bce_loss_op.cu 4.1 KB
Newer Older
C
ceci3 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28
/* Copyright (c) 2020 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 <algorithm>
#include "paddle/fluid/operators/bce_loss_op.h"
#include "paddle/fluid/operators/math.h"
#include "paddle/fluid/platform/cuda_primitives.h"
#include "paddle/fluid/platform/gpu_launch_config.h"
#include "paddle/fluid/platform/hostdevice.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;

template <typename T>
__global__ void GPUBCELossForward(const T* x_data, const T* label_data,
                                  T* out_data, const int in_numel) {
29
  CUDA_KERNEL_LOOP(i, in_numel) {
C
ceci3 已提交
30 31 32 33 34
    T x = x_data[i];
    T label = label_data[i];
    T one = static_cast<T>(1.);
    T neg_100 = static_cast<T>(-100.);

35 36 37 38 39
    PADDLE_ENFORCE(
        (x >= static_cast<T>(0)) && (x <= one),
        "Input is expected to be within the interval [0, 1], but recieved %f.",
        x);

C
ceci3 已提交
40 41 42 43 44 45 46 47 48 49 50
    T term1 = max(real_log(x), neg_100);
    T term2 = max(real_log(one - x), neg_100);

    out_data[i] = ((label - one) * term2) - (label * term1);
  }
}

template <typename T>
__global__ void GPUBCELossBackward(const T* x_data, const T* label_data,
                                   const T* dout_data, T* dx_data,
                                   const int in_numel) {
51
  CUDA_KERNEL_LOOP(i, in_numel) {
C
ceci3 已提交
52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
    T x = x_data[i];
    T label = label_data[i];
    T dout = dout_data[i];
    T one = static_cast<T>(1.);
    T eps = static_cast<T>(1e-12);

    T term1 = max((one - x) * x, eps);

    dx_data[i] = dout * (x - label) / term1;
  }
}

template <typename DeviceContext, typename T>
class BCELossCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* x = ctx.Input<Tensor>("X");
    auto* labels = ctx.Input<Tensor>("Label");
    auto* out = ctx.Output<Tensor>("Out");

72 73
    const auto* x_data = x->data<T>();
    auto* out_data = out->mutable_data<T>(ctx.GetPlace());
74 75
    auto x_numel = x->numel();

C
ceci3 已提交
76
    auto& dev_ctx = ctx.cuda_device_context();
77 78
    platform::GpuLaunchConfig config =
        platform::GetGpuLaunchConfig1D(dev_ctx, x_numel);
C
ceci3 已提交
79

80 81 82
    GPUBCELossForward<T><<<config.block_per_grid, config.thread_per_block, 0,
                           dev_ctx.stream()>>>(x_data, labels->data<T>(),
                                               out_data, x_numel);
C
ceci3 已提交
83 84 85 86 87 88 89 90 91 92 93 94 95
  }
};

template <typename DeviceContext, typename T>
class BCELossGradCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* x = ctx.Input<Tensor>("X");
    auto* labels = ctx.Input<Tensor>("Label");
    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));

    int x_numel = x->numel();
96 97
    auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());

C
ceci3 已提交
98
    auto& dev_ctx = ctx.cuda_device_context();
99 100
    platform::GpuLaunchConfig config =
        platform::GetGpuLaunchConfig1D(dev_ctx, x_numel);
C
ceci3 已提交
101

102 103
    GPUBCELossBackward<T><<<config.block_per_grid, config.thread_per_block, 0,
                            dev_ctx.stream()>>>(
C
ceci3 已提交
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
        x->data<T>(), labels->data<T>(), dout->data<T>(), dx_data, x_numel);
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
    bce_loss,
    ops::BCELossCUDAKernel<paddle::platform::CUDADeviceContext, float>,
    ops::BCELossCUDAKernel<paddle::platform::CUDADeviceContext, double>);
REGISTER_OP_CUDA_KERNEL(
    bce_loss_grad,
    ops::BCELossGradCUDAKernel<paddle::platform::CUDADeviceContext, float>,
    ops::BCELossGradCUDAKernel<paddle::platform::CUDADeviceContext, double>);