DeformConv_cpu.cpp 33.1 KB
Newer Older
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
/*!
 ******************* BEGIN Caffe Copyright Notice and Disclaimer
 *****************
 *
 * COPYRIGHT
 *
 * All contributions by the University of California:
 * Copyright (c) 2014-2017 The Regents of the University of California (Regents)
 * All rights reserved.
 *
 * All other contributions:
 * Copyright (c) 2014-2017, the respective contributors
 * All rights reserved.
 *
 * Caffe uses a shared copyright model: each contributor holds copyright over
 * their contributions to Caffe. The project versioning records all such
 * contribution and copyright details. If a contributor wants to further mark
 * their specific copyright on a particular contribution, they should indicate
 * their copyright solely in the commit message of the change when it is
 * committed.
 *
 * LICENSE
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice,
 *this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 * this list of conditions and the following disclaimer in the documentation
 * and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 *AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 *IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE
 *FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 *DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
 *SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
 *CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
 *OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 *OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * CONTRIBUTION AGREEMENT
 *
 * By contributing to the BVLC/caffe repository through pull-request, comment,
 * or otherwise, the contributor releases their content to the
 * license and copyright terms herein.
 *
 ***************** END Caffe Copyright Notice and Disclaimer
 *********************
 *
 * Copyright (c) 2018 Microsoft
 * Licensed under The MIT License [see LICENSE for details]
 * \file modulated_deformable_im2col.cuh
 * \brief Function definitions of converting an image to
 * column matrix based on kernel, padding, dilation, and offset.
 * These functions are mainly used in deformable convolution operators.
 * \ref: https://arxiv.org/abs/1703.06211
 * \author Yuwen Xiong, Haozhi Qi, Jifeng Dai, Xizhou Zhu, Han Hu, Dazhi Cheng
 */

// modified from
// https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda_kernel.cu

// modified from
// https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda.cpp

#include <ATen/ATen.h>
#include <ATen/TensorUtils.h>
#include <TH/TH.h>

#include <cmath>
#include <iostream>
#include <tuple>

const int kMaxParallelImgs = 32;

template <typename scalar_t>
static scalar_t bilinear_interpolate(
    const scalar_t* in,
82
83
    int height,
    int width,
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
    scalar_t h,
    scalar_t w) {
  if (h <= -1 || height <= h || w <= -1 || width <= w) {
    return 0;
  }

  int h_low = floor(h);
  int w_low = floor(w);
  int h_high = h_low + 1;
  int w_high = w_low + 1;

  scalar_t lh = h - h_low;
  scalar_t lw = w - w_low;
  scalar_t hh = 1 - lh, hw = 1 - lw;

  scalar_t v1 = 0;
  if (h_low >= 0 && w_low >= 0)
    v1 = in[h_low * width + w_low];
  scalar_t v2 = 0;
  if (h_low >= 0 && w_high <= width - 1)
    v2 = in[h_low * width + w_high];
  scalar_t v3 = 0;
  if (h_high <= height - 1 && w_low >= 0)
    v3 = in[h_high * width + w_low];
  scalar_t v4 = 0;
  if (h_high <= height - 1 && w_high <= width - 1)
    v4 = in[h_high * width + w_high];

  scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw;

  scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
  return val;
}

template <typename scalar_t>
static void deformable_im2col_kernel(
120
    int n,
121
122
    const scalar_t* input,
    const scalar_t* offset,
123
    const scalar_t* mask,
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
    int height,
    int width,
    int weight_h,
    int weight_w,
    int pad_h,
    int pad_w,
    int stride_h,
    int stride_w,
    int dil_h,
    int dil_w,
    int batch_sz,
    int n_in_channels,
    int n_offset_grps,
    int out_h,
    int out_w,
139
    bool use_mask,
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
    scalar_t* columns) {
  for (int index = 0; index != n; ++index) {
    const int out_x = index % out_w;
    const int out_y = (index / out_w) % out_h;
    const int out_b = (index / (out_w * out_h)) % batch_sz;
    const int in_c = index / (out_w * out_h * batch_sz);
    const int out_c = in_c * weight_h * weight_w;

    int c_per_offset_grp = n_in_channels / n_offset_grps;
    const int grp_idx = in_c / c_per_offset_grp;

    auto columns_ptr = columns +
        (out_c * (batch_sz * out_h * out_w) + out_b * (out_h * out_w) +
         out_y * out_w + out_x);

    auto input_ptr = input +
        (out_b * (n_in_channels * height * width) + in_c * (height * width));

    auto offset_ptr = offset +
        (out_b * n_offset_grps + grp_idx) * 2 * weight_h * weight_w * out_h *
            out_w;

162
163
164
165
166
167
    auto mask_ptr = mask;
    if (use_mask) {
      mask_ptr += (out_b * n_offset_grps + grp_idx) * weight_h * weight_w *
          out_h * out_w;
    }

168
169
    for (int i = 0; i < weight_h; ++i) {
      for (int j = 0; j < weight_w; ++j) {
170
171
172
173
174
175
176
177
178
        const int mask_idx = i * weight_w + j;
        const int offset_idx = 2 * mask_idx;

        scalar_t mask_value = 1;
        if (use_mask) {
          mask_value =
              mask_ptr[mask_idx * (out_h * out_w) + out_y * out_w + out_x];
        }

179
180
181
182
183
184
        const scalar_t offset_h =
            offset_ptr[offset_idx * (out_h * out_w) + out_y * out_w + out_x];
        const scalar_t offset_w = offset_ptr
            [(offset_idx + 1) * (out_h * out_w) + out_y * out_w + out_x];
        const scalar_t y = (out_y * stride_h - pad_h) + i * dil_h + offset_h;
        const scalar_t x = (out_x * stride_w - pad_w) + j * dil_w + offset_w;
185
186
        *columns_ptr =
            mask_value * bilinear_interpolate(input_ptr, height, width, y, x);
187
188
189
190
191
192
193
        columns_ptr += batch_sz * out_h * out_w;
      }
    }
  }
}

static void deformable_im2col(
194
195
    const at::Tensor& input,
    const at::Tensor& data_offset,
196
    const at::Tensor& data_mask,
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
    int n_in_channels,
    int height,
    int width,
    int weight_h,
    int weight_w,
    int pad_h,
    int pad_w,
    int stride_h,
    int stride_w,
    int dil_h,
    int dil_w,
    int out_h,
    int out_w,
    int parallel_imgs,
    int deformable_group,
212
    bool use_mask,
213
214
215
216
217
218
219
220
221
    at::Tensor data_col) {
  int num_kernels = n_in_channels * out_h * out_w * parallel_imgs;

  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
      input.scalar_type(), "deformable_im2col", ([&] {
        deformable_im2col_kernel(
            num_kernels,
            input.data_ptr<scalar_t>(),
            data_offset.data_ptr<scalar_t>(),
222
            data_mask.data_ptr<scalar_t>(),
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
            height,
            width,
            weight_h,
            weight_w,
            pad_h,
            pad_w,
            stride_h,
            stride_w,
            dil_h,
            dil_w,
            parallel_imgs,
            n_in_channels,
            deformable_group,
            out_h,
            out_w,
238
            use_mask,
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
            data_col.data_ptr<scalar_t>());
      }));
}

static int get_greatest_divisor_below_bound(int n, int bound) {
  for (int k = bound; k > 1; --k) {
    if (n % k == 0) {
      return k;
    }
  }
  return 1;
}

at::Tensor DeformConv2d_forward_cpu(
    const at::Tensor& input_param,
    const at::Tensor& weight_param,
    const at::Tensor& offset_param,
256
    const at::Tensor& mask_param,
257
258
259
260
261
262
263
264
    const at::Tensor& bias_param,
    int64_t stride_h,
    int64_t stride_w,
    int64_t pad_h,
    int64_t pad_w,
    int64_t dil_h,
    int64_t dil_w,
    int64_t n_weight_grps,
265
266
    int64_t n_offset_grps,
    bool use_mask) {
267
268
269
  at::Tensor input = input_param.contiguous();
  at::Tensor offset = offset_param.contiguous();
  at::Tensor weight = weight_param.contiguous();
270
  at::Tensor mask = mask_param.contiguous();
271
  at::Tensor bias = bias_param.contiguous();
272
273
274

  TORCH_CHECK(input.ndimension() == 4);
  TORCH_CHECK(offset.ndimension() == 4);
275
  TORCH_CHECK(!use_mask || mask.ndimension() == 4);
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
  TORCH_CHECK(weight.ndimension() == 4);
  TORCH_CHECK(input.device().is_cpu(), "input must be a CPU tensor");

  int batch_sz = input.size(0);
  int n_in_channels = input.size(1);
  int in_h = input.size(2);
  int in_w = input.size(3);

  int n_parallel_imgs =
      get_greatest_divisor_below_bound(batch_sz, kMaxParallelImgs);

  // Unpack shapes and args
  int out_channels = weight.size(0);
  int weight_h = weight.size(2);
  int weight_w = weight.size(3);

  int ker_h = dil_h * (weight_h - 1) + 1;
  int ker_w = dil_w * (weight_w - 1) + 1;
  int out_h = ((in_h + 2 * pad_h - ker_h) / stride_h) + 1;
  int out_w = ((in_w + 2 * pad_w - ker_w) / stride_w) + 1;

  TORCH_CHECK(
      weight_h > 0 && weight_w > 0,
      "weight_h: ",
      weight_h,
      " weight_w: ",
      weight_w);
  TORCH_CHECK(
      stride_h > 0 && stride_w > 0,
      "stride_h: ",
      stride_h,
      " stride_w: ",
      stride_w);
  TORCH_CHECK(pad_h >= 0 && pad_w >= 0, "pad_h: ", pad_h, " pad_w: ", pad_w);
  TORCH_CHECK(dil_h > 0 && dil_w > 0, "dil_h: ", dil_h, " dil_w: ", dil_w);

  TORCH_CHECK(weight.size(1) * n_weight_grps == input.size(1));
  TORCH_CHECK(weight.size(0) % n_weight_grps == 0);
  TORCH_CHECK(
      (offset.size(1) == n_offset_grps * 2 * weight_h * weight_w),
316
      "offset.shape[1] is not valid: got: ",
317
318
319
      offset.size(1),
      " expected: ",
      n_offset_grps * 2 * weight_h * weight_w);
320
321
322
323
324
325
  TORCH_CHECK(
      (!use_mask || mask.size(1) == n_offset_grps * weight_h * weight_w),
      "mask.shape[1] is not valid: got: ",
      mask.size(1),
      " expected: ",
      n_offset_grps * weight_h * weight_w);
326
327
328
329
  TORCH_CHECK(input.size(1) % n_offset_grps == 0);

  TORCH_CHECK(
      (offset.size(0) == input.size(0)), "invalid batch size of offset");
330
331
332
333
334
335
336
337
338
339
340
341
  TORCH_CHECK(
      (offset.size(2) == out_h && offset.size(3) == out_w),
      "offset output dims: (",
      offset.size(2),
      ", ",
      offset.size(3),
      ") - ",
      "computed output dims: (",
      out_h,
      ", ",
      out_w,
      ")");
342
343
344
345
346
347
348
349
350
351
352
353
354
  TORCH_CHECK((mask.size(0) == input.size(0)), "invalid batch size of mask");
  TORCH_CHECK(
      (!use_mask || (mask.size(2) == out_h && mask.size(3) == out_w)),
      "offset output dims: (",
      mask.size(2),
      ", ",
      mask.size(3),
      ") - ",
      "computed output dims: (",
      out_h,
      ", ",
      out_w,
      ")");
355
356
357
358
359
360
361
362
  TORCH_CHECK(
      out_h > 0 && out_w > 0,
      "Calculated output size too small - out_h: ",
      out_h,
      " out_w: ",
      out_w);

  auto out = at::zeros({batch_sz, out_channels, out_h, out_w}, input.options());
363
364
365
  if (batch_sz == 0) {
    return out;
  }
366
367
368
369
370
371
372
373
374

  // Separate batches into blocks
  out = out.view({batch_sz / n_parallel_imgs,
                  n_parallel_imgs,
                  out_channels,
                  out_h,
                  out_w});
  input = input.view(
      {batch_sz / n_parallel_imgs, n_parallel_imgs, n_in_channels, in_h, in_w});
375

376
377
378
379
380
  offset = offset.view({batch_sz / n_parallel_imgs,
                        n_parallel_imgs,
                        n_offset_grps * 2 * weight_h * weight_w,
                        out_h,
                        out_w});
381
382
383
384
385
386
387
388
389

  if (use_mask) {
    mask = mask.view({batch_sz / n_parallel_imgs,
                      n_parallel_imgs,
                      n_offset_grps * weight_h * weight_w,
                      out_h,
                      out_w});
  }

390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
  at::Tensor out_buf = at::zeros(
      {batch_sz / n_parallel_imgs,
       out_channels,
       n_parallel_imgs * out_h,
       out_w},
      out.options());

  // Separate channels into convolution groups
  out_buf = out_buf.view({out_buf.size(0),
                          n_weight_grps,
                          out_buf.size(1) / n_weight_grps,
                          out_buf.size(2),
                          out_buf.size(3)});
  weight = weight.view({n_weight_grps,
                        weight.size(0) / n_weight_grps,
                        weight.size(1),
                        weight.size(2),
                        weight.size(3)});

  // Sample points and perform convolution
  auto columns = at::zeros(
      {n_in_channels * weight_h * weight_w, n_parallel_imgs * out_h * out_w},
      input.options());
  for (int b = 0; b < batch_sz / n_parallel_imgs; b++) {
    deformable_im2col(
        input[b],
        offset[b],
417
        mask[b],
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
        n_in_channels,
        in_h,
        in_w,
        weight_h,
        weight_w,
        pad_h,
        pad_w,
        stride_h,
        stride_w,
        dil_h,
        dil_w,
        out_h,
        out_w,
        n_parallel_imgs,
        n_offset_grps,
433
        use_mask,
434
435
436
437
438
439
440
441
442
443
        columns);

    columns = columns.view(
        {n_weight_grps, columns.size(0) / n_weight_grps, columns.size(1)});
    for (int g = 0; g < n_weight_grps; g++) {
      out_buf[b][g] = out_buf[b][g]
                          .flatten(1)
                          .addmm_(weight[g].flatten(1), columns[g])
                          .view_as(out_buf[b][g]);
    }
Francisco Massa's avatar
Francisco Massa committed
444
    columns =
445
        columns.view({columns.size(0) * columns.size(1), columns.size(2)});
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
  }

  out_buf = out_buf.view({batch_sz / n_parallel_imgs,
                          out_channels,
                          n_parallel_imgs,
                          out_h,
                          out_w});
  out_buf.transpose_(1, 2);
  out.copy_(out_buf);
  out = out.view({batch_sz, out_channels, out_h, out_w});

  return out + bias.view({1, out_channels, 1, 1});
}

template <typename scalar_t>
static void deformable_col2im_kernel(
462
    int n,
463
464
    const scalar_t* col,
    const scalar_t* offset,
465
    const scalar_t* mask,
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
    int channels,
    int height,
    int width,
    int kernel_h,
    int kernel_w,
    int pad_h,
    int pad_w,
    int stride_h,
    int stride_w,
    int dilation_h,
    int dilation_w,
    int batch_sz,
    int n_offset_grps,
    int out_h,
    int out_w,
481
    bool use_mask,
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
    scalar_t* grad_im) {
  for (int index = 0; index != n; ++index) {
    const int out_x = index % out_w;
    const int out_y = (index / out_w) % out_h;
    const int b = (index / (out_w * out_h)) % batch_sz;
    const int j = (index / (out_w * out_h * batch_sz)) % kernel_w;
    const int i = (index / (out_w * out_h * batch_sz * kernel_w)) % kernel_h;
    const int c = index / (out_w * out_h * batch_sz * kernel_w * kernel_h);

    int c_per_offset_grp = channels / n_offset_grps;
    const int offset_grp = c / c_per_offset_grp;

    auto offset_ptr = offset +
        (b * n_offset_grps + offset_grp) * 2 * kernel_h * kernel_w * out_h *
            out_w;
497
498
499
500
501
502
503
504
505
506
507
508
509

    auto mask_ptr = mask;
    if (use_mask) {
      mask_ptr += (b * n_offset_grps + offset_grp) * kernel_h * kernel_w *
          out_h * out_w;
    }

    const int mask_idx = i * kernel_w + j;
    const int offset_idx = 2 * mask_idx;

    const int offset_h_ptr = ((offset_idx)*out_h + out_y) * out_w + out_x;
    const int offset_w_ptr = ((offset_idx + 1) * out_h + out_y) * out_w + out_x;

510
511
    const scalar_t offset_h = offset_ptr[offset_h_ptr];
    const scalar_t offset_w = offset_ptr[offset_w_ptr];
512
513
514
515
516
517

    scalar_t mask_value = 1;
    if (use_mask) {
      mask_value = mask_ptr[(mask_idx * out_h + out_y) * out_w + out_x];
    }

518
519
520
521
522
523
524
525
526
527
528
    const scalar_t y = (out_y * stride_h - pad_h) + i * dilation_h + offset_h;
    const scalar_t x = (out_x * stride_w - pad_w) + j * dilation_w + offset_w;

    for (int dy = -1; dy <= 1; dy++) {
      for (int dx = -1; dx <= 1; dx++) {
        int yp = int(y) + dy;
        int xp = int(x) + dx;
        if (0 <= yp && yp < height && 0 <= xp && xp < width &&
            std::abs(y - yp) < 1 && std::abs(x - xp) < 1) {
          int grad_pos = ((b * channels + c) * height + yp) * width + xp;
          scalar_t weight = (1 - std::abs(y - yp)) * (1 - std::abs(x - xp));
529
          grad_im[grad_pos] += mask_value * weight * col[index];
530
531
532
533
534
535
536
        }
      }
    }
  }
}

static void compute_grad_input(
537
538
    const at::Tensor& columns,
    const at::Tensor& offset,
539
    const at::Tensor& mask,
540
541
542
543
544
545
546
547
548
549
550
551
552
    int channels,
    int height,
    int width,
    int weight_h,
    int weight_w,
    int pad_h,
    int pad_w,
    int stride_h,
    int stride_w,
    int dilation_h,
    int dilation_w,
    int parallel_imgs,
    int n_offset_grps,
553
    bool use_mask,
554
555
556
557
558
559
560
561
562
563
564
565
566
567
    at::Tensor grad_im) {
  int out_h =
      (height + 2 * pad_h - (dilation_h * (weight_h - 1) + 1)) / stride_h + 1;
  int out_w =
      (width + 2 * pad_w - (dilation_w * (weight_w - 1) + 1)) / stride_w + 1;
  int num_kernels =
      channels * weight_h * weight_w * out_h * out_w * parallel_imgs;

  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
      columns.scalar_type(), "deformable_col2im", ([&] {
        deformable_col2im_kernel(
            num_kernels,
            columns.data_ptr<scalar_t>(),
            offset.data_ptr<scalar_t>(),
568
            mask.data_ptr<scalar_t>(),
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
            channels,
            height,
            width,
            weight_h,
            weight_w,
            pad_h,
            pad_w,
            stride_h,
            stride_w,
            dilation_h,
            dilation_w,
            parallel_imgs,
            n_offset_grps,
            out_h,
            out_w,
584
            use_mask,
585
586
587
588
589
590
591
            grad_im.data_ptr<scalar_t>());
      }));
}

template <typename scalar_t>
static scalar_t get_coordinate_weight(
    const scalar_t* im_data,
592
593
    int height,
    int width,
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
    scalar_t y,
    scalar_t x,
    bool is_y_direction) {
  int y_l = floor(y);
  int x_l = floor(x);
  int y_h = y_l + 1;
  int x_h = x_l + 1;

  bool valid_y_l = 0 <= y_l && y_l < height;
  bool valid_y_h = 0 <= y_h && y_h < height;
  bool valid_x_l = 0 <= x_l && x_l < width;
  bool valid_x_h = 0 <= x_h && x_h < width;

  scalar_t zero = 0;
  scalar_t v_yx = (valid_y_l && valid_x_l) ? im_data[y_l * width + x_l] : zero;
  scalar_t v_yX = (valid_y_l && valid_x_h) ? im_data[y_l * width + x_h] : zero;
  scalar_t v_Yx = (valid_y_h && valid_x_l) ? im_data[y_h * width + x_l] : zero;
  scalar_t v_YX = (valid_y_h && valid_x_h) ? im_data[y_h * width + x_h] : zero;

  if (is_y_direction) {
    scalar_t dx = x - x_l;
    return dx * (v_YX - v_yX) + (1 - dx) * (v_Yx - v_yx);
  } else {
    scalar_t dy = y - y_l;
    return dy * (v_YX - v_Yx) + (1 - dy) * (v_yX - v_yx);
  }
}

template <typename scalar_t>
static void deformable_col2im_coord_kernel(
624
    int n,
625
626
627
    const scalar_t* col,
    const scalar_t* im,
    const scalar_t* offset,
628
    const scalar_t* mask,
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
    int channels,
    int height,
    int width,
    int weight_h,
    int weight_w,
    int pad_h,
    int pad_w,
    int stride_h,
    int stride_w,
    int dilation_h,
    int dilation_w,
    int batch_sz,
    int offset_channels,
    int n_offset_grps,
    int out_h,
    int out_w,
645
646
647
    bool use_mask,
    scalar_t* grad_offset,
    scalar_t* grad_mask) {
648
  for (int index = 0; index != n; ++index) {
649
650
651
    scalar_t grad_offset_val = 0;
    scalar_t grad_mask_val = 0;

652
653
    int w = index % out_w;
    int h = (index / out_w) % out_h;
654
655
    int w_w = (index / (out_w * out_h * 2)) % weight_w;
    int w_h = (index / (out_w * out_h * 2 * weight_w)) % weight_h;
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
    int c = (index / (out_w * out_h)) % offset_channels;
    int b = index / (out_w * out_h * offset_channels);

    const int offset_grp = c / (2 * weight_h * weight_w);
    const int col_step = weight_h * weight_w;

    int c_per_offset_grp = channels / n_offset_grps;

    auto col_ptr = col +
        offset_grp * c_per_offset_grp * weight_h * weight_w * batch_sz * out_w *
            out_h;
    auto im_ptr = im +
        (b * n_offset_grps + offset_grp) * c_per_offset_grp * height * width;
    auto offset_ptr = offset +
        (b * n_offset_grps + offset_grp) * 2 * weight_h * weight_w * out_h *
            out_w;

673
674
675
676
677
678
    auto mask_ptr = mask;
    if (use_mask) {
      mask_ptr += (b * n_offset_grps + offset_grp) * weight_h * weight_w *
          out_h * out_w;
    }

679
    const int offset_c = c - offset_grp * 2 * weight_h * weight_w;
680
    const bool is_y_direction = offset_c % 2 == 0;
681
682
683
684
685
686
687
688
689
690

    const int c_bound = c_per_offset_grp * weight_h * weight_w;
    for (int col_c = (offset_c / 2); col_c < c_bound; col_c += col_step) {
      const int col_pos = (((col_c * batch_sz + b) * out_h) + h) * out_w + w;

      int out_x = col_pos % out_w;
      int out_y = (col_pos / out_w) % out_h;
      int j = (col_pos / (out_w * out_h * batch_sz)) % weight_w;
      int i = (col_pos / (out_w * out_h * batch_sz * weight_w)) % weight_h;

691
692
      const int mask_idx = i * weight_w + j;

693
      const int offset_h_idx =
694
          (((2 * mask_idx) * out_h + out_y) * out_w + out_x);
695
      const int offset_w_idx =
696
          (((2 * mask_idx + 1) * out_h + out_y) * out_w + out_x);
697
698
699
      const scalar_t offset_h = offset_ptr[offset_h_idx];
      const scalar_t offset_w = offset_ptr[offset_w_idx];

700
701
702
703
704
      scalar_t mask_value = 1;
      if (use_mask) {
        mask_value = mask_ptr[(mask_idx * out_h + out_y) * out_w + out_x];
      }

705
706
707
708
709
      scalar_t y = (out_y * stride_h - pad_h) + i * dilation_h + offset_h;
      scalar_t x = (out_x * stride_w - pad_w) + j * dilation_w + offset_w;

      const scalar_t weight =
          get_coordinate_weight(im_ptr, height, width, y, x, is_y_direction);
710
711
712
713
714
715
716
      grad_offset_val += mask_value * weight * col_ptr[col_pos];

      if (use_mask && is_y_direction) {
        grad_mask_val += col_ptr[col_pos] *
            bilinear_interpolate(im_ptr, height, width, y, x);
      }

717
718
719
      im_ptr += height * width;
    }

720
721
722
723
724
725
726
727
728
729
730
731
    grad_offset[index] = grad_offset_val;

    if (use_mask && is_y_direction) {
      const int idx =
          ((((b * n_offset_grps + offset_grp) * weight_h + w_h) * weight_w +
            w_w) *
               out_h +
           h) *
              out_w +
          w;
      grad_mask[idx] = grad_mask_val;
    }
732
733
734
  }
}

735
static void compute_grad_offset_and_mask(
736
737
738
    const at::Tensor& columns,
    const at::Tensor& input,
    const at::Tensor& offset,
739
    const at::Tensor& mask,
740
741
742
743
744
745
746
747
748
749
750
751
752
    int channels,
    int height,
    int width,
    int weight_h,
    int weight_w,
    int pad_h,
    int pad_w,
    int stride_h,
    int stride_w,
    int dilation_h,
    int dilation_w,
    int parallel_imgs,
    int n_offset_grps,
753
754
755
    bool use_mask,
    at::Tensor grad_offset,
    at::Tensor grad_mask) {
756
757
758
759
760
761
762
763
764
765
766
767
768
769
  int out_h =
      (height + 2 * pad_h - (dilation_h * (weight_h - 1) + 1)) / stride_h + 1;
  int out_w =
      (width + 2 * pad_w - (dilation_w * (weight_w - 1) + 1)) / stride_w + 1;
  int num_kernels =
      out_h * out_w * 2 * weight_h * weight_w * n_offset_grps * parallel_imgs;

  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
      columns.scalar_type(), "deformable_col2im_coord", ([&] {
        deformable_col2im_coord_kernel(
            num_kernels,
            columns.data_ptr<scalar_t>(),
            input.data_ptr<scalar_t>(),
            offset.data_ptr<scalar_t>(),
770
            mask.data_ptr<scalar_t>(),
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
            channels,
            height,
            width,
            weight_h,
            weight_w,
            pad_h,
            pad_w,
            stride_h,
            stride_w,
            dilation_h,
            dilation_w,
            parallel_imgs,
            2 * weight_h * weight_w * n_offset_grps,
            n_offset_grps,
            out_h,
            out_w,
787
788
789
            use_mask,
            grad_offset.data_ptr<scalar_t>(),
            grad_mask.data_ptr<scalar_t>());
790
791
792
      }));
}

793
794
static std::tuple<at::Tensor, at::Tensor, at::Tensor>
deform_conv2d_backward_input_cpu(
795
796
797
    at::Tensor input,
    at::Tensor weight,
    at::Tensor offset,
798
    at::Tensor mask,
799
    at::Tensor grad_out,
800
801
802
803
804
805
    int stride_h,
    int stride_w,
    int pad_h,
    int pad_w,
    int dil_h,
    int dil_w,
806
807
    int n_weight_grps,
    int n_offset_grps,
808
809
    int n_parallel_imgs,
    bool use_mask) {
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
  int batch_sz = input.size(0);
  int n_in_channels = input.size(1);
  int in_h = input.size(2);
  int in_w = input.size(3);

  n_parallel_imgs = std::min(batch_sz, n_parallel_imgs);

  long n_out_channels = weight.size(0);
  int weight_h = weight.size(2);
  int weight_w = weight.size(3);

  long out_h = (in_h + 2 * pad_h - (dil_h * (weight_h - 1) + 1)) / stride_h + 1;
  long out_w = (in_w + 2 * pad_w - (dil_w * (weight_w - 1) + 1)) / stride_w + 1;

  auto grad_input = at::zeros_like(input);
  auto grad_offset = at::zeros_like(offset);
826
827
  auto grad_mask = at::zeros_like(mask);

828
  if (batch_sz == 0) {
829
    return std::make_tuple(grad_input, grad_offset, grad_mask);
830
  }
831

832
  auto columns = at::empty(
833
834
835
836
      {n_in_channels * weight_w * weight_h, n_parallel_imgs * out_h * out_w},
      input.options());

  // Separate into blocks
837
  grad_input = grad_input.reshape(
838
      {batch_sz / n_parallel_imgs, n_parallel_imgs, n_in_channels, in_h, in_w});
839
  input = input.reshape(
840
      {batch_sz / n_parallel_imgs, n_parallel_imgs, n_in_channels, in_h, in_w});
841

842
843
844
845
846
847
848
849
850
851
852
  grad_offset = grad_offset.reshape({batch_sz / n_parallel_imgs,
                                     n_parallel_imgs,
                                     n_offset_grps * 2 * weight_h * weight_w,
                                     out_h,
                                     out_w});
  offset = offset.reshape({batch_sz / n_parallel_imgs,
                           n_parallel_imgs,
                           n_offset_grps * 2 * weight_h * weight_w,
                           out_h,
                           out_w});

853
854
855
856
857
858
859
860
861
862
863
864
865
  if (use_mask) {
    grad_mask = grad_mask.reshape({batch_sz / n_parallel_imgs,
                                   n_parallel_imgs,
                                   n_offset_grps * weight_h * weight_w,
                                   out_h,
                                   out_w});
    mask = mask.reshape({batch_sz / n_parallel_imgs,
                         n_parallel_imgs,
                         n_offset_grps * weight_h * weight_w,
                         out_h,
                         out_w});
  }

866
867
868
869
870
871
872
873
  grad_out = grad_out
                 .reshape({batch_sz / n_parallel_imgs,
                           n_parallel_imgs,
                           n_weight_grps,
                           n_out_channels / n_weight_grps,
                           out_h,
                           out_w})
                 .permute({0, 2, 3, 1, 4, 5});
874
875
876
877
878
879
880
881
882

  weight = weight.reshape({n_weight_grps,
                           weight.size(0) / n_weight_grps,
                           weight.size(1),
                           weight.size(2),
                           weight.size(3)});

  columns = columns.view(
      {n_weight_grps, columns.size(0) / n_weight_grps, columns.size(1)});
Francisco Massa's avatar
Francisco Massa committed
883

884
  for (int elt = 0; elt < batch_sz / n_parallel_imgs; elt++) {
885
    columns.zero_();
886
887
888
889
890
891
    // Separate into weight groups
    for (int g = 0; g < n_weight_grps; g++) {
      columns[g] = columns[g].addmm_(
          weight[g].flatten(1).transpose(0, 1), grad_out[elt][g].flatten(1));
    }

892
    compute_grad_offset_and_mask(
893
894
895
        columns,
        input[elt],
        offset[elt],
896
        mask[elt],
897
898
899
900
901
902
903
904
905
906
907
908
909
        n_in_channels,
        in_h,
        in_w,
        weight_h,
        weight_w,
        pad_h,
        pad_w,
        stride_h,
        stride_w,
        dil_h,
        dil_w,
        n_parallel_imgs,
        n_offset_grps,
910
911
912
        use_mask,
        grad_offset[elt],
        grad_mask[elt]);
913
914
915
916

    compute_grad_input(
        columns,
        offset[elt],
917
        mask[elt],
918
919
920
921
922
923
924
925
926
927
928
929
930
        n_in_channels,
        in_h,
        in_w,
        weight_h,
        weight_w,
        pad_h,
        pad_w,
        stride_h,
        stride_w,
        dil_h,
        dil_w,
        n_parallel_imgs,
        n_offset_grps,
931
        use_mask,
932
933
934
935
936
937
938
        grad_input[elt]);
  }

  grad_input = grad_input.view({batch_sz, n_in_channels, in_h, in_w});
  grad_offset = grad_offset.view(
      {batch_sz, n_offset_grps * 2 * weight_h * weight_w, out_h, out_w});

939
940
941
942
943
944
  if (use_mask) {
    grad_mask = grad_mask.view(
        {batch_sz, n_offset_grps * weight_h * weight_w, out_h, out_w});
  }

  return std::make_tuple(grad_input, grad_offset, grad_mask);
945
946
947
948
}

static at::Tensor deform_conv2d_backward_parameters_cpu(
    at::Tensor input,
949
    const at::Tensor& weight,
950
    at::Tensor offset,
951
    at::Tensor mask,
952
    const at::Tensor& grad_out,
953
954
955
956
957
958
    int stride_h,
    int stride_w,
    int pad_h,
    int pad_w,
    int dil_h,
    int dil_w,
959
960
    int n_weight_grps,
    int n_offset_grps,
961
962
    int n_parallel_imgs,
    bool use_mask) {
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
  int batch_sz = input.size(0);
  int n_in_channels = input.size(1);
  int in_h = input.size(2);
  int in_w = input.size(3);

  n_parallel_imgs = std::min(batch_sz, n_parallel_imgs);

  long n_out_channels = weight.size(0);
  int weight_h = weight.size(2);
  int weight_w = weight.size(3);

  long out_h = grad_out.size(2);
  long out_w = grad_out.size(3);

  auto grad_weight = at::zeros_like(weight);
978
979
980
  if (batch_sz == 0) {
    return grad_weight;
  }
981

982
983
984
985
986
987
988
989
990
  at::Tensor grad_out_buf = grad_out
                                .reshape({batch_sz / n_parallel_imgs,
                                          n_parallel_imgs,
                                          n_weight_grps,
                                          n_out_channels / n_weight_grps,
                                          out_h,
                                          out_w})
                                .permute({0, 2, 3, 1, 4, 5})
                                .contiguous();
991
992

  input = input.reshape(
993
      {batch_sz / n_parallel_imgs, n_parallel_imgs, n_in_channels, in_h, in_w});
994

995
996
997
998
999
  offset = offset.reshape({batch_sz / n_parallel_imgs,
                           n_parallel_imgs,
                           n_offset_grps * 2 * weight_h * weight_w,
                           out_h,
                           out_w});
1000

1001
1002
1003
1004
1005
1006
1007
1008
  if (use_mask) {
    mask = mask.reshape({batch_sz / n_parallel_imgs,
                         n_parallel_imgs,
                         n_offset_grps * weight_h * weight_w,
                         out_h,
                         out_w});
  }

1009
1010
1011
1012
1013
  grad_weight = grad_weight.view({n_weight_grps,
                                  grad_weight.size(0) / n_weight_grps,
                                  grad_weight.size(1),
                                  grad_weight.size(2),
                                  grad_weight.size(3)});
1014
1015
1016
1017
1018
1019
1020

  auto columns = at::empty(
      {n_weight_grps,
       n_in_channels * weight_w * weight_h / n_weight_grps,
       n_parallel_imgs * out_h * out_w},
      input.options());

1021
1022
1023
1024
  for (int elt = 0; elt < batch_sz / n_parallel_imgs; elt++) {
    deformable_im2col(
        input[elt],
        offset[elt],
1025
        mask[elt],
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
        n_in_channels,
        in_h,
        in_w,
        weight_h,
        weight_w,
        pad_h,
        pad_w,
        stride_h,
        stride_w,
        dil_h,
        dil_w,
        out_h,
        out_w,
        n_parallel_imgs,
        n_offset_grps,
1041
        use_mask,
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
        columns);

    for (int g = 0; g < n_weight_grps; g++) {
      grad_weight[g] =
          grad_weight[g]
              .flatten(1)
              .addmm_(
                  grad_out_buf[elt][g].flatten(1), columns[g].transpose(1, 0))
              .view_as(grad_weight[g]);
    }
  }

  grad_weight = grad_weight.view({grad_weight.size(0) * grad_weight.size(1),
                                  grad_weight.size(2),
                                  grad_weight.size(3),
                                  grad_weight.size(4)});
  return grad_weight;
}

1061
std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
1062
DeformConv2d_backward_cpu(
1063
1064
1065
1066
    const at::Tensor& grad_out_param,
    const at::Tensor& input_param,
    const at::Tensor& weight_param,
    const at::Tensor& offset_param,
1067
    const at::Tensor& mask_param,
1068
1069
1070
1071
1072
1073
1074
1075
    const at::Tensor& bias_param,
    int64_t stride_h,
    int64_t stride_w,
    int64_t pad_h,
    int64_t pad_w,
    int64_t dil_h,
    int64_t dil_w,
    int64_t n_weight_grps,
1076
1077
    int64_t n_offset_grps,
    bool use_mask) {
1078
1079
1080
1081
  at::Tensor grad_out = grad_out_param.contiguous();
  at::Tensor input = input_param.contiguous();
  at::Tensor weight = weight_param.contiguous();
  at::Tensor offset = offset_param.contiguous();
1082
  at::Tensor mask = mask_param.contiguous();
1083
1084
  at::Tensor bias = bias_param.contiguous();

1085
1086
1087
1088
  const int batch_sz = input.size(0);
  const int n_parallel_imgs =
      get_greatest_divisor_below_bound(batch_sz, kMaxParallelImgs);

1089
  auto grad_input_and_offset_and_mask = deform_conv2d_backward_input_cpu(
1090
1091
1092
      input,
      weight,
      offset,
1093
      mask,
1094
      grad_out,
1095
1096
1097
1098
1099
1100
      stride_h,
      stride_w,
      pad_h,
      pad_w,
      dil_h,
      dil_w,
1101
1102
      n_weight_grps,
      n_offset_grps,
1103
1104
      n_parallel_imgs,
      use_mask);
1105

1106
1107
1108
  auto grad_input = std::get<0>(grad_input_and_offset_and_mask);
  auto grad_offset = std::get<1>(grad_input_and_offset_and_mask);
  auto grad_mask = std::get<2>(grad_input_and_offset_and_mask);
1109
1110
1111
1112
1113

  auto grad_weight = deform_conv2d_backward_parameters_cpu(
      input,
      weight,
      offset,
1114
      mask,
1115
      grad_out,
1116
1117
1118
1119
1120
1121
      stride_h,
      stride_w,
      pad_h,
      pad_w,
      dil_h,
      dil_w,
1122
1123
      n_weight_grps,
      n_offset_grps,
1124
1125
      n_parallel_imgs,
      use_mask);
1126
1127
1128

  auto grad_bias = at::ones_like(bias) * grad_out.sum({0, 2, 3});

1129
1130
  return std::make_tuple(
      grad_input, grad_weight, grad_offset, grad_mask, grad_bias);
1131
}