few-shot-learning for object detection

github  https://github.com/LiuXinyu12378/few-shot-learning-for-object-detection

train.py

?

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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231

from __future__ import print_function import sys   import time import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim import torch.backends.cudnn as cudnn from torchvision import datasets, transforms from torch.autograd import Variable from tqdm import tqdm   import dataset import random import math import os from utils import * from cfg import parse_cfg, cfg from darknet import Darknet import pdb   # Training settings # datacfg = sys.argv[1] # darknetcfg = parse_cfg(sys.argv[2]) # learnetcfg = parse_cfg(sys.argv[3])   datacfg = "cfg/fewyolov3_voc.data" darknetcfg = parse_cfg("cfg/darknet_yolov3_spp.cfg") learnetcfg = parse_cfg("cfg/reweighting_net.cfg") weightfile = "tmp/000050.weights" if len(sys.argv) == 5:     weightfile = sys.argv[4]   data_options = read_data_cfg(datacfg) net_options = darknetcfg[0] meta_options = learnetcfg[0]   # Configure options cfg.config_data(data_options) cfg.config_meta(meta_options) cfg.config_net(net_options)   # Parameters metadict = data_options['meta'] trainlist = data_options['train']   testlist = data_options['valid'] backupdir = data_options['backup'] gpus = data_options['gpus']  # e.g. 0,1,2,3 ngpus = len(gpus.split(',')) num_workers = int(data_options['num_workers'])   batch_size = int(net_options['batch']) print("batch_size:",batch_size) max_batches = int(net_options['max_batches']) learning_rate = float(data_options['learning_rate']) momentum = float(net_options['momentum']) decay = float(net_options['decay']) steps = [float(step) for step in data_options['steps'].split(',')] scales = [float(scale) for scale in data_options['scales'].split(',')]   # Train parameters use_cuda = True seed = int(time.time())   ## -------------------------------------------------------------------------- ## MAIN backupdir = cfg.backup print('logging to ' + backupdir) if not os.path.exists(backupdir):     os.makedirs(backupdir)   torch.manual_seed(seed) if use_cuda:     os.environ['CUDA_VISIBLE_DEVICES'] = gpus     torch.cuda.manual_seed(seed)   model = Darknet(darknetcfg, learnetcfg) region_loss = model.loss   model.print_network() # if len(sys.argv) == 5: model.load_weights(weightfile)   ################################################### ### Meta-model parameters region_loss.seen = model.seen processed_batches = 0 if cfg.tuning else model.seen / batch_size trainlist = dataset.build_dataset(data_options) nsamples = len(trainlist) init_width = model.width init_height = model.height init_epoch = 0 if cfg.tuning else model.seen / nsamples max_epochs = max_batches * batch_size / nsamples + 1 max_epochs = int(math.ceil(cfg.max_epoch * 1. / cfg.repeat)) if cfg.tuning else max_epochs print(cfg.repeat, nsamples, max_batches, batch_size) print(num_workers)   kwargs = {'num_workers': num_workers, 'pin_memory': True} if use_cuda else {}   if use_cuda:     if ngpus > 1:         model = torch.nn.DataParallel(model).cuda()     else:         model = model.cuda()   optimizer = optim.Adam(model.parameters(), lr=learning_rate)     def adjust_learning_rate(optimizer, processed_batches):     """Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""     lr = learning_rate     for i in range(len(steps)):         scale = scales[i] if i < len(scales) else 1         if processed_batches >= steps[i]:             lr = lr * scale             if processed_batches == steps[i]:                 break         else:             break     for param_group in optimizer.param_groups:         param_group['lr'] = lr     return lr     def train(epoch):     global processed_batches     t0 = time.time()     if ngpus > 1:         cur_model = model.module     else:         cur_model = model       train_loader = torch.utils.data.DataLoader(         dataset.listDataset(trainlist, shape=(init_width, init_height),                             shuffle=False,                             transform=transforms.Compose([                                 transforms.ToTensor(),                             ]),                             train=True,                             seen=cur_model.seen,                             batch_size=batch_size,                             num_workers=num_workers),         batch_size=batch_size, shuffle=False, **kwargs)       metaset = dataset.MetaDataset(metafiles=metadict, train=True)     metaloader = torch.utils.data.DataLoader(         metaset,         batch_size=metaset.batch_size,         shuffle=False,         num_workers=num_workers,         pin_memory=True     )     metaloader = iter(metaloader)       lr = adjust_learning_rate(optimizer, processed_batches)     logging('epoch %d/%d, processed %d samples, lr %e' % (epoch, max_epochs, epoch * len(train_loader.dataset), lr))       model.train()     t1 = time.time()     avg_time = torch.zeros(9)     with tqdm(total=train_loader.__len__()) as t:           for batch_idx, (data, target) in enumerate(train_loader):             metax, mask = metaloader.next()             t2 = time.time()             adjust_learning_rate(optimizer, processed_batches)             processed_batches = processed_batches + 1             if use_cuda:                 data = data.cuda()                 metax = metax.cuda()                 mask = mask.cuda()                 # target= target.cuda()             t3 = time.time()             data, target = Variable(data), Variable(target)             metax, mask = Variable(metax), Variable(mask)             t4 = time.time()             optimizer.zero_grad()             t5 = time.time()             output = model(data, metax, mask)             t6 = time.time()             region_loss.seen = region_loss.seen + data.data.size(0)             cur_model.seen = region_loss.seen             region_loss.input_size = (data.data.size(2), data.data.size(3))             loss,loss_box,loss_conf,loss_cls,cls_acc,recall50,recall75,nProposals = region_loss(output, target)             t.set_description('Epoch %d' % epoch)             t.set_postfix(loss=loss.item(), loss_bbox=loss_box,loss_conf=loss_conf,loss_cls=loss_cls,                           cls_acc=cls_acc, recall50=recall50, recall75=recall75,Proposals=nProposals)             t.update()               t7 = time.time()             loss.backward()             t8 = time.time()             optimizer.step()             t9 = time.time()             if False and batch_idx > 1:                 avg_time[0] = avg_time[0] + (t2 - t1)                 avg_time[1] = avg_time[1] + (t3 - t2)                 avg_time[2] = avg_time[2] + (t4 - t3)                 avg_time[3] = avg_time[3] + (t5 - t4)                 avg_time[4] = avg_time[4] + (t6 - t5)                 avg_time[5] = avg_time[5] + (t7 - t6)                 avg_time[6] = avg_time[6] + (t8 - t7)                 avg_time[7] = avg_time[7] + (t9 - t8)                 avg_time[8] = avg_time[8] + (t9 - t1)                 print('-------------------------------')                 print('       load data : %f' % (avg_time[0] / (batch_idx)))                 print('     cpu to cuda : %f' % (avg_time[1] / (batch_idx)))                 print('cuda to variable : %f' % (avg_time[2] / (batch_idx)))                 print('       zero_grad : %f' % (avg_time[3] / (batch_idx)))                 print(' forward feature : %f' % (avg_time[4] / (batch_idx)))                 print('    forward loss : %f' % (avg_time[5] / (batch_idx)))                 print('        backward : %f' % (avg_time[6] / (batch_idx)))                 print('            step : %f' % (avg_time[7] / (batch_idx)))                 print('           total : %f' % (avg_time[8] / (batch_idx)))             t1 = time.time()         print('')         t1 = time.time()         logging('training with %f samples/s' % (len(train_loader.dataset) / (t1 - t0)))           if (epoch + 1) % cfg.save_interval == 0:             logging('save weights to %s/%06d.weights' % (backupdir, epoch + 1))             cur_model.save_weights('%s/%06d.weights' % (backupdir, epoch + 1))   init_epoch = int(init_epoch) max_epochs = int(max_epochs) print("init_epoch:",init_epoch) print("max_epochs:",max_epochs) for epoch in range(init_epoch, max_epochs):     train(epoch)

 

region_loss.py

?

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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475

import time import torch import math import torch.nn as nn import torch.nn.functional as F import numpy as np from torch.autograd import Variable from utils import * from cfg import cfg from numbers import Number from random import random, randint import pdb     def neg_filter(pred_boxes, target, withids=False):     assert pred_boxes.size(0) == target.size(0)     if cfg.neg_ratio == 'full':         inds = list(range(pred_boxes.size(0)))     elif isinstance(cfg.neg_ratio, Number):         flags = torch.sum(target, 1) != 0         flags = flags.cpu().data.tolist()         ratio = cfg.neg_ratio * sum(flags) * 1. / (len(flags) - sum(flags))         if ratio >= 1:             inds = list(range(pred_boxes.size(0)))         else:             flags = [0 if f == 0 and random() > ratio else 1 for f in flags]             inds = np.argwhere(flags).squeeze()             pred_boxes, target = pred_boxes[inds], target[inds]     else:         raise NotImplementedError('neg_ratio not recognized')     if withids:         return pred_boxes, target, inds     else:         return pred_boxes, target     def neg_filter_v2(pred_boxes, target, withids=False):     assert pred_boxes.size(0) == target.size(0)     if cfg.neg_ratio == 'full':         inds = list(range(pred_boxes.size(0)))     elif isinstance(cfg.neg_ratio, Number):         flags = torch.sum(target, 1) != 0         flags = flags.cpu().data.tolist()         ratio = cfg.neg_ratio * sum(flags) * 1. / (len(flags) - sum(flags))         if ratio >= 1:             inds = list(range(pred_boxes.size(0)))         else:             flags = [0 if f == 0 and random() > ratio else 1 for f in flags]             if sum(flags) == 0:                 flags[randint(0, len(flags) - 1)] = 1             inds = np.nonzero(flags)[0]             pred_boxes, target = pred_boxes[inds], target[inds]     else:         raise NotImplementedError('neg_ratio not recognized')     if withids:         return pred_boxes, target, inds     else:         return pred_boxes, target     def build_targets(pred_boxes, target, conf, anchors, num_anchors, feature_size, input_size, ignore_thresh):     nB = target.size(0)     nA = num_anchors     # print('anchor_step: ', anchor_step)     obj_mask = torch.cuda.ByteTensor(nB, nA, feature_size[0], feature_size[1]).fill_(0)     noobj_mask = torch.cuda.ByteTensor(nB, nA, feature_size[0], feature_size[1]).fill_(1)     tx = torch.zeros(nB, nA, feature_size[0], feature_size[1]).cuda()     ty = torch.zeros(nB, nA, feature_size[0], feature_size[1]).cuda()     tw = torch.zeros(nB, nA, feature_size[0], feature_size[1]).cuda()     th = torch.zeros(nB, nA, feature_size[0], feature_size[1]).cuda()     tcls = torch.zeros(nB, nA, feature_size[0], feature_size[1]).cuda()     iou_scores = torch.zeros(nB, nA, feature_size[0], feature_size[1]).cuda()       tboxes = target.view(-1, 5)     nonzero_ind = tboxes[:, 3] > 0     tboxes = tboxes[nonzero_ind.unsqueeze(1).repeat(1, 5)].view(-1, 5)     ind_B = torch.linspace(0, nB - 1, nB).unsqueeze(1).repeat(1, 50).view(-1).long().cuda()     ind_B = ind_B[nonzero_ind]     gx = (tboxes[:, 1] * feature_size[1]).float()     gy = (tboxes[:, 2] * feature_size[0]).float()     gw = (tboxes[:, 3] * input_size[1]).float()     gh = (tboxes[:, 4] * input_size[0]).float()     aw = anchors[:, 0]     ah = anchors[:, 1]     nbox = tboxes.size(0)     gt_box = torch.cat([torch.zeros(1, nbox).cuda(), torch.zeros(1, nbox).cuda(), gw.unsqueeze(0), gh.unsqueeze(0)], 0)     anchor_box = torch.cat([torch.zeros(1, nA).cuda(), torch.zeros(1, nA).cuda(), aw.unsqueeze(0), ah.unsqueeze(0)], 0)     ious = bbox_ious(gt_box.unsqueeze(2).repeat(1, 1, nA), anchor_box.unsqueeze(1).repeat(1, nbox, 1), x1y1x2y2=False)     best_ious, best_a = ious.max(1)     gj = gy.long()     gi = gx.long()     obj_mask[ind_B, best_a, gj, gi] = 1     noobj_mask[ind_B, best_a, gj, gi] = 0       for i, iou in enumerate(ious):         if (iou > ignore_thresh).sum():             noobj_mask[ind_B[i:i + 1], (iou > ignore_thresh).nonzero().squeeze(1), gj[i:i + 1], gi[i:i + 1]] = 0       tx[ind_B, best_a, gj, gi] = gx - gx.floor()     ty[ind_B, best_a, gj, gi] = gy - gy.floor()     tw[ind_B, best_a, gj, gi] = torch.log(gw / anchors[best_a][:, 0])     th[ind_B, best_a, gj, gi] = torch.log(gh / anchors[best_a][:, 1])     tcls[ind_B, best_a, gj, gi] = tboxes[:, 0].float()     tconf = obj_mask.float()     pred_boxes = pred_boxes.contiguous().view(nB, nA, feature_size[0], feature_size[1], 4).cuda()     conf = conf.contiguous().view(nB, nA, feature_size[0], feature_size[1]).data     target_boxes = torch.cat([(tboxes[:, 1] * input_size[1]).float().unsqueeze(0),                               (tboxes[:, 2] * input_size[0]).float().unsqueeze(0),                               gw.unsqueeze(0),                               gh.unsqueeze(0)], 0)       iou_scores[ind_B, best_a, gj, gi] = bbox_ious(pred_boxes[ind_B, best_a, gj, gi].t(), target_boxes, x1y1x2y2=False)     conf50 = (conf[ind_B, best_a, gj, gi] > 0.5).float()     detected50 = (iou_scores[ind_B, best_a, gj, gi] > 0.5).float() * conf50     detected75 = (iou_scores[ind_B, best_a, gj, gi] > 0.75).float() * conf50       return nbox, iou_scores, obj_mask, noobj_mask, tx, ty, tw, th, tconf, tcls, detected50, detected75     class RegionLoss(nn.Module):     def __init__(self, num_classes=0, anchors=[], num_anchors=1):         super(RegionLoss, self).__init__()         self.num_classes = num_classes         self.anchors = anchors         self.num_anchors = num_anchors         self.anchor_step = len(anchors) / num_anchors         self.coord_scale = 1         self.noobject_scale = 1         self.object_scale = 5         self.class_scale = 1         self.thresh = 0.6         self.seen = 0       def forward(self, output, target):         # import pdb; pdb.set_trace()         # output : BxAs*(4+1+num_classes)*H*W           # if target.dim() == 3:         #     # target : B * n_cls * l         #     l = target.size(-1)         #     target = target.permute(1,0,2).contiguous().view(-1, l)         if target.dim() == 3:             target = target.view(-1, target.size(-1))         bef = target.size(0)         output, target = neg_filter(output, target)         # print("{}/{}".format(target.size(0), bef))           t0 = time.time()         nB = output.data.size(0)         nA = self.num_anchors         nC = self.num_classes         nH = output.data.size(2)         nW = output.data.size(3)           output = output.view(nB, nA, (5 + nC), nH, nW)         x = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([0]))).view(nB, nA, nH, nW))         y = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([1]))).view(nB, nA, nH, nW))         w = output.index_select(2, Variable(torch.cuda.LongTensor([2]))).view(nB, nA, nH, nW)         h = output.index_select(2, Variable(torch.cuda.LongTensor([3]))).view(nB, nA, nH, nW)         conf = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([4]))).view(nB, nA, nH, nW))         # [nB, nA, nC, nW, nH] | (bs, 5, 1, 13, 13)         cls = output.index_select(2, Variable(torch.linspace(5, 5 + nC - 1, nC).long().cuda()))         cls = cls.view(nB * nA, nC, nH * nW).transpose(1, 2).contiguous().view(nB * nA * nH * nW, nC)           t1 = time.time()           pred_boxes = torch.cuda.FloatTensor(4, nB * nA * nH * nW)         grid_x = torch.linspace(0, nW - 1, nW).repeat(nH, 1).repeat(nB * nA, 1, 1).view(nB * nA * nH * nW).cuda()         grid_y = torch.linspace(0, nH - 1, nH).repeat(nW, 1).t().repeat(nB * nA, 1, 1).view(nB * nA * nH * nW).cuda()         anchor_w = torch.Tensor(self.anchors).view(nA, self.anchor_step).index_select(1, torch.LongTensor([0])).cuda()         anchor_h = torch.Tensor(self.anchors).view(nA, self.anchor_step).index_select(1, torch.LongTensor([1])).cuda()         anchor_w = anchor_w.repeat(nB, 1).repeat(1, 1, nH * nW).view(nB * nA * nH * nW)         anchor_h = anchor_h.repeat(nB, 1).repeat(1, 1, nH * nW).view(nB * nA * nH * nW)         pred_boxes[0] = x.data + grid_x         pred_boxes[1] = y.data + grid_y         pred_boxes[2] = torch.exp(w.data) * anchor_w         pred_boxes[3] = torch.exp(h.data) * anchor_h         pred_boxes = convert2cpu(pred_boxes.transpose(0, 1).contiguous().view(-1, 4))         t2 = time.time()           nGT, nCorrect, coord_mask, conf_mask, cls_mask, tx, ty, tw, th, tconf, tcls = build_targets(pred_boxes,                                                                                                     target.data,                                                                                                     self.anchors, nA,                                                                                                     nC, \                                                                                                     nH, nW,                                                                                                     self.noobject_scale,                                                                                                     self.object_scale,                                                                                                     self.thresh,                                                                                                     self.seen)         cls_mask = (cls_mask == 1)         if cfg.metayolo:             tcls.zero_()         nProposals = int((conf > 0.25).float().sum().data[0])           tx = Variable(tx.cuda())         ty = Variable(ty.cuda())         tw = Variable(tw.cuda())         th = Variable(th.cuda())         tconf = Variable(tconf.cuda())         tcls = Variable(tcls.view(-1)[cls_mask].long().cuda())           coord_mask = Variable(coord_mask.cuda())         conf_mask = Variable(conf_mask.cuda().sqrt())         cls_mask = Variable(cls_mask.view(-1, 1).repeat(1, nC).cuda())         cls = cls[cls_mask].view(-1, nC)           t3 = time.time()           loss_x = self.coord_scale * nn.MSELoss(size_average=False)(x * coord_mask, tx * coord_mask) / 2.0         loss_y = self.coord_scale * nn.MSELoss(size_average=False)(y * coord_mask, ty * coord_mask) / 2.0         loss_w = self.coord_scale * nn.MSELoss(size_average=False)(w * coord_mask, tw * coord_mask) / 2.0         loss_h = self.coord_scale * nn.MSELoss(size_average=False)(h * coord_mask, th * coord_mask) / 2.0         loss_conf = nn.MSELoss(size_average=False)(conf * conf_mask, tconf * conf_mask) / 2.0         loss_cls = self.class_scale * nn.CrossEntropyLoss(size_average=False)(cls, tcls)         loss = loss_x + loss_y + loss_w + loss_h + loss_conf + loss_cls         t4 = time.time()         if False:             print('-----------------------------------')             print('        activation : %f' % (t1 - t0))             print(' create pred_boxes : %f' % (t2 - t1))             print('     build targets : %f' % (t3 - t2))             print('       create loss : %f' % (t4 - t3))             print('             total : %f' % (t4 - t0))         print('%d: nGT %d, recall %d, proposals %d, loss: x %f, y %f, w %f, h %f, conf %f, cls %f, total %f' % (             self.seen, nGT, nCorrect, nProposals, loss_x.data[0], loss_y.data[0], loss_w.data[0], loss_h.data[0],             loss_conf.data[0], loss_cls.data[0], loss.data[0]))         return loss     class RegionLossV2(nn.Module):     """     Yolo region loss + Softmax classification across meta-inputs     """       def __init__(self, num_classes=0, anchors=[], num_anchors=1, input_size=(832, 832)):         super(RegionLossV2, self).__init__()         self.num_classes = num_classes         self.anchors = anchors         self.num_anchors = num_anchors         self.coord_scale = 1         self.class_scale = 1         self.obj_scale = 1         self.noobj_scale = 100         self.thresh = 0.5         self.seen = 0         self.input_size = input_size         self.feature_scale = [32, 16, 8]         print('class_scale', self.class_scale)       def forward(self, output, target):         # output : (bs*cs, nA*(5+1), N)         # target : (bs, cs, 50*5)         # Get all classification prediction         # pdb.set_trace()         bs = target.size(0)         cs = target.size(1)         nA = self.num_anchors         nC = self.num_classes         N = output.data.size(2)         # feature_size = [[26, 26], [52, 52], [104, 104]]         cls = output.view(output.size(0), nA, (5 + nC), N)         cls = cls.index_select(2, Variable(torch.linspace(5, 5 + nC - 1, nC).long().cuda())).squeeze()         cls = cls.view(bs, cs, nA * N).transpose(1, 2).contiguous().view(bs * nA * N, cs)         cls_conf = F.softmax(cls, 1)         _, cls_max_ids = torch.max(cls_conf, 1)         cls_max_ids = cls_max_ids.data         pre_cls_mask = torch.zeros(bs * nA * N, cs).cuda()         pre_cls_mask[torch.linspace(0, bs * nA * N - 1, bs * nA * N).long().cuda(), cls_max_ids] = 1         pre_cls_mask = pre_cls_mask.view(bs, nA * N, cs).transpose(1, 2).contiguous().view(bs * cs, nA, N)           # Rearrange target and perform filtering operation         target = target.view(-1, target.size(-1))         # bef = target.size(0)         output, target, inds = neg_filter_v2(output, target, withids=True)         counts, _ = np.histogram(inds, bins=bs, range=(0, bs * cs))         # print("{}/{}".format(target.size(0), bef))         pre_cls_mask = pre_cls_mask[inds]           t0 = time.time()         nB = output.data.size(0)           output = output.view(nB, nA, (5 + nC), N)  # (nB, nA, (5+nC), N)         x = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([0]))).squeeze(2))  # (nB, nA, N)         y = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([1]))).squeeze(2))         w = output.index_select(2, Variable(torch.cuda.LongTensor([2]))).squeeze(2)         h = output.index_select(2, Variable(torch.cuda.LongTensor([3]))).squeeze(2)         conf = F.sigmoid(output.index_select(2, Variable(torch.cuda.LongTensor([4]))).squeeze(2))         # [nB, nA, nC, nW, nH] | (bs, 5, 1, 13, 13)         # cls  = output.index_select(2, Variable(torch.linspace(5,5+nC-1,nC).long().cuda()))         # cls  = cls.view(nB*nA, nC, nH*nW).transpose(1,2).contiguous().view(nB*nA*nH*nW, nC)         t1 = time.time()           pred_boxes = torch.cuda.FloatTensor(4, nB, nA, N)         grid_x = []         grid_y = []         anchor_w = []         anchor_h = []         scale = []         feature_size = []         for fs in self.feature_scale:             feature_h = self.input_size[0] // fs             feature_w = self.input_size[1] // fs             # print("feature_h:",feature_h)             # print("feature_w:",feature_w)             feature_size.append([feature_h, feature_w])             grid_x.append(torch.linspace(0, feature_w - 1, feature_w).repeat(feature_h, 1) \                           .repeat(nB * nA, 1, 1).view(nB, nA, feature_h * feature_w).cuda())             grid_y.append(torch.linspace(0, feature_h - 1, feature_h).repeat(feature_w, 1).t() \                           .repeat(nB * nA, 1, 1).view(nB, nA, feature_h * feature_w).cuda())             scale.append((torch.ones(nB, nA, feature_h * feature_w) * fs).cuda())         grid_x = torch.cat(grid_x, 2)  # (nB, nA, N)         grid_y = torch.cat(grid_y, 2)         scale = torch.cat(scale, 2)         for i in range(3):             aw = torch.Tensor(self.anchors[6 * i:6 * (i + 1)]).view(nA, -1) \                 .index_select(1, torch.LongTensor([0])).cuda()             ah = torch.Tensor(self.anchors[6 * i:6 * (i + 1)]).view(nA, -1) \                 .index_select(1, torch.LongTensor([1])).cuda()             anchor_w.append(aw.repeat(nB, feature_size[i][0] * feature_size[i][1]) \                             .view(nB, nA, feature_size[i][0] * feature_size[i][1]))             anchor_h.append(ah.repeat(nB, feature_size[i][0] * feature_size[i][1]) \                             .view(nB, nA, feature_size[i][0] * feature_size[i][1]))         anchor_w = torch.cat(anchor_w, 2)         anchor_h = torch.cat(anchor_h, 2)         pred_boxes[0] = (x.data + grid_x) * scale         pred_boxes[1] = (y.data + grid_y) * scale         pred_boxes[2] = torch.exp(w.data) * anchor_w         pred_boxes[3] = torch.exp(h.data) * anchor_h         pred_boxes = convert2cpu(pred_boxes.permute(1, 2, 3, 0).contiguous())  # (nB, nA, N, 4)         t2 = time.time()         nGT = 0         iou_scores = []         obj_mask = []         noobj_mask = []         tx = []         ty = []         tw = []         th = []         tconf = []         tcls = []         start_N = 0         detected50 = torch.zeros(0)         detected75 = torch.zeros(0)         for imap in range(3):             nGT, iou_scores_temp, obj_mask_temp, noobj_mask_temp, tx_temp, ty_temp, tw_temp, th_temp, tconf_temp, \             tcls_temp, detected50_temp, detected75_temp = build_targets(                 pred_boxes[:, :, start_N:start_N + feature_size[imap][0] * feature_size[imap][1], :],                 target.data.cuda(),                 conf[:, :, start_N:start_N + feature_size[imap][0] * feature_size[imap][1]],                 torch.Tensor(self.anchors[6 * imap:6 * (imap + 1)]).view(nA, -1).cuda(),                 nA,                 feature_size[imap],                 self.input_size,                 self.thresh)             if not len(detected50):                 detected50 = torch.zeros(nGT).cuda()             if not len(detected75):                 detected75 = torch.zeros(nGT).cuda()             detected50 += detected50_temp             detected75 += detected75_temp             start_N += feature_size[imap][0] * feature_size[imap][1]             iou_scores.append(iou_scores_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))             obj_mask.append(obj_mask_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))             noobj_mask.append(noobj_mask_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))             tx.append(tx_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))             ty.append(ty_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))             tw.append(tw_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))             th.append(th_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))             tconf.append(tconf_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))             tcls.append(tcls_temp.view(nB, nA, feature_size[imap][0] * feature_size[imap][1]))           iou_scores = torch.cat(iou_scores, 2)         obj_mask = torch.cat(obj_mask, 2)         noobj_mask = torch.cat(noobj_mask, 2)         tx = torch.cat(tx, 2)         ty = torch.cat(ty, 2)         tw = torch.cat(tw, 2)         th = torch.cat(th, 2)         tconf = torch.cat(tconf, 2)         tcls = torch.cat(tcls, 2)           # Take care of class mask         idx_start = 0         cls_mask_list = []         tcls_list = []         for i in range(len(counts)):             if counts[i] == 0:                 cur_mask = torch.zeros(nA, N).cuda()                 cur_tcls = torch.zeros(nA, N).cuda()             else:                 cur_mask = torch.sum(obj_mask[idx_start:idx_start + counts[i]].float(), dim=0)                 cur_tcls = torch.sum(tcls[idx_start:idx_start + counts[i]], dim=0)             cls_mask_list.append(cur_mask)             tcls_list.append(cur_tcls)             idx_start += counts[i]         cls_mask = torch.stack(cls_mask_list)  # (bs, nA, N)         tcls = torch.stack(tcls_list)           cls_mask = (cls_mask == 1)         conf50 = (conf > 0.5).float().data         iou50 = (iou_scores > 0.5).float()         detected_mask = conf50 * tconf         precision = torch.sum(iou50 * detected_mask) / (conf50.sum() + 1e-16)         detected50 = (detected50 > 0).float()         detected75 = (detected75 > 0).float()         recall50 = detected50.sum() / (nGT + 1e-16)         recall75 = detected75.sum() / (nGT + 1e-16)         nProposals = int((conf > 0.25).float().sum().item())         tx = Variable(tx)         ty = Variable(ty)         tw = Variable(tw)         th = Variable(th)         tconf = Variable(tconf)           obj_mask = Variable(obj_mask.bool())         noobj_mask = Variable(noobj_mask.bool())         # cls_mask   = Variable(cls_mask.view(-1, 1).repeat(1,cs).cuda())         cls = cls[Variable(cls_mask.view(-1, 1).repeat(1, cs))].view(-1, cs)         cls_max_ids = cls_max_ids[cls_mask.view(-1)]         tcls = Variable(tcls[cls_mask].long())         cls_acc = float(torch.sum(cls_max_ids == tcls.data)) / (cls_max_ids.numel() + 1e-16)           ClassificationLoss = nn.CrossEntropyLoss()         MseLoss = nn.MSELoss()         BceLoss = nn.BCELoss()           t3 = time.time()           loss_x = self.coord_scale * MseLoss(x[obj_mask], tx[obj_mask])         loss_y = self.coord_scale * MseLoss(y[obj_mask], ty[obj_mask])         loss_w = self.coord_scale * MseLoss(w[obj_mask], tw[obj_mask])         loss_h = self.coord_scale * MseLoss(h[obj_mask], th[obj_mask])         loss_conf_obj = BceLoss(conf[obj_mask], tconf[obj_mask])         loss_conf_noobj = BceLoss(conf[noobj_mask], tconf[noobj_mask])         loss_conf = self.obj_scale * loss_conf_obj + self.noobj_scale * loss_conf_noobj         if len(cls):             loss_cls = self.class_scale * ClassificationLoss(cls, tcls)         else:             loss_cls = Variable(torch.Tensor([0]).float().cuda())           # # pdb.set_trace()         # ids = [9,11,12,16]         # new_cls, new_tcls = select_classes(cls, tcls, ids)         # new_tcls = Variable(torch.from_numpy(new_tcls).long().cuda())         # loss_cls_new = self.class_scale * nn.CrossEntropyLoss(size_average=False)(new_cls, new_tcls)         # loss_cls_new *= 10         # loss_cls += loss_cls_new           loss = loss_x + loss_y + loss_w + loss_h + loss_conf + loss_cls         t4 = time.time()         if False:             print('-----------------------------------')             print('        activation : %f' % (t1 - t0))             print(' create pred_boxes : %f' % (t2 - t1))             print('     build targets : %f' % (t3 - t2))             print('       create loss : %f' % (t4 - t3))             print('             total : %f' % (t4 - t0))         # print(         #     '%d: nGT %d, precision %f, recall50 %f, recall75 %f, cls_acc %f, loss: x %f, y %f, w %f, h %f, conf %f, cls %f, total %f' % \         #     (self.seen, nGT, precision, recall50, recall75, cls_acc, loss_x.item(), loss_y.item(), \         #      loss_w.item(), loss_h.item(), loss_conf.item(), loss_cls.item(), loss.item()))         # print('%d: nGT %d, recall %d, proposals %d, loss: x %f, y %f, w %f, h %f, conf %f, cls %f, cls_new %f, total %f' % (self.seen, nGT, nCorrect, nProposals, loss_x.data[0], loss_y.data[0], loss_w.data[0], loss_h.data[0], loss_conf.data[0], loss_cls.data[0], loss_cls_new.data[0], loss.data[0]))         return loss,loss_x.item() + loss_y.item() + loss_w.item() + loss_h.item(),loss_conf.item(),loss_cls.item(),cls_acc,recall50.item(),recall75.item(),nProposals     def select_classes(pred, tgt, ids):     # convert tgt to numpy     tgt = tgt.cpu().data.numpy()     new_tgt = [(tgt == d) * i for i, d in enumerate(ids)]     new_tgt = np.max(np.stack(new_tgt), axis=0)     idxes = np.argwhere(new_tgt > 0).squeeze()     new_pred = pred[idxes]     new_pred = new_pred[:, ids]     new_tgt = new_tgt[idxes]     return new_pred, new_tgt