warpctc_op.h 17.4 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

namespace paddle {
namespace operators {

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

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
template <typename DeviceContext, typename T>
class ComputeCtcLossFunctor {
 public:
  ctcStatus_t operator()(const T* const activations, T* gradients,
                         const int* const flat_labels,
                         const int* const label_lengths,
                         const int* const input_lengths, int alphabet_size,
                         int minibatch, T* costs, void* workspace,
                         ctcOptions options) {
    return CTC_STATUS_EXECUTION_FAILED;
  }
};

template <typename DeviceContext>
class ComputeCtcLossFunctor<DeviceContext, float> {
 public:
  ctcStatus_t operator()(const float* const activations, float* gradients,
                         const int* const flat_labels,
                         const int* const label_lengths,
                         const int* const input_lengths, int alphabet_size,
                         int minibatch, float* costs, void* workspace,
                         ctcOptions options) {
    return platform::dynload::compute_ctc_loss(
        activations, gradients, flat_labels, label_lengths, input_lengths,
        static_cast<int>(alphabet_size), static_cast<int>(minibatch), costs,
        workspace, options);
  }
};

Y
Yiqun Liu 已提交
59
template <typename DeviceContext>
60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
class ComputeCtcLossFunctor<DeviceContext, double> {
 public:
  ctcStatus_t operator()(const double* const activations, double* gradients,
                         const int* const flat_labels,
                         const int* const label_lengths,
                         const int* const input_lengths, int alphabet_size,
                         int minibatch, double* costs, void* workspace,
                         ctcOptions options) {
    return platform::dynload::compute_ctc_loss_double(
        activations, gradients, flat_labels, label_lengths, input_lengths,
        static_cast<int>(alphabet_size), static_cast<int>(minibatch), costs,
        workspace, options);
  }
};

template <typename DeviceContext, typename T>
Y
Yiqun Liu 已提交
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
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.
   */
99 100
  void operator()(const framework::ExecutionContext& ctx, const T* input,
                  T* gradient, const int* cpu_labels,
Y
Yiqun Liu 已提交
101 102
                  const int* cpu_label_lengths, const int* cpu_input_lengths,
                  const size_t sequence_width, const size_t num_sequences,
103
                  const size_t blank, T* cpu_loss) {
Y
Yiqun Liu 已提交
104 105 106 107 108 109
    // 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;
110 111 112 113 114 115 116 117 118 119 120 121
    ctcStatus_t status = CTC_STATUS_UNKNOWN_ERROR;
    if (sizeof(T) == 4) {
      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);
    } else {
      status = platform::dynload::get_workspace_size_double(
          cpu_label_lengths, cpu_input_lengths,
          static_cast<int>(sequence_width), static_cast<int>(num_sequences),
          options_, &workspace_bytes);
    }
122 123 124 125 126 127 128 129 130 131 132
    PADDLE_ENFORCE_EQ(
        CTC_STATUS_SUCCESS, status,
        platform::errors::PreconditionNotMet(
            "warp-ctc [version %d] Error in get_workspace_size: %s",
            warpctc_version_, platform::dynload::ctcGetStatusString(status)));
    PADDLE_ENFORCE_GT(
        workspace_bytes, 0UL,
        platform::errors::InvalidArgument(
            "Bytes of workspace got by warp-ctc function, "
            "get_workspace_size() should be larger than 0, but received %d",
            workspace_bytes));
Y
Yiqun Liu 已提交
133

L
Li Fuchen 已提交
134
    auto& dev_ctx = ctx.template device_context<DeviceContext>();
135 136
    size_t workspace_elements = workspace_bytes / sizeof(T) + 1UL;
    Tensor workspace = ctx.AllocateTmpTensor<T, DeviceContext>(
Y
Yiqun Liu 已提交
137
        framework::make_ddim({static_cast<int64_t>(workspace_elements)}),
L
Li Fuchen 已提交
138
        dev_ctx);
139 140
    T* workspace_data = workspace.data<T>();
    math::SetConstant<DeviceContext, T>()(
Y
Yiqun Liu 已提交
141
        ctx.template device_context<DeviceContext>(), &workspace,
142
        static_cast<T>(0));
Y
Yiqun Liu 已提交
143 144

    // compute loss and gradient
145
    status = ComputeCtcLossFunctor<DeviceContext, T>()(
Y
Yiqun Liu 已提交
146 147 148
        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_);
149 150 151 152 153 154

    PADDLE_ENFORCE_EQ(
        CTC_STATUS_SUCCESS, status,
        platform::errors::PreconditionNotMet(
            "warp-ctc [version %d] Error in get_workspace_size: %s",
            warpctc_version_, platform::dynload::ctcGetStatusString(status)));
Y
Yiqun Liu 已提交
155 156 157 158 159 160 161
  }

 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())) {
162
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
Y
Yiqun Liu 已提交
163 164 165 166 167
      options_.loc = CTC_GPU;
      options_.stream = reinterpret_cast<const platform::CUDADeviceContext&>(
                            ctx.device_context())
                            .stream();
#else
168 169
      PADDLE_THROW(platform::errors::PreconditionNotMet(
          "[warpctc init] GPU is not enabled."));
Y
Yiqun Liu 已提交
170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192
#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");

193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212
    size_t num_sequences, sequence_width, max_sequence_length;
    framework::Vector<size_t> logits_lod;
    framework::Vector<size_t> label_lod;

    if (ctx.HasInput("LogitsLength") && ctx.HasInput("LabelLength")) {
      num_sequences = logits->dims()[1];
      sequence_width = logits->dims()[2];
      max_sequence_length = logits->dims()[0];

      auto* logits_length = ctx.Input<framework::Tensor>("LogitsLength");
      auto* labels_length = ctx.Input<framework::Tensor>("LabelLength");
      framework::Tensor logits_length_cpu;
      framework::Tensor labels_length_cpu;
      framework::TensorCopy(*logits_length, platform::CPUPlace(),
                            &logits_length_cpu);
      framework::TensorCopy(*labels_length, platform::CPUPlace(),
                            &labels_length_cpu);

      logits_lod.push_back(0);
      label_lod.push_back(0);
213
      for (size_t i = 0; i < num_sequences; i++) {
214 215 216 217 218 219
        logits_lod.push_back(logits_lod[i] +
                             logits_length_cpu.data<int64_t>()[i]);
        label_lod.push_back(label_lod[i] +
                            labels_length_cpu.data<int64_t>()[i]);
      }
    } else {
220 221 222 223 224 225 226 227 228
      PADDLE_ENFORCE_GT(logits->NumLevels(), 0UL,
                        platform::errors::InvalidArgument(
                            "Input(Logits) Tensor of WarpCTC "
                            "does not contain LoD information."));
      PADDLE_ENFORCE_GT(label->NumLevels(), 0UL,
                        platform::errors::InvalidArgument(
                            "Input(Label) Tensor of WarpCTC "
                            "does not contain LoD information."));

229 230
      logits_lod = framework::ToAbsOffset(logits->lod())[0];
      auto logits_dims = logits->dims();
231

232 233
      PADDLE_ENFORCE_EQ(
          logits_dims[0], static_cast<int64_t>(logits_lod.back()),
234 235 236 237
          platform::errors::InvalidArgument(
              "The first dimension of Input(Logits) should be equal to "
              "the sum of all sequences' lengths = %d., but received %d. ",
              static_cast<int64_t>(logits_lod.back()), logits_dims[0]));
238 239 240

      label_lod = framework::ToAbsOffset(label->lod())[0];
      auto label_dims = label->dims();
241 242 243 244 245
      PADDLE_ENFORCE_EQ(label_dims[1], 1,
                        platform::errors::InvalidArgument(
                            "The last dimension of Input(Label) should be 1, "
                            "but received %d",
                            label_dims[1]));
246 247

      num_sequences = logits_lod.size() - 1;
248 249 250 251 252 253
      PADDLE_ENFORCE_EQ(
          num_sequences, label_lod.size() - 1,
          platform::errors::InvalidArgument(
              "The number of sequences of Input(Logits) should be "
              "equal to that of Input(Label) = %d, but received %d",
              label_lod.size() - 1, num_sequences));
254 255 256 257 258

      sequence_width = logits->numel() / logits_dims[0];
      max_sequence_length = math::MaximumSequenceLength(logits_lod);
    }

Y
Yiqun Liu 已提交
259 260 261 262
    auto loss_dims =
        framework::make_ddim({static_cast<int64_t>(num_sequences), 1});

    // warpctc needs sequences data stored in transposed padding format
F
fengjiayi 已提交
263
    LoDTensor warpctc_logits;
Y
Yiqun Liu 已提交
264 265 266 267
    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)});
L
Li Fuchen 已提交
268 269 270 271
    auto& dev_ctx = ctx.template device_context<DeviceContext>();
    Tensor warpctc_logits_tmp =
        ctx.AllocateTmpTensor<T, DeviceContext>(warpctc_logits_dims, dev_ctx);
    warpctc_logits.ShareDataWith(warpctc_logits_tmp);
272 273
    if (ctx.HasInput("LogitsLength")) {
      TensorCopySync(*logits, ctx.GetPlace(), &warpctc_logits);
F
fengjiayi 已提交
274
    } else {
275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
      LoDTensor cpu_pad_value;
      T* pad_value_data =
          cpu_pad_value.mutable_data<T>({1}, platform::CPUPlace());
      *pad_value_data = static_cast<T>(0);
      LoDTensor pad_value;
      if (platform::is_cpu_place(ctx.GetPlace())) {
        pad_value = cpu_pad_value;
      } else {
        TensorCopySync(cpu_pad_value, ctx.GetPlace(), &pad_value);
      }

      math::PaddingLoDTensorFunctor<DeviceContext, T>()(
          ctx.template device_context<DeviceContext>(), *logits,
          &warpctc_logits, pad_value, -1, 0, false /* norm_by_times */,
          math::kLengthBatchWidth);
F
fengjiayi 已提交
290
    }
Y
Yiqun Liu 已提交
291 292 293 294 295 296
    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) {
297 298
      warpctc_label_lengths[i] = label_lod[i + 1] - label_lod[i];
      warpctc_logits_lengths[i] = logits_lod[i + 1] - logits_lod[i];
Y
Yiqun Liu 已提交
299 300 301 302 303 304 305
    }

    // 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());

306 307 308 309
    math::SetConstant<DeviceContext, T>()(
        ctx.template device_context<DeviceContext>(), warpctc_grad,
        static_cast<T>(0));

Y
Yiqun Liu 已提交
310
    // warpctc accesses labels in CPU memory
W
whs 已提交
311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339
    LoDTensor warpctc_label;
    if (ctx.HasInput("LogitsLength")) {
      warpctc_label.mutable_data<int>(
          {static_cast<int64_t>(math::TotalSequenceLength(label_lod)), 1},
          platform::CPUPlace());
      std::vector<framework::Vector<size_t>> lod;
      lod.push_back(label_lod);
      warpctc_label.set_lod(lod);

      if (platform::is_cpu_place(ctx.GetPlace())) {
        math::UnpaddingLoDTensorFunctor<DeviceContext, int>()(
            ctx.template device_context<DeviceContext>(), *label,
            &warpctc_label, label->dims()[1] /*pad_seq_len*/, 0 /*lod_level*/,
            false /*norm_by_times*/, math::kBatchLengthWidth);
      } else {
        LoDTensor gpu_label;
        gpu_label.mutable_data<int>(
            {static_cast<int64_t>(math::TotalSequenceLength(label_lod)), 1},
            ctx.GetPlace());
        gpu_label.set_lod(lod);
        math::UnpaddingLoDTensorFunctor<DeviceContext, int>()(
            ctx.template device_context<DeviceContext>(), *label, &gpu_label,
            label->dims()[1] /*pad_seq_len*/, 0 /*lod_level*/,
            false /*norm_by_times*/, math::kBatchLengthWidth);
        TensorCopySync(gpu_label, platform::CPUPlace(), &warpctc_label);
      }
    } else {
      TensorCopySync(*label, platform::CPUPlace(), &warpctc_label);
    }
340

Y
Yiqun Liu 已提交
341 342 343 344 345 346 347 348
    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"));

349
    WarpCTCFunctor<DeviceContext, T>()(
Y
Yiqun Liu 已提交
350 351 352 353 354
        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
355
    TensorCopy(warpctc_loss, ctx.GetPlace(), ctx.device_context(), loss);
Y
Yiqun Liu 已提交
356 357 358 359 360 361 362
  }
};

template <typename DeviceContext, typename T>
class WarpCTCGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
F
fengjiayi 已提交
363
    auto* warpctc_grad = ctx.Input<LoDTensor>("WarpCTCGrad");
Y
Yiqun Liu 已提交
364
    auto* logits_grad = ctx.Output<LoDTensor>(framework::GradVarName("Logits"));
W
wanghaoshuang 已提交
365 366 367
    const Tensor* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));

    logits_grad->mutable_data<T>(ctx.GetPlace());
Y
Yiqun Liu 已提交
368
    bool norm_by_times = ctx.Attr<bool>("norm_by_times");
369 370 371 372 373 374

    if (ctx.HasInput("LogitsLength")) {
      size_t max_seq_length = warpctc_grad->dims()[0];
      size_t num_sequences = warpctc_grad->dims()[1];
      size_t seq_width = warpctc_grad->dims()[2];

375 376 377 378 379
      auto* logits_length = ctx.Input<framework::Tensor>("LogitsLength");
      framework::Tensor logits_length_cpu;
      framework::TensorCopy(*logits_length, platform::CPUPlace(),
                            &logits_length_cpu);

380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401
      LoDTensor logits_grad_with_lod;
      auto logits_grad_dims =
          framework::make_ddim({static_cast<int64_t>(max_seq_length),
                                static_cast<int64_t>(num_sequences),
                                static_cast<int64_t>(seq_width)});
      T* logits_grad_cpu_data = logits_grad_with_lod.mutable_data<T>(
          logits_grad_dims, platform::CPUPlace());

      TensorCopySync(*warpctc_grad, platform::CPUPlace(),
                     &logits_grad_with_lod);

      Tensor loss_grad_cpu;
      loss_grad_cpu.mutable_data<T>(loss_grad->dims(), platform::CPUPlace());
      TensorCopySync(*loss_grad, platform::CPUPlace(), &loss_grad_cpu);

      LoDTensor scaled_logits;
      T* scaled_logits_data =
          scaled_logits.mutable_data<T>(logits_grad_dims, platform::CPUPlace());

      const T* loss_grad_data = loss_grad_cpu.data<T>();
      for (size_t i = 0; i < max_seq_length; ++i) {
        for (size_t j = 0; j < num_sequences; ++j) {
402 403 404 405
          T scale = 1.0;
          if (norm_by_times) {
            scale = 1.0 / static_cast<T>(logits_length_cpu.data<int64_t>()[j]);
          }
406 407 408
          for (size_t k = 0; k < seq_width; ++k) {
            size_t idx = i * (num_sequences * seq_width) + j * seq_width + k;
            scaled_logits_data[idx] =
409
                logits_grad_cpu_data[idx] * loss_grad_data[j] * scale;
410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
          }
        }
      }

      TensorCopySync(scaled_logits, ctx.GetPlace(), logits_grad);
    } else {
      math::UnpaddingLoDTensorFunctor<DeviceContext, T>()(
          ctx.template device_context<DeviceContext>(), *warpctc_grad,
          logits_grad, -1, 0, norm_by_times, math::kLengthBatchWidth);

      const T* loss_grad_data = loss_grad->data<T>();
      math::ScaleLoDTensorFunctor<DeviceContext, T>()(
          ctx.template device_context<DeviceContext>(), loss_grad_data,
          logits_grad);
    }
Y
Yiqun Liu 已提交
425 426 427 428 429
  }
};

}  // namespace operators
}  // namespace paddle