File size: 5,956 Bytes
128757a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
// This file is modified from  https://github.com/pytorch/pytorch/blob/master/modules/detectron/sigmoid_focal_loss_op.cu
// Cheng-Yang Fu
// cyfu@cs.unc.edu
#include <ATen/ATen.h>
#include <ATen/cuda/CUDAContext.h>

#include <THC/THC.h>
#include <THC/THCAtomics.cuh>
#include <THC/THCDeviceUtils.cuh>

#include <cfloat>

// TODO make it in a common file
#define CUDA_1D_KERNEL_LOOP(i, n)                            \
  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
       i += blockDim.x * gridDim.x)


template <typename T>
__global__ void SigmoidFocalLossForward(const int nthreads, 

    const T* logits,

    const int* targets,

    const int num_classes,

    const float gamma, 

    const float alpha,

    const int num, 

    T* losses) {
  CUDA_1D_KERNEL_LOOP(i, nthreads) {

    int n = i / num_classes;
    int d = i % num_classes; // current class[0~79]; 
    int t = targets[n]; // target class [1~80];

    // Decide it is positive or negative case. 
    T c1 = (t == (d+1)); 
    T c2 = (t>=0 & t != (d+1));

    T zn = (1.0 - alpha);
    T zp = (alpha);

    // p = 1. / 1. + expf(-x); p = sigmoid(x)
    T  p = 1. / (1. + expf(-logits[i]));

    // (1-p)**gamma * log(p) where
    T term1 = powf((1. - p), gamma) * logf(max(p, FLT_MIN));

    // p**gamma * log(1-p)
    T term2 = powf(p, gamma) *
            (-1. * logits[i] * (logits[i] >= 0) -   
             logf(1. + expf(logits[i] - 2. * logits[i] * (logits[i] >= 0))));

    losses[i] = 0.0;
    losses[i] += -c1 * term1 * zp;
    losses[i] += -c2 * term2 * zn;

  } // CUDA_1D_KERNEL_LOOP
} // SigmoidFocalLossForward


template <typename T>
__global__ void SigmoidFocalLossBackward(const int nthreads,

                const T* logits,

                const int* targets,

                const T* d_losses,

                const int num_classes,

                const float gamma,

                const float alpha,

                const int num,

                T* d_logits) {
  CUDA_1D_KERNEL_LOOP(i, nthreads) {

    int n = i / num_classes;
    int d = i % num_classes; // current class[0~79]; 
    int t = targets[n]; // target class [1~80], 0 is background;

    // Decide it is positive or negative case. 
    T c1 = (t == (d+1));
    T c2 = (t>=0 & t != (d+1));

    T zn = (1.0 - alpha);
    T zp = (alpha);
    // p = 1. / 1. + expf(-x); p = sigmoid(x)
    T  p = 1. / (1. + expf(-logits[i]));

    // (1-p)**g * (1 - p - g*p*log(p)
    T term1 = powf((1. - p), gamma) *
                      (1. - p - (p * gamma * logf(max(p, FLT_MIN))));

    // (p**g) * (g*(1-p)*log(1-p) - p)
    T term2 = powf(p, gamma) *
                  ((-1. * logits[i] * (logits[i] >= 0) -
                      logf(1. + expf(logits[i] - 2. * logits[i] * (logits[i] >= 0)))) *
                      (1. - p) * gamma - p);
    d_logits[i] = 0.0;
    d_logits[i] += -c1 * term1 * zp;
    d_logits[i] += -c2 * term2 * zn;
    d_logits[i] = d_logits[i] * d_losses[i];

  } // CUDA_1D_KERNEL_LOOP
} // SigmoidFocalLossBackward


at::Tensor SigmoidFocalLoss_forward_cuda(

		const at::Tensor& logits,

                const at::Tensor& targets,

		const int num_classes, 

		const float gamma, 

		const float alpha) {
  AT_ASSERTM(logits.device().is_cuda(), "logits must be a CUDA tensor");
  AT_ASSERTM(targets.device().is_cuda(), "targets must be a CUDA tensor");
  AT_ASSERTM(logits.dim() == 2, "logits should be NxClass");

  const int num_samples = logits.size(0);
	
  auto losses = at::empty({num_samples, logits.size(1)}, logits.options());
  auto losses_size = num_samples * logits.size(1);
  cudaStream_t stream = at::cuda::getCurrentCUDAStream();

  dim3 grid(std::min(THCCeilDiv(losses_size, 512L), 4096L));
  dim3 block(512);

  if (losses.numel() == 0) {
    THCudaCheck(cudaGetLastError());
    return losses;
  }

  AT_DISPATCH_FLOATING_TYPES(logits.scalar_type(), "SigmoidFocalLoss_forward", [&] {
    SigmoidFocalLossForward<scalar_t><<<grid, block, 0, stream>>>(
         losses_size,
         logits.contiguous().data_ptr<scalar_t>(),
	 targets.contiguous().data_ptr<int>(),
         num_classes,
	 gamma,
	 alpha,
	 num_samples,
         losses.data_ptr<scalar_t>());
  });
  THCudaCheck(cudaGetLastError());
  return losses;   
}	


at::Tensor SigmoidFocalLoss_backward_cuda(

		const at::Tensor& logits,

                const at::Tensor& targets,

		const at::Tensor& d_losses,

		const int num_classes, 

		const float gamma, 

		const float alpha) {
  AT_ASSERTM(logits.device().is_cuda(), "logits must be a CUDA tensor");
  AT_ASSERTM(targets.device().is_cuda(), "targets must be a CUDA tensor");
  AT_ASSERTM(d_losses.device().is_cuda(), "d_losses must be a CUDA tensor");

  AT_ASSERTM(logits.dim() == 2, "logits should be NxClass");

  const int num_samples = logits.size(0);
  AT_ASSERTM(logits.size(1) == num_classes, "logits.size(1) should be num_classes");
	
  auto d_logits = at::zeros({num_samples, num_classes}, logits.options());
  auto d_logits_size = num_samples * logits.size(1);
  cudaStream_t stream = at::cuda::getCurrentCUDAStream();

  dim3 grid(std::min(THCCeilDiv(d_logits_size, 512L), 4096L));
  dim3 block(512);

  if (d_logits.numel() == 0) {
    THCudaCheck(cudaGetLastError());
    return d_logits;
  }

  AT_DISPATCH_FLOATING_TYPES(logits.scalar_type(), "SigmoidFocalLoss_backward", [&] {
    SigmoidFocalLossBackward<scalar_t><<<grid, block, 0, stream>>>(
         d_logits_size,
         logits.contiguous().data_ptr<scalar_t>(),
	 targets.contiguous().data_ptr<int>(),
	 d_losses.contiguous().data_ptr<scalar_t>(),
         num_classes,
	 gamma,
	 alpha,
	 num_samples,
         d_logits.data_ptr<scalar_t>());
  });

  THCudaCheck(cudaGetLastError());
  return d_logits;   
}