 protected:
  void init(const framework::ExecutionContext& ctx, const size_t blank) {
    warpctc_version_ = platform::dynload::get_warpctc_version();

    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
      options_.loc = CTC_GPU;
      options_.stream = reinterpret_cast<const platform::CUDADeviceContext&>(
                            ctx.device_context())
                            .stream();
#else
      PADDLE_THROW("[warpctc init] GPU is not enabled.");
#endif
    } else {
      options_.loc = CTC_CPU;
      options_.num_threads = 1;
    }

    options_.blank_label = blank;
  }

 private:
  int warpctc_version_;
  ctcOptions options_;
};

template <typename DeviceContext, typename T>
class WarpCTCKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* logits = ctx.Input<LoDTensor>("Logits");
    auto* label = ctx.Input<LoDTensor>("Label");
    auto* warpctc_grad = ctx.Output<Tensor>("WarpCTCGrad");
    auto* loss = ctx.Output<Tensor>("Loss");

    const size_t level = 0;

    auto logits_lod = framework::ToAbsOffset(logits->lod());
    auto logits_dims = logits->dims();
    PADDLE_ENFORCE_EQ(logits_dims[0],
                      static_cast<int64_t>(logits_lod[level].back()),
                      "The first dimension of Input(Logits) should be equal to "
                      "the sum of all sequences' lengths.");

    auto label_lod = framework::ToAbsOffset(label->lod());
    auto label_dims = label->dims();
    PADDLE_ENFORCE_EQ(
        label_dims[0], label->numel(),
        "The width of each timestep in Input(Label) should be 1.");

    const size_t num_sequences = logits_lod[level].size() - 1;
    PADDLE_ENFORCE_EQ(num_sequences, label_lod[level].size() - 1,
                      "The number of sequences of Input(Logits) should be "
                      "equal to that of Input(Label).");

    const size_t sequence_width = logits->numel() / logits_dims[0];
    auto loss_dims =
        framework::make_ddim({static_cast<int64_t>(num_sequences), 1});

    // warpctc needs sequences data stored in transposed padding format
    Tensor warpctc_logits;
    const size_t max_sequence_length =
        math::MaximumSequenceLength(logits_lod, level);
    auto warpctc_logits_dims =
        framework::make_ddim({static_cast<int64_t>(max_sequence_length),
                              static_cast<int64_t>(num_sequences),
                              static_cast<int64_t>(sequence_width)});
    warpctc_logits.mutable_data<T>(warpctc_logits_dims, ctx.GetPlace());
164
    math::PaddingLoDTensorFunctor<DeviceContext, T, math::LENGTH_BATCH_WIDTH>()(
165
        ctx.template device_context<DeviceContext>(), *logits, &warpctc_logits,
Y
Yiqun Liu 已提交
166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
        false);
    const T* warpctc_logits_data = warpctc_logits.data<T>();

    std::vector<int> warpctc_label_lengths(num_sequences);
    std::vector<int> warpctc_logits_lengths(num_sequences);

    for (size_t i = 0; i < num_sequences; ++i) {
      warpctc_label_lengths[i] = label_lod[level][i + 1] - label_lod[level][i];
      warpctc_logits_lengths[i] =
          logits_lod[level][i + 1] - logits_lod[level][i];
    }

    // warpctc computes loss and gradient in one call, gradient data also stored
    // in batch format
    T* warpctc_grad_data =
        warpctc_grad->mutable_data<T>(warpctc_logits.dims(), ctx.GetPlace());

183 184 185 186
    math::SetConstant<DeviceContext, T>()(
        ctx.template device_context<DeviceContext>(), warpctc_grad,
        static_cast<T>(0));

Y
Yiqun Liu 已提交
187 188
    // warpctc accesses labels in CPU memory
    Tensor warpctc_label;
Y
Yi Wang 已提交
189 190
    TensorCopy(*label, platform::CPUPlace(), ctx.device_context(),
               &warpctc_label);
Y
Yiqun Liu 已提交
191 192 193 194 195 196 197 198 199 200 201 202 203 204
    const int* warpctc_label_data = warpctc_label.data<int>();
    // warpctc stores loss in CPU memory
    Tensor warpctc_loss;
    T* warpctc_loss_data =
        warpctc_loss.mutable_data<T>(loss_dims, platform::CPUPlace());

    const size_t blank = static_cast<size_t>(ctx.Attr<int>("blank"));

    WarpCTCFunctor<DeviceContext>()(
        ctx, warpctc_logits_data, warpctc_grad_data, warpctc_label_data,
        warpctc_label_lengths.data(), warpctc_logits_lengths.data(),
        sequence_width, num_sequences, blank, warpctc_loss_data);

    // Copy the loss back
Y
Yi Wang 已提交
205
    TensorCopy(warpctc_loss, ctx.GetPlace(), ctx.device_context(), loss);
Y
Yiqun Liu 已提交
206 207 208 209 210 211 212 213 214
  }
};

template <typename DeviceContext, typename T>
class WarpCTCGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* warpctc_grad = ctx.Input<Tensor>("WarpCTCGrad");
    auto* logits_grad = ctx.Output<LoDTensor>(framework::GradVarName("Logits"));
W
wanghaoshuang 已提交
215 216 217
    const Tensor* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));

    logits_grad->mutable_data<T>(ctx.GetPlace());
Y
Yiqun Liu 已提交
218
    bool norm_by_times = ctx.Attr<bool>("norm_by_times");
219 220
    math::UnpaddingLoDTensorFunctor<DeviceContext, T,
                                    math::LENGTH_BATCH_WIDTH>()(
221
        ctx.template device_context<DeviceContext>(), logits_grad,
Y
Yiqun Liu 已提交
222
        *warpctc_grad, norm_by_times);
W
wanghaoshuang 已提交
223 224 225

    const T* loss_grad_data = loss_grad->data<T>();
    math::ScaleLoDTensorFunctor<DeviceContext, T>()(
226 227
        ctx.template device_context<DeviceContext>(), loss_grad_data,
        logits_grad);
Y
Yiqun Liu 已提交
228 229 230 231 232
  }
};

}  // namespace operators
}  // namespace paddle