def train(train_loader, net, criterion, optimizer, epoch, train_args):
train_loss = AverageMeter()
curr_iter = (epoch - 1) * len(train_loader)
for i, data in enumerate(train_loader):
inputs, labels = data
assert inputs.size()[2:] == labels.size()[1:]
N = inputs.size(0)
inputs = Variable(inputs).cuda()
labels = Variable(labels).cuda()
optimizer.zero_grad()
outputs = net(inputs)
assert outputs.size()[2:] == labels.size()[1:]
assert outputs.size()[1] == voc.num_classes
loss = criterion(outputs, labels) / N
loss.backward()
optimizer.step()
train_loss.update(loss.data[0], N)
curr_iter += 1
writer.add_scalar('train_loss', train_loss.avg, curr_iter)
if (i + 1) % train_args['print_freq'] == 0:
print('[epoch %d], [iter %d / %d], [train loss %.5f]' % (
epoch, i + 1, len(train_loader), train_loss.avg
))
def validate(val_loader, model, criterion, log):
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to evaluate mode
model.eval()
for i, (input, target) in enumerate(val_loader):
if args.use_cuda:
target = target.cuda(async=True)
input = input.cuda()
input_var = torch.autograd.Variable(input, volatile=True)
target_var = torch.autograd.Variable(target, volatile=True)
# compute output
output = model(input_var)
loss = criterion(output, target_var)
# measure accuracy and record loss
prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
losses.update(loss.data[0], input.size(0))
top1.update(prec1[0], input.size(0))
top5.update(prec5[0], input.size(0))
print_log(' **Test** Prec@1 {top1.avg:.3f} Prec@5 {top5.avg:.3f} Error@1 {error1:.3f}'.format(top1=top1, top5=top5, error1=100-top1.avg), log)
return top1.avg, losses.avg
def main(args):
# Parse device ids
default_dev, *parallel_dev = parse_devices(args.devices)
all_devs = parallel_dev + [default_dev]
all_devs = [x.replace('gpu', '') for x in all_devs]
all_devs = [int(x) for x in all_devs]
nr_devs = len(all_devs)
with open(args.list_val, 'r') as f:
lines = f.readlines()
nr_files = len(lines)
if args.num_val > 0:
nr_files = min(nr_files, args.num_val)
nr_files_per_dev = math.ceil(nr_files / nr_devs)
pbar = tqdm(total=nr_files)
acc_meter = AverageMeter()
intersection_meter = AverageMeter()
union_meter = AverageMeter()
result_queue = Queue(500)
procs = []
for dev_id in range(nr_devs):
start_idx = dev_id * nr_files_per_dev
end_idx = min(start_idx + nr_files_per_dev, nr_files)
proc = Process(target=worker, args=(args, dev_id, start_idx, end_idx, result_queue))
print('process:%d, start_idx:%d, end_idx:%d' % (dev_id, start_idx, end_idx))
proc.start()
procs.append(proc)
# master fetches results
processed_counter = 0
while processed_counter < nr_files:
if result_queue.empty():
continue
(acc, pix, intersection, union) = result_queue.get()
acc_meter.update(acc, pix)
intersection_meter.update(intersection)
union_meter.update(union)
processed_counter += 1
pbar.update(1)
for p in procs:
p.join()
iou = intersection_meter.sum / (union_meter.sum + 1e-10)
for i, _iou in enumerate(iou):
print('class [{}], IoU: {}'.format(i, _iou))
print('[Eval Summary]:')
print('Mean IoU: {:.4}, Accuracy: {:.2f}%'
.format(iou.mean(), acc_meter.average()*100))
print('Evaluation Done!')
def validate(val_loader, model, i, silence=False):
batch_time = AverageMeter()
coco_gt = val_loader.dataset.coco
coco_pred = COCO()
coco_pred.dataset['images'] = [img for img in coco_gt.datasets['images']]
coco_pred.dataset['categories'] = copy.deepcopy(coco_gt.dataset['categories'])
id = 0
# switch to evaluate mode
model.eval()
end = time.time()
for i, (inputs, anns) in enumerate(val_loader):
# forward images one by one (TODO: support batch mode later, or
# multiprocess)
for j, input in enumerate(inputs):
input_anns= anns[j] # anns of this input
gt_bbox= np.vstack([ann['bbox'] + [ann['ordered_id']] for ann in input_anns])
im_info= [[input.size(1), input.size(2),
input_anns[0]['scale_ratio']]]
input_var= Variable(input.unsqueeze(0),
requires_grad=False).cuda()
cls_prob, bbox_pred, rois = model(input_var, im_info)
scores, pred_boxes = model.interpret_outputs(cls_prob, bbox_pred, rois, im_info)
print(scores, pred_boxes)
# for i in range(scores.shape[0]):
# measure elapsed time
batch_time.update(time.time() - end)
end= time.time()
coco_pred.createIndex()
coco_eval = COCOeval(coco_gt, coco_pred, 'bbox')
coco_eval.params.imgIds= sorted(coco_gt.getImgIds())
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
print('iter: [{0}] '
'Time {batch_time.avg:.3f} '
'Val Stats: {1}'
.format(i, coco_eval.stats,
batch_time=batch_time))
return coco_eval.stats[0]
def train(train_loader, net, criterion, optimizer, curr_epoch, train_args, val_loader, visualize):
while True:
train_main_loss = AverageMeter()
train_aux_loss = AverageMeter()
curr_iter = (curr_epoch - 1) * len(train_loader)
for i, data in enumerate(train_loader):
optimizer.param_groups[0]['lr'] = 2 * train_args['lr'] * (1 - float(curr_iter) / train_args['max_iter']
) ** train_args['lr_decay']
optimizer.param_groups[1]['lr'] = train_args['lr'] * (1 - float(curr_iter) / train_args['max_iter']
) ** train_args['lr_decay']
inputs, gts, _ = data
assert len(inputs.size()) == 5 and len(gts.size()) == 4
inputs.transpose_(0, 1)
gts.transpose_(0, 1)
assert inputs.size()[3:] == gts.size()[2:]
slice_batch_pixel_size = inputs.size(1) * inputs.size(3) * inputs.size(4)
for inputs_slice, gts_slice in zip(inputs, gts):
inputs_slice = Variable(inputs_slice).cuda()
gts_slice = Variable(gts_slice).cuda()
optimizer.zero_grad()
outputs, aux = net(inputs_slice)
assert outputs.size()[2:] == gts_slice.size()[1:]
assert outputs.size()[1] == cityscapes.num_classes
main_loss = criterion(outputs, gts_slice)
aux_loss = criterion(aux, gts_slice)
loss = main_loss + 0.4 * aux_loss
loss.backward()
optimizer.step()
train_main_loss.update(main_loss.data[0], slice_batch_pixel_size)
train_aux_loss.update(aux_loss.data[0], slice_batch_pixel_size)
curr_iter += 1
writer.add_scalar('train_main_loss', train_main_loss.avg, curr_iter)
writer.add_scalar('train_aux_loss', train_aux_loss.avg, curr_iter)
writer.add_scalar('lr', optimizer.param_groups[1]['lr'], curr_iter)
if (i + 1) % train_args['print_freq'] == 0:
print('[epoch %d], [iter %d / %d], [train main loss %.5f], [train aux loss %.5f]. [lr %.10f]' % (
curr_epoch, i + 1, len(train_loader), train_main_loss.avg, train_aux_loss.avg,
optimizer.param_groups[1]['lr']))
if curr_iter >= train_args['max_iter']:
return
if curr_iter % train_args['val_freq'] == 0:
validate(val_loader, net, criterion, optimizer, curr_epoch, i + 1, train_args, visualize)
curr_epoch += 1
def test(data_loader, model, opt, class_names):
print('test')
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
end_time = time.time()
output_buffer = []
previous_video_id = ''
test_results = {'results': {}}
for i, (inputs, targets) in enumerate(data_loader):
data_time.update(time.time() - end_time)
inputs = Variable(inputs, volatile=True)
outputs = model(inputs)
if not opt.no_softmax_in_test:
outputs = F.softmax(outputs)
for j in range(outputs.size(0)):
if not (i == 0 and j == 0) and targets[j] != previous_video_id:
calculate_video_results(output_buffer, previous_video_id,
test_results, class_names)
output_buffer = []
output_buffer.append(outputs[j].data.cpu())
previous_video_id = targets[j]
if (i % 100) == 0:
with open(
os.path.join(opt.result_path, '{}.json'.format(
opt.test_subset)), 'w') as f:
json.dump(test_results, f)
batch_time.update(time.time() - end_time)
end_time = time.time()
print('[{}/{}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'.format(
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time))
with open(
os.path.join(opt.result_path, '{}.json'.format(opt.test_subset)),
'w') as f:
json.dump(test_results, f)
def train(train_loader, model, optimizer, start_iter, num_iters):
batch_time = AverageMeter()
data_time = AverageMeter()
total_losses = AverageMeter()
rpn_losses = AverageMeter()
odn_losses = AverageMeter()
rpn_ce_losses = AverageMeter()
rpn_box_losses = AverageMeter()
odn_ce_losses = AverageMeter()
odn_box_losses = AverageMeter()
# switch to train mode
end_iter = start_iter + num_iters - 1
model.train()
end = time.time()
# for i in range(start_iter, start_iter + num_iters):
for i, (inputs, anns) in enumerate(train_loader):
i += start_iter
# get minibatch
# inputs, anns = next(train_loader)
lr = adjust_learning_rate(optimizer, args.lr, args.decay_rate,
i, args.niters) # TODO: add custom
# measure data loading time
data_time.update(time.time() - end)
optimizer.zero_grad()
# forward images one by one (TODO: support batch mode later, or
# multiprocess)
for j, input in enumerate(inputs):
input_anns = anns[j] # anns of this input
if len(input_anns) == 0:
continue
gt_bbox = np.vstack([ann['bbox'] + [ann['ordered_id']] for ann in input_anns])
im_info= [[input.size(1), input.size(2),
input_anns[0]['scale_ratio']]]
input_var= torch.autograd.Variable(input.unsqueeze(0).cuda(),
requires_grad=False)
cls_prob, bbox_pred, rois= model(input_var, im_info, gt_bbox)
loss= model.loss
loss.backward()
# record loss
total_losses.update(loss.data[0], input_var.size(0))
rpn_losses.update(model.rpn.loss.data[0], input_var.size(0))
rpn_ce_losses.update(
model.rpn.cross_entropy.data[0], input_var.size(0))
rpn_box_losses.update(
model.rpn.loss_box.data[0], input_var.size(0))
odn_losses.update(model.odn.loss.data[0], input_var.size(0))
odn_ce_losses.update(
model.odn.cross_entropy.data[0], input_var.size(0))
odn_box_losses.update(
model.odn.loss_box.data[0], input_var.size(0))
# do SGD step
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if args.print_freq > 0 and (i + 1) % args.print_freq == 0:
print('iter: [{0}] '
'Time {batch_time.val:.3f} '
'Data {data_time.val:.3f} '
'Loss {total_losses.val:.4f} '
'RPN {rpn_losses.val:.4f} '
'{rpn_ce_losses.val:.4f} '
'{rpn_box_losses.val:.4f} '
'ODN {odn_losses.val:.4f} '
'{odn_ce_losses.val:.4f} '
'{odn_box_losses.val:.4f} '
.format(i, batch_time=batch_time,
data_time=data_time,
total_losses=total_losses,
rpn_losses=rpn_losses,
rpn_ce_losses=rpn_ce_losses,
rpn_box_losses=rpn_box_losses,
odn_losses=odn_losses,
odn_ce_losses=odn_ce_losses,
odn_box_losses=odn_box_losses))
del inputs
del anns
if i == end_iter:
break
print('iter: [{0}-{1}] '
'Time {batch_time.avg:.3f} '
'Data {data_time.avg:.3f} '
'Loss {total_losses.avg:.4f} '
'RPN {rpn_losses.avg:.4f} '
'{rpn_ce_losses.avg:.4f} '
'{rpn_box_losses.avg:.4f} '
'ODN {odn_losses.avg:.4f} '
'{odn_ce_losses.avg:.4f} '
'{odn_box_losses.avg:.4f} '
.format(start_iter, end_iter,
batch_time=batch_time,
data_time=data_time,
total_losses=total_losses,
rpn_losses=rpn_losses,
rpn_ce_losses=rpn_ce_losses,
rpn_box_losses=rpn_box_losses,
odn_losses=odn_losses,
odn_ce_losses=odn_ce_losses,
odn_box_losses=odn_box_losses))
if args.tensorboard:
log_value('train_total_loss', total_losses.avg, end_iter)
log_value('train_rpn_loss', rpn_losses.avg, end_iter)
log_value('train_rpn_ce_loss', rpn_ce_losses.avg, end_iter)
log_value('train_rpn_box_loss', rpn_box_losses.avg, end_iter)
log_value('train_odn_loss', odn_losses.avg, end_iter)
log_value('train_odn_ce_loss', odn_ce_losses.avg, end_iter)
log_value('train_odn_box_loss', odn_box_losses.avg, end_iter)
return total_losses.avg
def train(trainloader, model, criterion, optimizer, epoch, use_cuda):
# switch to train mode
model.train()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
end = time.time()
for batch_idx, (inputs, targets) in enumerate(trainloader):
# measure data loading time
data_time.update(time.time() - end)
if use_cuda:
inputs, targets = inputs.cuda(), targets.cuda(async=True)
inputs, targets = torch.autograd.Variable(
inputs), torch.autograd.Variable(targets)
# compute output
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))
losses.update(loss.item(), inputs.size(0))
top1.update(prec1.item(), inputs.size(0))
top5.update(prec5.item(), inputs.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
progress_bar(
batch_idx, len(trainloader),
'Loss: %.2f | Top1: %.2f | Top5: %.2f' %
(losses.avg, top1.avg, top5.avg))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
return (losses.avg, top1.avg, top5.avg)
if args.other_distill is not None:
if args.other_distill == 'AT':
criterion = Attention()
weight = 1000
elif args.other_distill == 'SP':
criterion = Similarity()
weight = 3000
best_acc = 0
counter = torch.zeros(args.epoch, 50000).cuda()
epoch = 0
for epoch in range(args.epoch):
record = {name: AverageMeter() for name in items}
center = cal_center(val_loader, args, s_model)
for f**k, (x, y, k) in enumerate(train_loader):
s_model.train()
x = x.cuda()
y = y.cuda()
k = k.cuda()
with torch.no_grad():
s_feats, logits = s_model(x, is_feat=True)
probs = F.softmax(logits, dim=1)
# confidence
conf = probs.max(dim=1)[0]
# margin
def validate(val_loader, model, criterion, args, logger, writer, epoch,
local_rank):
batch_times = AverageMeter('Time', ':6.3f')
losses = AverageMeter('Loss', ':.4e') # 4e表示科学记数法中的4位小数
top1 = AverageMeter('Acc@1', ':6.2f')
# switch to evaluate mode
model.eval()
with torch.no_grad():
end = time.time()
for i, (images, target) in enumerate(val_loader):
images = images.cuda(local_rank, non_blocking=True)
target = target.cuda(local_rank, non_blocking=True)
# compute output
output = model(images)
loss = criterion(output, target)
# measure accuracy and record loss
acc1 = accuracy(output, target, 1)
# DDP: data synchronization
dist.barrier()
reduced_loss = reduce_mean(loss, args.nprocs)
reduced_acc1 = reduce_mean(acc1, args.nprocs)
losses.update(reduced_loss.item(), images.size(0))
top1.update(reduced_acc1, images.size(0))
# measure elapsed time
batch_time = time.time() - end
batch_times.update(batch_time)
end = time.time()
if i % args.print_freq == 0:
ddp_print(
'Val epoch: [{:d}/{:d}][{:d}/{:d}]\tce_loss={:.4f}\ttop1_acc={:.4f}\tbatch_time={:6.3f}s'
.format(epoch, args.epochs, i, len(val_loader), losses.avg,
top1.avg, batch_times.avg), logger, local_rank)
break
ddp_print(
'||==> Val epoch: [{:d}/{:d}]\tce_loss={:.4f}\ttop1_acc={:.4f}\tbatch_time={:6.3f}s'
.format(epoch, args.epochs, losses.avg, top1.avg,
batch_times.avg), logger, local_rank)
if args.local_rank == 0:
# save tensorboard
writer.add_scalar('Val_ce_loss', losses.avg, epoch)
writer.add_scalar('Val_top1_accuracy', top1.avg, epoch)
return top1.avg
def test(testloader, model, criterion, epoch, use_cuda):
global best_acc
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to evaluate mode
model.eval()
for batch_idx, (inputs, targets) in enumerate(testloader):
if use_cuda:
inputs, targets = inputs.cuda(local_rank), targets.cuda(local_rank)
inputs, targets = torch.autograd.Variable(inputs), torch.autograd.Variable(targets)
# compute output
with torch.no_grad():
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))
losses.update(loss.data.item(), inputs.size(0))
top1.update(prec1.item(), inputs.size(0))
top5.update(prec5.item(), inputs.size(0))
return (losses.avg, top1.avg)
def train(self, data_loader, epoch, args, result_dict):
total_loss = 0
count = 0
losses = AverageMeter()
top1 = AverageMeter()
self.model.train()
for batch_idx, (inputs, labels) in enumerate(data_loader):
inputs, labels = inputs.cuda(), labels.cuda()
if args.amp:
with torch.cuda.amp.autocast():
outputs = self.model(inputs)
loss = self.criterion(outputs, labels)
else:
outputs = self.model(inputs)
loss = self.criterion(outputs, labels)
if len(labels.size()) > 1:
labels = torch.argmax(labels, axis=1)
prec1, prec3 = accuracy(outputs.data, labels, topk=(1, 3))
losses.update(loss.item(), inputs.size(0))
top1.update(prec1.item(), inputs.size(0))
self.optimizer.zero_grad()
if args.amp:
self.scaler.scale(loss).backward()
self.scaler.step(self.optimizer)
self.scaler.update()
else:
loss.backward()
self.optimizer.step()
total_loss += loss.tolist()
count += labels.size(0)
if batch_idx % args.log_interval == 0:
_s = str(len(str(len(data_loader.sampler))))
ret = [
('epoch: {:0>3} [{: >' + _s + '}/{} ({: >3.0f}%)]').format(
epoch, count, len(data_loader.sampler),
100 * count / len(data_loader.sampler)),
'train_loss: {: >4.2e}'.format(total_loss / count),
'train_accuracy : {:.2f}%'.format(top1.avg)
]
print(', '.join(ret))
if not args.decay_type == 'swa':
self.scheduler.step()
else:
if epoch <= args.swa_start:
self.scheduler.step()
if epoch > args.swa_start and args.decay_type == 'swa':
self.swa_model.update_parameters(self.model)
self.swa_scheduler.step()
result_dict['train_loss'].append(losses.avg)
result_dict['train_acc'].append(top1.avg)
return result_dict
def eval_model_verbose(model,
test_loader,
decoder,
cuda,
out_path,
item_info_array,
warmups=0,
meta=False):
"""
Model evaluation -- used during inference.
returns wer, cer, batch time array and warm up time
"""
# Warming up
end = time.time()
total_trials = len(test_loader)
for i, data in enumerate(test_loader):
if i >= warmups:
break
sys.stdout.write("\rWarmups ({}/{}) ".format(i+1, warmups))
sys.stdout.flush()
if meta:
inputs, targets, input_percentages, target_sizes, batch_meta, item_meta = data
else:
inputs, targets, input_percentages, target_sizes = data
inputs = torch.autograd.Variable(inputs, volatile=False)
# unflatten targets
split_targets = []
offset = 0
for size in target_sizes:
split_targets.append(targets[offset:offset + size])
offset += size
if cuda:
inputs = inputs.cuda()
out = model(inputs)
warmup_time = time.time() - end
if warmups > 0: print("Warmed up in {}s").format(warmup_time)
# the actual inference trial
total_cer, total_wer = 0, 0
word_count, char_count = 0, 0
model.eval()
batch_time = AverageMeter()
# For each batch in the test_loader, make a prediction and calculate the WER CER
item_num = 1
with open(out_path, 'wb') as f:
csvwriter = csv.DictWriter(f, fieldnames=csv_header)
csvwriter.writeheader()
for i, data in enumerate(test_loader):
batch_num = i + 1
if meta:
inputs, targets, input_percentages, target_sizes, batch_meta, item_meta = data
else:
inputs, targets, input_percentages, target_sizes = data
inputs = torch.autograd.Variable(inputs, volatile=False)
# unflatten targets
split_targets = []
offset = 0
for size in target_sizes:
split_targets.append(targets[offset:offset + size])
offset += size
if cuda:
inputs = inputs.cuda()
end = time.time() # Timing start (Inference only)
out = model(inputs)
batch_time.update(time.time() - end) # Timing end (Inference only)
out = out.transpose(0, 1) # TxNxH
seq_length = out.size(0)
sizes = input_percentages.mul_(int(seq_length)).int()
# Decode the ouput to actual strings and compare to label
# Get the LEV score and the word, char count
decoded_output = decoder.decode(out.data, sizes)
target_strings = decoder.process_strings(decoder.convert_to_strings(split_targets))
batch_we = batch_wc = batch_ce = batch_cc = 0
for x in range(len(target_strings)):
this_we = decoder.wer(decoded_output[x], target_strings[x])
this_ce = decoder.cer(decoded_output[x], target_strings[x])
this_wc = len(target_strings[x].split())
this_cc = len(target_strings[x])
this_pred = decoded_output[x]
this_true = target_strings[x]
if item_num <= len(item_info_array):
item_latency = item_info_array[item_num - 1]['batch_latency']
else:
item_latency = "-9999"
out_data = [batch_num,
batch_time.array[-1],
batch_meta[2],
batch_meta[4],
batch_meta[3],
item_num,
item_latency,
item_meta[x][2],
item_meta[x][4],
item_meta[x][3],
this_wc, this_cc,
this_we, this_ce,
this_pred, this_true]
csv_dict = {k:v for k, v in zip(csv_header, out_data)}
csvwriter.writerow(csv_dict)
item_num += 1
batch_we += this_we
batch_ce += this_ce
batch_wc += this_wc
batch_cc += this_cc
total_wer += batch_we
total_cer += batch_ce
word_count += batch_wc
char_count += batch_cc
print('[{0}/{1}]\t'
'Batch: latency (running average) {batch_time.val:.4f} ({batch_time.avg:.3f})\t\t'
'WER {2:.1f} \t CER {3:.1f}'
.format((i + 1), total_trials,
batch_we / float(batch_wc),
batch_ce / float(batch_cc),
batch_time=batch_time))
if cuda:
torch.cuda.synchronize()
del out
# WER, CER
wer = total_wer / float(word_count)
cer = total_cer / float(char_count)
wer *= 100
cer *= 100
return wer, cer, batch_time, warmup_time
def train(epoch, model, criterion_xent, criterion_htri, optimizer, trainloader,
use_gpu):
losses = AverageMeter()
batch_time = AverageMeter()
data_time = AverageMeter()
# add two meters
xent_losses = AverageMeter()
htri_losses = AverageMeter()
#global_losses = AverageMeter()
#local_losses = AverageMeter()
model.train()
end = time.time()
for batch_idx, (imgs, pids, _) in enumerate(trainloader):
if use_gpu:
imgs, pids = imgs.cuda(), pids.cuda()
# measure data loading time
data_time.update(time.time() - end)
outputs, features = model(imgs)
if args.htri_only:
if isinstance(features, tuple):
loss = DeepSupervision(criterion_htri, features, pids)
else:
loss = criterion_htri(features, pids)
else:
if isinstance(outputs, tuple):
xent_loss = DeepSupervision(criterion_xent, outputs, pids)
else:
xent_loss = criterion_xent(outputs, pids) # use this one
if isinstance(features, tuple):
htri_loss = DeepSupervision(criterion_htri, features, pids)
else:
htri_loss = criterion_htri(features, pids) # use this one
loss = xent_loss + htri_loss # use this one
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
losses.update(loss.item(), pids.size(0))
xent_losses.update(xent_loss.item(), pids.size(0))
htri_losses.update(htri_loss.item(), pids.size(0))
if (batch_idx + 1) % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'CLoss {xent_loss.val:.4f} ({xent_loss.avg:.4f})\t'
'MLoss {htri_loss.val:.4f} ({htri_loss.avg:.4f})\t'.format(
epoch + 1,
batch_idx + 1,
len(trainloader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
xent_loss=xent_losses,
htri_loss=htri_losses))
def test(model, queryloader, galleryloader, use_gpu, ranks=[1, 5, 10, 20]):
batch_time = AverageMeter()
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
if use_gpu:
imgs = imgs.cuda()
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".
format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu:
imgs = imgs.cuda()
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print("Extracted features for gallery set, obtained {}-by-{} matrix".
format(gf.size(0), gf.size(1)))
print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(
batch_time.avg, args.test_batch))
m, n = qf.size(0), gf.size(0)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat,
q_pids,
g_pids,
q_camids,
g_camids,
use_metric_cuhk03=args.use_metric_cuhk03)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def valid(valid_loader, model, criterions, logger):
model.eval() # eval mode (dropout and batchnorm is NOT used)
losses = AverageMeter()
reg_losses = AverageMeter()
expression_accs = AverageMeter()
gender_accs = AverageMeter()
glasses_accs = AverageMeter()
race_accs = AverageMeter()
L1Loss, CrossEntropyLoss = criterions
# Batches
for i, (img, reg, expression, gender, glasses,
race) in enumerate(valid_loader):
# Move to GPU, if available
img = img.to(device)
reg_label = reg.type(torch.FloatTensor).to(device) # [N, 5]
expression_label = expression.type(torch.LongTensor).to(
device) # [N, 3]
gender_label = gender.type(torch.LongTensor).to(device) # [N, 2]
glasses_label = glasses.type(torch.LongTensor).to(device) # [N, 3]
race_label = race.type(torch.LongTensor).to(device) # [N, 4]
# Forward prop.
reg_out, expression_out, gender_out, glasses_out, race_out = model(img)
# Calculate loss
reg_loss = L1Loss(reg_out, reg_label) * loss_ratio
expression_loss = CrossEntropyLoss(expression_out, expression_label)
gender_loss = CrossEntropyLoss(gender_out, gender_label)
glasses_loss = CrossEntropyLoss(glasses_out, glasses_label)
race_loss = CrossEntropyLoss(race_out, race_label)
loss = reg_loss + expression_loss + gender_loss + glasses_loss + race_loss
# Keep track of metrics
losses.update(loss.item())
reg_losses.update(reg_loss.item())
expression_accuracy = accuracy(expression_out, expression_label)
expression_accs.update(expression_accuracy)
gender_accuracy = accuracy(gender_out, gender_label)
gender_accs.update(gender_accuracy)
glasses_accuracy = accuracy(glasses_out, glasses_label)
glasses_accs.update(glasses_accuracy)
race_accuracy = accuracy(race_out, race_label)
race_accs.update(race_accuracy)
# Print status
status = 'Validation: Loss {loss.avg:.4f}\t' \
'Reg Loss {reg_loss.val:.4f} ({reg_loss.avg:.4f})\t' \
'Expression Accuracy {expression_acc.val:.4f} ({expression_acc.avg:.4f})\t' \
'Gender Accuracy {gender_acc.val:.4f} ({gender_acc.avg:.4f})\t' \
'Glasses Accuracy {expression_acc.val:.4f} ({expression_acc.avg:.4f})\t' \
'Race Accuracy {expression_acc.val:.4f} ({expression_acc.avg:.4f})\n'.format(loss=losses,
reg_loss=reg_losses,
expression_acc=expression_accs,
gender_acc=gender_accs,
glasses_acc=glasses_accs,
race_acc=race_accs)
logger.info(status)
return losses.avg
def trainModel(data_loader,
model,
normalizer_global,
normalizer_local,
epoch=None,
evaluation=False,
testing=False):
"""
The function to train/test the model for one epoch. Also, writes the test results to a file 'test_results.csv' in the end
Parameters
----------
data_loader : The data iterator to generate batches
model : The model to train
normalizer_global : The normalizer for global gdt targets
normalizer_local : The normalizer for local lddt targets
epoch : The current epoch
evaluation : (bool) Denotes if the model is in eval mode (True for both testing and validation)
testing : (bool) Denotes if the model is in test mode (True only while testing)
Returns
-------
avg_errors_global : The average global MAE error
avg_errors_local : The average local MAE error
losses : The average MSE loss
"""
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
avg_errors_global = AverageMeter()
avg_errors_local = AverageMeter()
# placeholders to store results to write to file
if testing:
test_targets_global = []
test_preds_global = []
test_targets_local = []
test_preds_local = []
test_cif_ids = []
test_amino_crystal = []
end = time.time()
for protein_batch_iter, (input_data, batch_data,
target_tuples) in enumerate(data_loader):
batch_protein_ids = batch_data[0]
batch_amino_crystal = batch_data[1]
batch_size = len(batch_protein_ids)
# measure data loading time
data_time.update(time.time() - end)
# move inputs and targets to cuda
input_var, target_var = getInputs(input_data, target_tuples,
normalizer_global, normalizer_local)
if not evaluation and not testing:
# Switch to train mode
model.train()
out = model(input_var)
out = model.module.mask_remove(out)
assert out[1].shape[0] == target_var[1].shape[
0], "Predicted Outputs Amino & Target Outputs Amino don't match"
model.module.fit(out, target_var, batch_protein_ids)
else:
# evaluate one iteration
with torch.no_grad():
# Switch to evaluation mode
model.eval()
predicted = model(input_var)
predicted = model.module.mask_remove(predicted)
assert predicted[1].shape[0] == target_var[1].shape[
0], "Predicted Outputs Amino & Target Outputs Amino don't match"
model.module.fit(predicted,
target_var,
batch_protein_ids,
pred=True)
# Calculate the accuracy between the denormalized values
model.module.accuracy[
0] = model.module.accuracy[0] * normalizer_global.std
model.module.accuracy[
1] = model.module.accuracy[1] * normalizer_local.std
# measure accuracy and record loss
losses.update(model.module.loss.item(), batch_size)
avg_errors_global.update(model.module.accuracy[0].item(), batch_size)
avg_errors_local.update(model.module.accuracy[1].item(), batch_size)
# Collect all the results that needs to be written to file
if testing and batch_size != 1:
test_pred_global = normalizer_global.denorm(
model.module.outputs[0].data).squeeze().tolist()
test_target_global = target_tuples[0].squeeze()
test_preds_global += test_pred_global
test_targets_global += test_target_global.tolist()
test_amino_crystal += batch_amino_crystal.tolist()
test_pred_local = normalizer_local.denorm(
model.module.outputs[1].data).squeeze().tolist()
test_target_local = target_tuples[1].squeeze().tolist()
res1, res2 = OrderedDict(), OrderedDict()
for i, idx in enumerate(batch_amino_crystal):
if idx not in res1: res1[idx] = []
if idx not in res2: res2[idx] = []
res1[idx].append(test_target_local[i])
res2[idx].append(test_pred_local[i])
test_target_local = [v for _, v in res1.items()]
test_pred_local = [v for _, v in res2.items()]
test_preds_local += test_pred_local
test_targets_local += test_target_local
test_cif_ids += batch_protein_ids
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# print progress between steps
if protein_batch_iter % args.print_freq == 0:
if evaluation or testing:
print(
'Test: [{0}][{1}]/{2}\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'ERRG {avg_errors_global.val:.3f} ({avg_errors_global.avg:.3f})\t'
'ERRL {avg_errors_local.val:.3f} ({avg_errors_local.avg:.3f})'
.format(epoch,
protein_batch_iter,
len(data_loader),
batch_time=batch_time,
loss=losses,
avg_errors_global=avg_errors_global,
avg_errors_local=avg_errors_local))
else:
print(
'Epoch: [{0}][{1}]/{2}\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'ERRG {avg_errors_global.val:.3f} ({avg_errors_global.avg:.3f})\t'
'ERRL {avg_errors_local.val:.3f} ({avg_errors_local.avg:.3f})'
.format(epoch,
protein_batch_iter,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
avg_errors_global=avg_errors_global,
avg_errors_local=avg_errors_local))
if protein_batch_iter % args.print_freq == 0:
clearCache()
# write results to file
if testing:
star_label = '**'
with open(savepath + 'test_results.csv', 'w') as f:
writer = csv.writer(f)
for cif_id, targets_global, preds_global, targets_local, preds_local in zip(
test_cif_ids, test_targets_global, test_preds_global,
test_targets_local, test_preds_local):
writer.writerow((cif_id, targets_global, preds_global,
targets_local, preds_local))
elif evaluation:
star_label = '*'
else:
star_label = '##'
print(
' {star} ERRG {avg_errors_global.avg:.3f} ERRL {avg_errors_local.avg:.3f} LOSS {avg_loss.avg:.3f}'
.format(star=star_label,
avg_errors_global=avg_errors_global,
avg_errors_local=avg_errors_local,
avg_loss=losses))
return avg_errors_global.avg, avg_errors_local.avg, losses.avg
def train(train_loader, model, criterions, optimizer, epoch, logger):
model.train() # train mode (dropout and batchnorm is used)
losses = AverageMeter()
reg_losses = AverageMeter()
expression_accs = AverageMeter()
gender_accs = AverageMeter()
glasses_accs = AverageMeter()
race_accs = AverageMeter()
L1Loss, CrossEntropyLoss = criterions
# Batches
for i, (img, reg, expression, gender, glasses,
race) in enumerate(train_loader):
# Move to GPU, if available
img = img.to(device)
reg_label = reg.type(torch.FloatTensor).to(device) # [N, 5]
expression_label = expression.type(torch.LongTensor).to(
device) # [N, 3]
gender_label = gender.type(torch.LongTensor).to(device) # [N, 2]
glasses_label = glasses.type(torch.LongTensor).to(device) # [N, 3]
race_label = race.type(torch.LongTensor).to(device) # [N, 4]
# Forward prop.
reg_out, expression_out, gender_out, glasses_out, race_out = model(
img) # embedding => [N, 17]
# Calculate loss
reg_loss = L1Loss(reg_out, reg_label) * loss_ratio
expression_loss = CrossEntropyLoss(expression_out, expression_label)
gender_loss = CrossEntropyLoss(gender_out, gender_label)
glasses_loss = CrossEntropyLoss(glasses_out, glasses_label)
race_loss = CrossEntropyLoss(race_out, race_label)
loss = reg_loss + expression_loss + gender_loss + glasses_loss + race_loss
# Back prop.
optimizer.zero_grad()
loss.backward()
# Clip gradients
clip_gradient(optimizer, grad_clip)
# Update weights
optimizer.step()
# Keep track of metrics
losses.update(loss.item())
reg_losses.update(reg_loss.item())
expression_accuracy = accuracy(expression_out, expression_label)
expression_accs.update(expression_accuracy)
gender_accuracy = accuracy(gender_out, gender_label)
gender_accs.update(gender_accuracy)
glasses_accuracy = accuracy(glasses_out, glasses_label)
glasses_accs.update(glasses_accuracy)
race_accuracy = accuracy(race_out, race_label)
race_accs.update(race_accuracy)
# Print status
if i % print_freq == 0:
status = 'Epoch: [{0}][{1}/{2}]\t' \
'Loss {loss.val:.4f} ({loss.avg:.4f})\t' \
'Reg Loss {reg_loss.val:.4f} ({reg_loss.avg:.4f})\t' \
'Expression Accuracy {expression_acc.val:.4f} ({expression_acc.avg:.4f})\t' \
'Gender Accuracy {gender_acc.val:.4f} ({gender_acc.avg:.4f})\t' \
'Glasses Accuracy {expression_acc.val:.4f} ({expression_acc.avg:.4f})\t' \
'Race Accuracy {expression_acc.val:.4f} ({expression_acc.avg:.4f})\t'.format(epoch, i,
len(train_loader),
loss=losses,
reg_loss=reg_losses,
expression_acc=expression_accs,
gender_acc=gender_accs,
glasses_acc=glasses_accs,
race_acc=race_accs)
logger.info(status)
return losses.avg
def validate(val_loader, model, criterion, gpu_avail, print_freq, f):
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
# switch to evaluate mode
model.eval()
with torch.no_grad():
end = time.time()
for i, (input, target) in enumerate(val_loader):
if gpu_avail:
input = input.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
# compute output
output = model(input)
loss = criterion(output, target)
# measure accuracy and record loss
acc1 = accuracy(output, target, topk=(1, ))
losses.update(loss.item(), input.size(0))
top1.update(acc1[0], input.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % print_freq == 0:
print('Test: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {top1.val:.3f} ({top1.avg:.3f})'.format(
i,
len(val_loader),
batch_time=batch_time,
loss=losses,
top1=top1))
f.write('Test: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\n'.format(
i,
len(val_loader),
batch_time=batch_time,
loss=losses,
top1=top1))
print(' * Acc@1 {top1.avg:.3f}'.format(top1=top1))
f.write(' * Acc@1 {top1.avg:.3f}\n'.format(top1=top1))
return top1.avg
def train(train_loader, model, criterion, optimizer, epoch, gpu_avail,
print_freq, f):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
# switch to train mode
model.train()
end = time.time()
for i, (input, target) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
if gpu_avail:
input = input.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
# compute output
output = model(input)
loss = criterion(output, target)
# measure accuracy and record loss
acc1 = accuracy(output, target, topk=(1, ))
losses.update(loss.item(), input.size(0))
top1.update(acc1[0], input.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {top1.val:.3f} ({top1.avg:.3f})'.format(
epoch,
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
top1=top1))
f.write('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\n'.format(
epoch,
i,
len(train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
top1=top1))
def validate(val_loader, net, criterion, optimizer, epoch, train_args, restore, visualize):
net.eval()
val_loss = AverageMeter()
inputs_all, gts_all, predictions_all = [], [], []
for vi, data in enumerate(val_loader):
inputs, gts = data
N = inputs.size(0)
inputs = Variable(inputs, volatile=True).cuda()
gts = Variable(gts, volatile=True).cuda()
outputs = net(inputs)
predictions = outputs.data.max(1)[1].squeeze_(1).squeeze_(0).cpu().numpy()
val_loss.update(criterion(outputs, gts).data[0] / N, N)
if random.random() > train_args['val_img_sample_rate']:
inputs_all.append(None)
else:
inputs_all.append(inputs.data.squeeze_(0).cpu())
gts_all.append(gts.data.squeeze_(0).cpu().numpy())
predictions_all.append(predictions)
acc, acc_cls, mean_iu, fwavacc = evaluate(predictions_all, gts_all, voc.num_classes)
if mean_iu > train_args['best_record']['mean_iu']:
train_args['best_record']['val_loss'] = val_loss.avg
train_args['best_record']['epoch'] = epoch
train_args['best_record']['acc'] = acc
train_args['best_record']['acc_cls'] = acc_cls
train_args['best_record']['mean_iu'] = mean_iu
train_args['best_record']['fwavacc'] = fwavacc
snapshot_name = 'epoch_%d_loss_%.5f_acc_%.5f_acc-cls_%.5f_mean-iu_%.5f_fwavacc_%.5f_lr_%.10f' % (
epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc, optimizer.param_groups[1]['lr']
)
torch.save(net.state_dict(), os.path.join(ckpt_path, exp_name, snapshot_name + '.pth'))
torch.save(optimizer.state_dict(), os.path.join(ckpt_path, exp_name, 'opt_' + snapshot_name + '.pth'))
if train_args['val_save_to_img_file']:
to_save_dir = os.path.join(ckpt_path, exp_name, str(epoch))
check_mkdir(to_save_dir)
val_visual = []
for idx, data in enumerate(zip(inputs_all, gts_all, predictions_all)):
if data[0] is None:
continue
input_pil = restore(data[0])
gt_pil = voc.colorize_mask(data[1])
predictions_pil = voc.colorize_mask(data[2])
if train_args['val_save_to_img_file']:
input_pil.save(os.path.join(to_save_dir, '%d_input.png' % idx))
predictions_pil.save(os.path.join(to_save_dir, '%d_prediction.png' % idx))
gt_pil.save(os.path.join(to_save_dir, '%d_gt.png' % idx))
val_visual.extend([visualize(input_pil.convert('RGB')), visualize(gt_pil.convert('RGB')),
visualize(predictions_pil.convert('RGB'))])
val_visual = torch.stack(val_visual, 0)
val_visual = vutils.make_grid(val_visual, nrow=3, padding=5)
writer.add_image(snapshot_name, val_visual)
print('--------------------------------------------------------------------')
print('[epoch %d], [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f]' % (
epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc))
print('best record: [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f], [epoch %d]' % (
train_args['best_record']['val_loss'], train_args['best_record']['acc'], train_args['best_record']['acc_cls'],
train_args['best_record']['mean_iu'], train_args['best_record']['fwavacc'], train_args['best_record']['epoch']))
print('--------------------------------------------------------------------')
writer.add_scalar('val_loss', val_loss.avg, epoch)
writer.add_scalar('acc', acc, epoch)
writer.add_scalar('acc_cls', acc_cls, epoch)
writer.add_scalar('mean_iu', mean_iu, epoch)
writer.add_scalar('fwavacc', fwavacc, epoch)
writer.add_scalar('lr', optimizer.param_groups[1]['lr'], epoch)
net.train()
return val_loss.avg
def train(train_loader, decoder, criterion_ce, criterion_dis, decoder_optimizer, epoch):
"""
Performs one epoch's training.
:param train_loader: DataLoader for training data
:param decoder: decoder model
:param criterion_ce: cross entropy loss layer
:param criterion_dis : discriminative loss layer
:param decoder_optimizer: optimizer to update decoder's weights
:param epoch: epoch number
"""
decoder.train() # train mode (dropout and batchnorm is used)
batch_time = AverageMeter() # forward prop. + back prop. time
data_time = AverageMeter() # data loading time
losses = AverageMeter() # loss (per word decoded)
top5accs = AverageMeter() # top5 accuracy
start = time.time()
# Batches
for i, sample in enumerate(train_loader):
if scene_graph:
(obj, rel, caps, caplens, obj_mask, rel_mask, pair_idx) = sample
obj = obj.to(device)
rel = rel.to(device)
obj_mask = obj_mask.to(device)
rel_mask = rel_mask.to(device)
pair_idx = pair_idx.to(device)
else:
(imgs, caps, caplens) = sample
imgs = imgs.to(device)
data_time.update(time.time() - start)
# Move to GPU, if available
caps = caps.to(device)
caplens = caplens.to(device)
# Forward prop.
if scene_graph:
scores, scores_d, caps_sorted, decode_lengths, sort_ind = decoder(object_features=obj,
relation_features=rel,
encoded_captions=caps,
caption_lengths=caplens,
object_mask=obj_mask,
relation_mask=rel_mask,
rel_pair_idx=pair_idx)
else:
scores, scores_d, caps_sorted, decode_lengths, sort_ind = decoder(imgs, caps, caplens)
# Max-pooling across predicted words across time steps for discriminative supervision
scores_d = scores_d.max(1)[0]
# Since we decoded starting with <start>, the targets are all words after <start>, up to <end>
targets = caps_sorted[:, 1:]
targets_d = torch.zeros(scores_d.size(0), scores_d.size(1)).to(device)
targets_d.fill_(-1)
for length in decode_lengths:
targets_d[:, :length - 1] = targets[:, :length - 1]
# Remove timesteps that we didn't decode at, or are pads
# pack_padded_sequence is an easy trick to do this
scores = pack_padded_sequence(scores, decode_lengths, batch_first=True, enforce_sorted=True).data
targets = pack_padded_sequence(targets, decode_lengths, batch_first=True, enforce_sorted=True).data
#scores, _ = pack_padded_sequence(scores, decode_lengths, batch_first=True)
#targets, _ = pack_padded_sequence(targets, decode_lengths, batch_first=True)
# Calculate loss
loss_d = criterion_dis(scores_d, targets_d.long())
loss_g = criterion_ce(scores, targets)
loss = loss_g + (10 * loss_d)
# Back prop.
decoder_optimizer.zero_grad()
loss.backward()
# Clip gradients when they are getting too large
torch.nn.utils.clip_grad_norm_(filter(lambda p: p.requires_grad, decoder.parameters()), 0.25)
# Update weights
decoder_optimizer.step()
# Keep track of metrics
top5 = accuracy(scores, targets, 5)
losses.update(loss.item(), sum(decode_lengths))
top5accs.update(top5, sum(decode_lengths))
batch_time.update(time.time() - start)
start = time.time()
# Print status
if i % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Batch Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data Load Time {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Top-5 Accuracy {top5.val:.3f} ({top5.avg:.3f})'.format(epoch, i, len(train_loader),
batch_time=batch_time,
data_time=data_time, loss=losses,
top5=top5accs))
def train_model(model, criterion, optimizer, log_saver, num_epochs=70):
since = time.time()
steps = 0
for epoch in range(num_epochs):
print('Epoch {}/{}'.format(epoch + 1, num_epochs))
print('-' * 10)
for phase in ['train', 'test']:
loss_meter = AverageMeter()
acc_meter = AverageMeter()
if phase == 'train':
model.train(True)
else:
model.train(False)
for i, data in enumerate(loaders[phase]):
inputs, labels = data
if use_gpu:
inputs = inputs.cuda()
labels = labels.cuda()
if phase == 'train':
inputs = Variable(inputs)
labels = Variable(labels)
else:
inputs = Variable(inputs, volatile=True)
labels = Variable(labels, volatile=True)
optimizer.zero_grad()
outputs = model(inputs)
_, preds = torch.max(outputs.data, 1)
loss = criterion(outputs, labels)
if phase == 'train':
loss.backward()
optimizer.step()
steps += 1
loss_meter.update(loss.data[0], outputs.size(0))
acc_meter.update(
accuracy(outputs.data, labels.data)[-1][0],
outputs.size(0))
epoch_loss = loss_meter.avg
epoch_error = 1 - acc_meter.avg / 100
if phase == 'train' and epoch == num_epochs - 1:
log_saver['train_loss'].append(epoch_loss)
log_saver['train_error'].append(epoch_error)
elif phase == 'test' and epoch == num_epochs - 1:
log_saver['test_loss'].append(epoch_loss)
log_saver['test_error'].append(epoch_error)
print('{} Loss: {:.4f} Error: {:.4f}'.format(
phase, epoch_loss, epoch_error))
if epoch % 30 == 0 or epoch == num_epochs - 1:
print('Saving..')
state = {'net': model, 'epoch': epoch, 'log': log_saver}
if not os.path.isdir('checkpoint_CNN'):
os.mkdir('checkpoint_CNN')
torch.save(
state, './checkpoint_CNN/ckpt_epoch_{}_{}.t7'.format(
epoch, log_saver['num_params'][-1]))
time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
return model, log_saver
def validate(val_loader, decoder, criterion_ce, criterion_dis, epoch):
"""
Performs one epoch's validation.
:param val_loader: DataLoader for validation data.
:param decoder: decoder model
:param criterion_ce: cross entropy loss layer
:param criterion_dis : discriminative loss layer
:return: BLEU-4 score
"""
decoder.eval() # eval mode (no dropout or batchnorm)
batch_time = AverageMeter()
losses = AverageMeter()
top5accs = AverageMeter()
start = time.time()
references = list() # references (true captions) for calculating BLEU-4 score
hypotheses = list() # hypotheses (predictions)
# Batches
with torch.no_grad():
# for i, (imgs, caps, caplens,allcaps) in enumerate(val_loader):
for i, sample in enumerate(val_loader):
if i % 5 != 0:
# only decode every 5th caption, starting from idx 0.
# this is because the iterator iterates over all captions in the dataset, not all images.
if i % args.print_freq_val == 0:
print('Validation: [{0}/{1}]\t'
'Batch Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Top-5 Accuracy {top5.val:.3f} ({top5.avg:.3f})\t'.format(i, len(val_loader),
batch_time=batch_time,
loss=losses, top5=top5accs))
continue
if scene_graph:
(obj, rel, caps, caplens, orig_caps, obj_mask, rel_mask, pair_idx) = sample
obj = obj.to(device)
rel = rel.to(device)
obj_mask = obj_mask.to(device)
rel_mask = rel_mask.to(device)
pair_idx = pair_idx.to(device)
else:
(imgs, caps, caplens, orig_caps) = sample
imgs = imgs.to(device)
# Move to device, if available
caps = caps.to(device)
caplens = caplens.to(device)
# Forward prop.
if scene_graph:
scores, scores_d, caps_sorted, decode_lengths, sort_ind = decoder(object_features=obj,
relation_features=rel,
encoded_captions=caps,
caption_lengths=caplens,
object_mask=obj_mask,
relation_mask=rel_mask,
rel_pair_idx=pair_idx)
else:
scores, scores_d, caps_sorted, decode_lengths, sort_ind = decoder(imgs, caps, caplens)
# Max-pooling across predicted words across time steps for discriminative supervision
scores_d = scores_d.max(1)[0]
# Since we decoded starting with <start>, the targets are all words after <start>, up to <end>
targets = caps_sorted[:, 1:]
targets_d = torch.zeros(scores_d.size(0), scores_d.size(1)).to(device)
targets_d.fill_(-1)
for length in decode_lengths:
targets_d[:, :length - 1] = targets[:, :length - 1]
# Remove timesteps that we didn't decode at, or are pads
# pack_padded_sequence is an easy trick to do this
scores_copy = scores.clone()
scores = pack_padded_sequence(scores, decode_lengths, batch_first=True, enforce_sorted=True).data
targets = pack_padded_sequence(targets, decode_lengths, batch_first=True, enforce_sorted=True).data
#scores, _ = pack_padded_sequence(scores, decode_lengths, batch_first=True)
#targets, _ = pack_padded_sequence(targets, decode_lengths, batch_first=True)
# Calculate loss
loss_d = criterion_dis(scores_d, targets_d.long())
loss_g = criterion_ce(scores, targets)
loss = loss_g + (10 * loss_d)
# Keep track of metrics
losses.update(loss.item(), sum(decode_lengths))
top5 = accuracy(scores, targets, 5)
top5accs.update(top5, sum(decode_lengths))
batch_time.update(time.time() - start)
start = time.time()
if i % args.print_freq_val == 0:
print('Validation: [{0}/{1}]\t'
'Batch Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Top-5 Accuracy {top5.val:.3f} ({top5.avg:.3f})\t'.format(i, len(val_loader),
batch_time=batch_time,
loss=losses, top5=top5accs))
# Store references (true captions), and hypothesis (prediction) for each image
# If for n images, we have n hypotheses, and references a, b, c... for each image, we need -
# references = [[ref1a, ref1b, ref1c], [ref2a, ref2b], ...], hypotheses = [hyp1, hyp2, ...]
# References
assert (len(sort_ind) == 1), "Cannot have batch_size>1 for validation."
# a reference is a list of lists:
# [['the', 'cat', 'sat', 'on', 'the', 'mat'], ['a', 'cat', 'on', 'the', 'mat']]
references.append(orig_caps)
# Hypotheses
_, preds = torch.max(scores_copy, dim=2)
preds = preds.tolist()
preds_idxs_no_pads = list()
for j, p in enumerate(preds):
preds_idxs_no_pads.append(preds[j][:decode_lengths[j]]) # remove pads
preds_idxs_no_pads = list(map(lambda c: [w for w in c if w not in {word_map['<start>'],
word_map['<pad>']}],
preds_idxs_no_pads))
temp_preds = list()
# remove <start> and pads and convert idxs to string
for hyp in preds_idxs_no_pads:
temp_preds.append([])
for w in hyp:
assert (not w == word_map['pad']), "Should have removed all pads."
if not w == word_map['<start>']:
temp_preds[-1].append(word_map_inv[w])
preds = temp_preds
hypotheses.extend(preds)
assert len(references) == len(hypotheses)
# Calculate BLEU-4 scores
# bleu4 = corpus_bleu(references, hypotheses)
# bleu4 = round(bleu4, 4)
# compute the metrics
hypotheses_file = os.path.join(args.outdir, 'hypotheses', 'Epoch{:0>3d}.Hypotheses.json'.format(epoch))
references_file = os.path.join(args.outdir, 'references', 'Epoch{:0>3d}.References.json'.format(epoch))
create_captions_file(range(len(hypotheses)), hypotheses, hypotheses_file)
create_captions_file(range(len(references)), references, references_file)
coco = COCO(references_file)
# add the predicted results to the object
coco_results = coco.loadRes(hypotheses_file)
# create the evaluation object with both the ground-truth and the predictions
coco_eval = COCOEvalCap(coco, coco_results)
# change to use the image ids in the results object, not those from the ground-truth
coco_eval.params['image_id'] = coco_results.getImgIds()
# run the evaluation
coco_eval.evaluate(verbose=False, metrics=['bleu', 'meteor', 'rouge', 'cider'])
# Results contains: "Bleu_1", "Bleu_2", "Bleu_3", "Bleu_4", "METEOR", "ROUGE_L", "CIDEr", "SPICE"
results = coco_eval.eval
results['loss'] = losses.avg
results['top5'] = top5accs.avg
for k, v in results.items():
print(k+':\t'+str(v))
# print('\n * LOSS - {loss.avg:.3f}, TOP-5 ACCURACY - {top5.avg:.3f}, BLEU-4 - {bleu}, CIDEr - {cider}\n'
# .format(loss=losses, top5=top5accs, bleu=round(results['Bleu_4'], 4), cider=round(results['CIDEr'], 1)))
return results
def val_epoch(epoch, data_loader, model, criterion, opt, logger, writer):
print('validation at epoch {}'.format(epoch))
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
recalls = AverageMeter()
writer = writer
end_time = time.time()
for i, (inputs, labels) in enumerate(data_loader):
data_time.update(time.time() - end_time)
labels = list(map(int, labels))
inputs = torch.unsqueeze(inputs, 1)
inputs = inputs.type(torch.FloatTensor)
if not opt.no_cuda:
labels = torch.LongTensor(labels).cuda(async=True)
with torch.no_grad():
inputs = Variable(inputs)
labels = Variable(labels)
outputs = model(inputs)
loss = criterion(outputs, labels)
acc = calculate_accuracy(outputs, labels)
recall = calculate_recall(outputs, labels)
losses.update(loss.data, inputs.size(0))
accuracies.update(acc, inputs.size(0))
recalls.update(recall, inputs.size(0))
batch_time.update(time.time() - end_time)
end_time = time.time()
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})\t'
'Recall {recall.val:.3f} ({recall.avg:.3f})'.format(
epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies,
recall=recalls))
logger.log({
'epoch': epoch,
'loss': round(losses.avg.item(), 4),
'acc': round(accuracies.avg.item(), 4),
'recall': round(recalls.avg.item(), 4)
})
writer.add_scalar('val/loss', losses.avg, epoch)
writer.add_scalar('val/accuracy', accuracies.avg, epoch)
writer.add_scalar('val/recall', recalls.avg, epoch)
return losses.avg
def train_one_epoch_mixup(train_loader,
mix_loader,
model,
criterion,
optimizer,
epoch,
meters,
since,
alpha=0.4,
log=None):
losses = AverageMeter()
f1 = AverageMeter()
model.train()
if len(meters['f1']):
previous_loss = meters['loss'][-1]
previous_f1 = meters['f1'][-1]
best_f1_epoch = np.argmax(meters['f1'])
best_f1_score = meters['f1'][best_f1_epoch]
best_loss_epoch = np.argmin(meters['loss'])
best_loss = meters['loss'][best_loss_epoch]
else:
best_f1_epoch = 0
best_f1_score = 0
best_loss_epoch = 0
best_loss = 0
previous_loss = 0
previous_f1 = 0
for batch_id, ((x1, y1), (x2,
y2)) in enumerate(zip(train_loader, mix_loader)):
batch_x1 = x1.cuda(non_blocking=True)
batch_x2 = x2.cuda(non_blocking=True)
lam = np.random.beta(alpha, alpha)
batch_x = lam * batch_x1 + (1.0 - lam) * batch_x2
batch_y1 = torch.Tensor(np.array(y1)).float().cuda(non_blocking=True)
batch_y2 = torch.Tensor(np.array(y2)).float().cuda(non_blocking=True)
batch_y = lam * batch_y1 + (1.0 - lam) * batch_y2
output = model(batch_x)
loss = criterion(output, batch_y)
losses.update(loss.item(), batch_x.size(0))
f1_batch = f1_score(batch_y.cpu() > 0.5,
output.sigmoid().cpu() > 0.15,
average='macro')
f1.update(f1_batch, batch_x.size(0))
optimizer.zero_grad()
loss.backward()
if cfg.grident_clip:
torch.nn.utils.clip_grad_norm(model.parameters(), 1.)
optimizer.step()
print('Epoch %3d\t' % epoch,
'Batch %3d|%3d\t' % (batch_id, len(train_loader)),
'Loss: %10.5f\t' % losses.avg,
'Metrics|F1 Score: %10.5f\t' % f1.avg,
'Previous Loss: %10.5f\t' % previous_loss,
'Previous F1 Score: %10.5f\t' % previous_f1,
'Best loss:%10.5f Epoch %3d\t' % (best_loss, best_loss_epoch),
'Besr F1:%10.5f Epoch %3d\t' % (best_f1_score, best_f1_epoch),
'Time: %s' % time_to_str((timer() - since), 'min'),
file=log)
meters['loss'].append(losses.avg)
meters['f1'].append(f1.avg)
return meters
def train(train_loader, model, criterion, optimizer, epoch, args, logger,
writer, local_rank):
batch_times = AverageMeter('Time', ':6.3f')
data_times = AverageMeter('Data', ':6.3f')
losses = AverageMeter('Loss', ':.4e') # 4e表示科学记数法中的4位小数
top1 = AverageMeter('Acc@1', ':6.2f')
# switch to train mode
model.train()
end = time.time()
for i, (images, target) in enumerate(train_loader):
# measure data loading time
data_time = time.time() - end
data_times.update(data_time)
images = images.cuda(local_rank, non_blocking=True)
target = target.cuda(local_rank, non_blocking=True)
# compute output
output = model(images)
loss = criterion(output, target)
# measure accuracy and record loss
acc1 = accuracy(output, target, 1)
# DDP: data synchronization
dist.barrier()
reduced_loss = reduce_mean(loss, args.nprocs)
reduced_acc1 = reduce_mean(acc1, args.nprocs)
losses.update(reduced_loss.item(), images.size(0))
top1.update(reduced_acc1, images.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_times.update(time.time() - end)
end = time.time()
if i % args.print_freq == 0:
ddp_print(
'Train epoch: [{:d}/{:d}][{:d}/{:d}]\tlr={:.6f}\tce_loss={:.4f}\ttop1_acc={:.4f}\tdata_time={:6.3f}s'
'\tbatch_time={:6.3f}s'.format(epoch, args.epochs, i,
len(train_loader),
get_learning_rate(optimizer),
losses.avg, top1.avg,
data_times.avg,
batch_times.avg), logger,
local_rank)
break
ddp_print(
'||==> Train epoch: [{:d}/{:d}]\tlr={:.6f}\tce_loss={:.4f}\ttop1_acc={:.4f}\tbatch_time={:6.3f}s'
.format(epoch, args.epochs, get_learning_rate(optimizer), losses.avg,
top1.avg, batch_times.avg), logger, local_rank)
if args.local_rank == 0:
# save tensorboard
writer.add_scalar('lr', get_learning_rate(optimizer), epoch)
writer.add_scalar('Train_ce_loss', losses.avg, epoch)
writer.add_scalar('Train_top1_accuracy', top1.avg, epoch)
def test(val_loader, model, epoch, use_cuda):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
model.eval()
save_objs = args.evaluate
import os
if not os.path.exists(args.save_path):
os.makedirs(
args.save_path) # /scratch/xiaolonw/davis_results_mask_mixfcn/')
# save_path = '/scratch/xiaolonw/davis_results_mask_mixfcn/'
save_path = args.save_path + '/'
# img_path = '/scratch/xiaolonw/vlog_frames/'
save_file = '%s/list.txt' % save_path
fileout = open(save_file, 'w')
end = time.time()
# bar = Bar('Processing', max=len(val_loader))
for batch_idx, (imgs_total, patch2_total, lbls,
meta) in enumerate(tqdm(val_loader)):
finput_num_ori = params['videoLen']
finput_num = finput_num_ori
# measure data loading time
data_time.update(time.time() - end)
imgs_total = torch.autograd.Variable(imgs_total.cuda())
# patch2_total = torch.autograd.Variable(patch2_total.cuda())
t00 = time.time()
bs = imgs_total.size(0)
total_frame_num = imgs_total.size(1)
channel_num = imgs_total.size(2)
height_len = imgs_total.size(3)
width_len = imgs_total.size(4)
assert (bs == 1)
folder_paths = meta['folder_path']
gridx = int(meta['gridx'].data.cpu().numpy()[0])
gridy = int(meta['gridy'].data.cpu().numpy()[0])
print('gridx: ' + str(gridx) + ' gridy: ' + str(gridy))
print('total_frame_num: ' + str(total_frame_num))
height_dim = int(params['cropSize'] / 8)
width_dim = int(params['cropSize'] / 8)
# processing labels
lbls = lbls[0].data.cpu().numpy()
print(lbls.shape)
# print(patch2_total.size())
lbls_new = []
lbl_set = []
lbl_set.append(np.zeros(3).astype(np.uint8))
count_lbls = []
count_lbls.append(0)
for i in range(lbls.shape[0]):
nowlbl = lbls[i].copy()
if i == 0:
for j in range(nowlbl.shape[0]):
for k in range(nowlbl.shape[1]):
pixellbl = nowlbl[j, k, :].astype(np.uint8)
flag = 0
for t in range(len(lbl_set)):
if lbl_set[t][0] == pixellbl[0] and lbl_set[t][
1] == pixellbl[1] and lbl_set[t][
2] == pixellbl[2]:
flag = 1
count_lbls[t] = count_lbls[t] + 1
break
if flag == 0:
lbl_set.append(pixellbl)
count_lbls.append(0)
lbls_new.append(nowlbl)
lbl_set_temp = []
for i in range(len(lbl_set)):
if count_lbls[i] > 10:
lbl_set_temp.append(lbl_set[i])
lbl_set = lbl_set_temp
print(lbl_set)
print(count_lbls)
t01 = time.time()
lbls_resize = np.zeros(
(lbls.shape[0], lbls.shape[1], lbls.shape[2], len(lbl_set)))
lbls_resize2 = np.zeros(
(lbls.shape[0], height_dim, width_dim, len(lbl_set)))
for i in range(lbls.shape[0]):
nowlbl = lbls[i].copy()
for j in range(nowlbl.shape[0]):
for k in range(nowlbl.shape[1]):
pixellbl = nowlbl[j, k, :].astype(np.uint8)
for t in range(len(lbl_set)):
if lbl_set[t][0] == pixellbl[0] and lbl_set[t][
1] == pixellbl[1] and lbl_set[t][
2] == pixellbl[2]:
lbls_resize[i, j, k, t] = 1
for i in range(lbls.shape[0]):
lbls_resize2[i] = cv2.resize(lbls_resize[i],
(height_dim, width_dim))
t02 = time.time()
print(t02 - t01, 'relabel', t01 - t00, 'label')
# print the images
imgs_set = imgs_total.data
imgs_set = imgs_set.cpu().numpy()
imgs_set = imgs_set[0]
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
imgs_toprint = []
# ref image
for t in range(imgs_set.shape[0]):
img_now = imgs_set[t]
for c in range(3):
img_now[c] = img_now[c] * std[c]
img_now[c] = img_now[c] + mean[c]
img_now = img_now * 255
img_now = np.transpose(img_now, (1, 2, 0))
img_now = cv2.resize(img_now,
(img_now.shape[0] * 2, img_now.shape[1] * 2))
imgs_toprint.append(img_now)
imname = save_path + str(batch_idx) + '_' + str(t) + '_frame.jpg'
scipy.misc.imsave(imname, img_now)
for t in range(finput_num_ori):
nowlbl = lbls_new[t]
imname = save_path + str(batch_idx) + '_' + str(t) + '_label.jpg'
scipy.misc.imsave(imname, nowlbl)
# now_batch_size = 4
now_batch_size = 1 # we use one gpu for eval
imgs_stack = []
patch2_stack = []
im_num = total_frame_num - finput_num_ori
trans_out_2_set = []
corrfeat2_set = []
imgs_tensor = torch.Tensor(now_batch_size, finput_num, 3,
params['cropSize'], params['cropSize'])
target_tensor = torch.Tensor(now_batch_size, 1, 3, params['cropSize'],
params['cropSize'])
imgs_tensor = torch.autograd.Variable(imgs_tensor.cuda())
target_tensor = torch.autograd.Variable(target_tensor.cuda())
t03 = time.time()
for iter in range(0, im_num, now_batch_size):
# print(iter)
startid = iter
endid = iter + now_batch_size
if endid > im_num:
endid = im_num
now_batch_size2 = endid - startid
for i in range(now_batch_size2):
imgs = imgs_total[:, iter + i + 1:iter + i +
finput_num_ori, :, :, :]
imgs2 = imgs_total[:, 0, :, :, :].unsqueeze(1)
imgs = torch.cat((imgs2, imgs), dim=1)
imgs_tensor[i] = imgs
target_tensor[i, 0] = imgs_total[0, iter + i + finput_num_ori]
corrfeat2_now = model(imgs_tensor, target_tensor)
corrfeat2_now = corrfeat2_now.view(now_batch_size, finput_num_ori,
corrfeat2_now.size(1),
corrfeat2_now.size(2),
corrfeat2_now.size(3))
for i in range(now_batch_size2):
corrfeat2_set.append(corrfeat2_now[i].data.cpu().numpy())
t04 = time.time()
print(t04 - t03, 'model forward', t03 - t02, 'image prep')
for iter in range(total_frame_num - finput_num_ori):
if iter % 10 == 0:
print(iter)
imgs = imgs_total[:, iter + 1:iter + finput_num_ori, :, :, :]
imgs2 = imgs_total[:, 0, :, :, :].unsqueeze(1)
imgs = torch.cat((imgs2, imgs), dim=1)
# trans_out_2, corrfeat2 = model(imgs, patch2)
corrfeat2 = corrfeat2_set[iter]
corrfeat2 = torch.from_numpy(corrfeat2)
out_frame_num = int(finput_num)
height_dim = corrfeat2.size(2)
width_dim = corrfeat2.size(3)
corrfeat2 = corrfeat2.view(corrfeat2.size(0), height_dim,
width_dim, height_dim, width_dim)
corrfeat2 = corrfeat2.data.cpu().numpy()
topk_vis = args.topk_vis
vis_ids_h = np.zeros((corrfeat2.shape[0], height_dim, width_dim,
topk_vis)).astype(np.int)
vis_ids_w = np.zeros((corrfeat2.shape[0], height_dim, width_dim,
topk_vis)).astype(np.int)
t05 = time.time()
atten1d = corrfeat2.reshape(corrfeat2.shape[0],
height_dim * width_dim, height_dim,
width_dim)
ids = np.argpartition(atten1d, -topk_vis, axis=1)[:, -topk_vis:]
# ids = np.argsort(atten1d, axis=1)[:, -topk_vis:]
hid = ids // width_dim
wid = ids % width_dim
vis_ids_h = wid.transpose(0, 2, 3, 1)
vis_ids_w = hid.transpose(0, 2, 3, 1)
t06 = time.time()
img_now = imgs_toprint[iter + finput_num_ori]
predlbls = np.zeros((height_dim, width_dim, len(lbl_set)))
# predlbls2 = np.zeros((height_dim * width_dim, len(lbl_set)))
for t in range(finput_num):
tt1 = time.time()
h, w, k = np.meshgrid(np.arange(height_dim),
np.arange(width_dim),
np.arange(topk_vis),
indexing='ij')
h, w = h.flatten(), w.flatten()
hh, ww = vis_ids_h[t].flatten(), vis_ids_w[t].flatten()
if t == 0:
lbl = lbls_resize2[0, hh, ww, :]
else:
lbl = lbls_resize2[t + iter, hh, ww, :]
np.add.at(predlbls, (h, w),
lbl * corrfeat2[t, ww, hh, h, w][:, None])
t07 = time.time()
# print(t07-t06, 'lbl proc', t06-t05, 'argsorts')
predlbls = predlbls / finput_num
for t in range(len(lbl_set)):
nowt = t
predlbls[:, :,
nowt] = predlbls[:, :, nowt] - predlbls[:, :,
nowt].min()
predlbls[:, :,
nowt] = predlbls[:, :, nowt] / predlbls[:, :,
nowt].max()
lbls_resize2[iter + finput_num_ori] = predlbls
predlbls_cp = predlbls.copy()
predlbls_cp = cv2.resize(predlbls_cp,
(params['imgSize'], params['imgSize']))
predlbls_val = np.zeros((params['imgSize'], params['imgSize'], 3))
ids = np.argmax(predlbls_cp[:, :, 1:len(lbl_set)], 2)
predlbls_val = np.array(lbl_set)[np.argmax(predlbls_cp, axis=-1)]
predlbls_val = predlbls_val.astype(np.uint8)
predlbls_val2 = cv2.resize(predlbls_val,
(img_now.shape[0], img_now.shape[1]),
interpolation=cv2.INTER_NEAREST)
# activation_heatmap = cv2.applyColorMap(predlbls, cv2.COLORMAP_JET)
img_with_heatmap = np.float32(img_now) * 0.5 + np.float32(
predlbls_val2) * 0.5
imname = save_path + str(batch_idx) + '_' + str(
iter + finput_num_ori) + '_label.jpg'
imname2 = save_path + str(batch_idx) + '_' + str(
iter + finput_num_ori) + '_mask.png'
scipy.misc.imsave(imname, np.uint8(img_with_heatmap))
scipy.misc.imsave(imname2, np.uint8(predlbls_val))
fileout.close()
return losses.avg
def train(trainloader, model, criterion, optimizer, epoch, use_cuda):
# switch to train mode
model.train()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
for batch_idx, (inputs, targets) in enumerate(trainloader):
if use_cuda:
inputs, targets = inputs.cuda(local_rank), targets.cuda(local_rank, async=True)
inputs, targets = torch.autograd.Variable(inputs), torch.autograd.Variable(targets)
# compute output
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))
losses.update(loss.data.item(), inputs.size(0))
top1.update(prec1.item(), inputs.size(0))
top5.update(prec5.item(), inputs.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
return (losses.avg, top1.avg)
def train(model,
train_dataloader,
lr=0.01,
save_dir='./weights',
num_epoches=200,
model_name="",
valid_dataloader=None,
batch_num=None,
train_type="clean",
random_layer=None):
if random_layer:
model_name = model_name + "_randomLayer"
print(train_dataloader, valid_dataloader)
optimizer = torch.optim.SGD(model.parameters(),
lr=lr,
momentum=0.9,
nesterov=True,
weight_decay=0.0005)
cost = nn.CrossEntropyLoss()
batch_time = AverageMeter()
losses = AverageMeter()
error = AverageMeter()
end = time.time()
best_error = 1.0
nobetter_num = 1
for epoch in range(num_epoches):
if nobetter_num >= 5:
print("train done .lr={},best_error={}".format(lr, best_error))
break
if nobetter_num >= 3:
lr = lr * 0.1
for param_group in optimizer.param_groups:
param_group['lr'] = lr
_, train_loss, train_error = train_epoch(
model,
train_dataloader if train_type == "clean" else combind_loader(
*train_dataloader),
optimizer,
cost,
epoch,
num_epoches,
batch_num=batch_num,
random_layer=random_layer)
if valid_dataloader:
with torch.no_grad():
_, valid_loss, valid_error = valid_epoch(
model,
valid_dataloader if train_type == "clean" else
combind_loader(*valid_dataloader),
cost,
batch_num=len(valid_dataloader)
if train_type == "clean" else len(valid_dataloader[0]),
random_layer=random_layer)
if valid_dataloader and valid_error < best_error:
best_error = valid_error
if valid_error + 0.005 < best_error:
nobetter_num += 1
else:
nobetter_num = 1
print('New best error: %.4f' % best_error)
torch.save(model.state_dict(),
os.path.join(save_dir, model_name + '_model.dat'))
else:
#torch.save(model.state_dict(), os.path.join(save_dir, 'vgg16_model.dat'))
nobetter_num += 1
with open(os.path.join(save_dir, model_name + '_results.csv'),
'a') as f:
f.write('%03d,%0.6f,%0.6f,%0.5f,%0.5f,\n' % (
(epoch + 1),
train_loss,
train_error,
valid_loss,
valid_error,
))
def train(train_loader, model, criterion, optimizer, epoch, log):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to train mode
model.train()
end = time.time()
for i, (input, target) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
if args.use_cuda:
target = target.cuda(async=True)
input = input.cuda()
input_var = torch.autograd.Variable(input)
target_var = torch.autograd.Variable(target)
# compute output
output = model(input_var)
loss = criterion(output, target_var)
# measure accuracy and record loss
prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
losses.update(loss.data[0], input.size(0))
top1.update(prec1[0], input.size(0))
top5.update(prec5[0], input.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
print_log(' Epoch: [{:03d}][{:03d}/{:03d}] '
'Time {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Data {data_time.val:.3f} ({data_time.avg:.3f}) '
'Loss {loss.val:.4f} ({loss.avg:.4f}) '
'Prec@1 {top1.val:.3f} ({top1.avg:.3f}) '
'Prec@5 {top5.val:.3f} ({top5.avg:.3f}) '.format(
epoch, i, len(train_loader), batch_time=batch_time,
data_time=data_time, loss=losses, top1=top1, top5=top5) + time_string(), log)
return top1.avg, losses.avg
def validate(val_loader, model, use_val_th=False):
batch_time = AverageMeter()
# switch to evaluate mode
model.eval()
with torch.no_grad():
end = time.time()
for i, (inputs, targets, infos) in enumerate(val_loader):
if default.generate_features_all:
logger.info('generating features, batch %d', i)
filenames = [info[0] for info in infos]
lesion_idxs = [info[1] for info in infos]
inputs = [input.cuda() for input in inputs]
unc_targets = targets[1]
targets = targets[0]
# compute output
out = model(inputs)
if config.SCORE_PROPAGATION:
prob_np = out['class_prob2'].detach().cpu().numpy()
scores_np = out['class_score2'].detach().cpu().numpy()
else:
prob_np = out['class_prob1'].detach().cpu().numpy()
scores_np = out['class_score1'].detach().cpu().numpy()
target1 = targets.numpy() > 0
pred_wt = unc_targets.numpy() == 0
if i == 0:
target_all = target1
prob_all = prob_np
score_all = scores_np
lesion_idx_all = lesion_idxs
pred_wt_all = pred_wt
if default.generate_features_all:
ft_all = out['emb']
else:
target_all = np.vstack((target_all, target1))
prob_all = np.vstack((prob_all, prob_np))
score_all = np.vstack((score_all, scores_np))
pred_wt_all = np.vstack((pred_wt_all, pred_wt))
lesion_idx_all.extend(lesion_idxs)
if default.generate_features_all:
ft_all = np.vstack((ft_all, out['emb']))
if default.generate_features_all:
save_ft_to_file(ft_all)
assert 0, 'all features have been generated and saved.'
if config.TEST.USE_CALIBRATED_TH:
accs, pred_label_all = compute_all_acc_wt_th(target_all, prob_all, pred_wt_all, use_val_th)
else:
pred_label_all = score2label(prob_all, config.TEST.SCORE_PARAM)
accs = compute_all_acc_wt(target_all, pred_label_all, prob_all, pred_wt_all)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % default.frequent == 0:
logger.info('Test: [{0}/{1}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'{crit} {accs:.3f}'
.format(
i, len(val_loader), batch_time=batch_time, crit=config.TEST.CRITERION,
accs=accs[config.TEST.CRITERION]
))
print_accs(accs)
accs['ex_neg'] = np.sum((target_all == 0) & pred_wt_all, axis=0)
if use_val_th: # only save for test set not val set
save_acc_to_file(accs, val_loader, 'all_terms')
if default.mode == 'infer' and use_val_th:
save_test_scores_to_file(score_all, pred_label_all, target_all, accs, lesion_idx_all)
return accs
def main():
if not os.path.isdir(args.save_path): os.makedirs(args.save_path)
log = open(os.path.join(args.save_path, 'log_seed_{}.txt'.format(args.manualSeed)), 'w')
print_log('save path : {}'.format(args.save_path), log)
state = {k: v for k, v in args._get_kwargs()}
print_log(state, log)
print_log("Random Seed: {}".format(args.manualSeed), log)
print_log("python version : {}".format(sys.version.replace('\n', ' ')), log)
print_log("torch version : {}".format(torch.__version__), log)
print_log("cudnn version : {}".format(torch.backends.cudnn.version()), log)
# Init dataset
if not os.path.isdir(args.data_path):
os.makedirs(args.data_path)
if args.dataset == 'cifar10':
mean = [x / 255 for x in [125.3, 123.0, 113.9]]
std = [x / 255 for x in [63.0, 62.1, 66.7]]
elif args.dataset == 'cifar100':
mean = [x / 255 for x in [129.3, 124.1, 112.4]]
std = [x / 255 for x in [68.2, 65.4, 70.4]]
else:
assert False, "Unknow dataset : {}".format(args.dataset)
train_transform = transforms.Compose(
[transforms.RandomHorizontalFlip(), transforms.RandomCrop(32, padding=4), transforms.ToTensor(),
transforms.Normalize(mean, std)])
test_transform = transforms.Compose(
[transforms.ToTensor(), transforms.Normalize(mean, std)])
if args.dataset == 'cifar10':
train_data = dset.CIFAR10(args.data_path, train=True, transform=train_transform, download=True)
test_data = dset.CIFAR10(args.data_path, train=False, transform=test_transform, download=True)
num_classes = 10
elif args.dataset == 'cifar100':
train_data = dset.CIFAR100(args.data_path, train=True, transform=train_transform, download=True)
test_data = dset.CIFAR100(args.data_path, train=False, transform=test_transform, download=True)
num_classes = 100
elif args.dataset == 'svhn':
train_data = dset.SVHN(args.data_path, split='train', transform=train_transform, download=True)
test_data = dset.SVHN(args.data_path, split='test', transform=test_transform, download=True)
num_classes = 10
elif args.dataset == 'stl10':
train_data = dset.STL10(args.data_path, split='train', transform=train_transform, download=True)
test_data = dset.STL10(args.data_path, split='test', transform=test_transform, download=True)
num_classes = 10
elif args.dataset == 'imagenet':
assert False, 'Do not finish imagenet code'
else:
assert False, 'Do not support dataset : {}'.format(args.dataset)
train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
test_loader = torch.utils.data.DataLoader(test_data, batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
# Init model, criterion, and optimizer
#net = models.__dict__[args.arch](num_classes).cuda()
net = SENet34()
# define loss function (criterion) and optimizer
criterion = F.nll_loss
optimizer = torch.optim.SGD(net.parameters(), state['learning_rate'], momentum=state['momentum'],
weight_decay=state['decay'], nesterov=True)
if args.use_cuda: net.cuda()
recorder = RecorderMeter(args.epochs)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print_log("=> loading checkpoint '{}'".format(args.resume), log)
checkpoint = torch.load(args.resume)
recorder = checkpoint['recorder']
args.start_epoch = checkpoint['epoch']
net.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print_log("=> loaded checkpoint '{}' (epoch {})" .format(args.resume, checkpoint['epoch']), log)
else:
print_log("=> no checkpoint found at '{}'".format(args.resume), log)
else:
print_log("=> do not use any checkpoint for model", log)
if args.evaluate:
validate(test_loader, net, criterion, log)
return
# Main loop
start_time = time.time()
epoch_time = AverageMeter()
for epoch in range(args.start_epoch, args.epochs):
current_learning_rate = adjust_learning_rate(optimizer, epoch, args.gammas, args.schedule)
need_hour, need_mins, need_secs = convert_secs2time(epoch_time.avg * (args.epochs-epoch))
need_time = '[Need: {:02d}:{:02d}:{:02d}]'.format(need_hour, need_mins, need_secs)
print_log('\n==>>{:s} [Epoch={:03d}/{:03d}] {:s} [learning_rate={:6.4f}]'.format(time_string(), epoch, args.epochs, need_time, current_learning_rate) \
+ ' [Best : Accuracy={:.2f}, Error={:.2f}]'.format(recorder.max_accuracy(False), 100-recorder.max_accuracy(False)), log)
# train for one epoch
train_acc, train_los = train(train_loader, net, criterion, optimizer, epoch, log)
# evaluate on validation set
val_acc, val_los = validate(test_loader, net, criterion, log)
is_best = recorder.update(epoch, train_los, train_acc, val_los, val_acc)
save_checkpoint({
'epoch': epoch + 1,
'state_dict': net.state_dict(),
'recorder': recorder,
'optimizer' : optimizer.state_dict(),
}, is_best, args.save_path, 'checkpoint.pth.tar')
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
recorder.plot_curve( os.path.join(args.save_path, 'curve.png') )
log.close()
def val_epoch(epoch, data_loader, model, criterion, opt, logger):
print('validation at epoch {}'.format(epoch))
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
end_time = time.time()
for i, (inputs, targets) in enumerate(data_loader):
data_time.update(time.time() - end_time)
if not opt.no_cuda:
targets = targets.cuda(async=True)
inputs = Variable(inputs, volatile=True)
targets = Variable(targets, volatile=True)
outputs = model(inputs)
loss = criterion(outputs, targets)
acc = calculate_accuracy(outputs, targets)
losses.update(loss.data[0], inputs.size(0))
accuracies.update(acc, inputs.size(0))
batch_time.update(time.time() - end_time)
end_time = time.time()
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})'.format(
epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies))
logger.log({'epoch': epoch, 'loss': losses.avg, 'acc': accuracies.avg})
return losses.avg
def validate(val_loader, net, criterion, optimizer, epoch, train_args, restore, visualize):
net.eval()
val_loss = AverageMeter()
inputs_all, gts_all, predictions_all = [], [], []
for vi, data in enumerate(val_loader):
inputs, gts = data
N = inputs.size(0)
inputs = Variable(inputs, volatile=True).cuda()
gts = Variable(gts, volatile=True).cuda()
outputs = net(inputs)
predictions = outputs.data.max(1)[1].squeeze_(1).squeeze_(0).cpu().numpy()
val_loss.update(criterion(outputs, gts).data[0] / N, N)
if random.random() > train_args['val_img_sample_rate']:
inputs_all.append(None)
else:
inputs_all.append(inputs.data.squeeze_(0).cpu())
gts_all.append(gts.data.squeeze_(0).cpu().numpy())
predictions_all.append(predictions)
acc, acc_cls, mean_iu, fwavacc = evaluate(predictions_all, gts_all, voc.num_classes)
if mean_iu > train_args['best_record']['mean_iu']:
train_args['best_record']['val_loss'] = val_loss.avg
train_args['best_record']['epoch'] = epoch
train_args['best_record']['acc'] = acc
train_args['best_record']['acc_cls'] = acc_cls
train_args['best_record']['mean_iu'] = mean_iu
train_args['best_record']['fwavacc'] = fwavacc
snapshot_name = 'epoch_%d_loss_%.5f_acc_%.5f_acc-cls_%.5f_mean-iu_%.5f_fwavacc_%.5f_lr_%.10f' % (
epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc, optimizer.param_groups[1]['lr']
)
torch.save(net.state_dict(), os.path.join(ckpt_path, exp_name, snapshot_name + '.pth'))
torch.save(optimizer.state_dict(), os.path.join(ckpt_path, exp_name, 'opt_' + snapshot_name + '.pth'))
if train_args['val_save_to_img_file']:
to_save_dir = os.path.join(ckpt_path, exp_name, str(epoch))
check_mkdir(to_save_dir)
val_visual = []
for idx, data in enumerate(zip(inputs_all, gts_all, predictions_all)):
if data[0] is None:
continue
input_pil = restore(data[0])
gt_pil = voc.colorize_mask(data[1])
predictions_pil = voc.colorize_mask(data[2])
if train_args['val_save_to_img_file']:
input_pil.save(os.path.join(to_save_dir, '%d_input.png' % idx))
predictions_pil.save(os.path.join(to_save_dir, '%d_prediction.png' % idx))
gt_pil.save(os.path.join(to_save_dir, '%d_gt.png' % idx))
val_visual.extend([visualize(input_pil.convert('RGB')), visualize(gt_pil.convert('RGB')),
visualize(predictions_pil.convert('RGB'))])
val_visual = torch.stack(val_visual, 0)
val_visual = vutils.make_grid(val_visual, nrow=3, padding=5)
writer.add_image(snapshot_name, val_visual)
print('--------------------------------------------------------------------')
print('[epoch %d], [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f]' % (
epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc))
print('best record: [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f], [epoch %d]' % (
train_args['best_record']['val_loss'], train_args['best_record']['acc'], train_args['best_record']['acc_cls'],
train_args['best_record']['mean_iu'], train_args['best_record']['fwavacc'], train_args['best_record']['epoch']))
print('--------------------------------------------------------------------')
writer.add_scalar('val_loss', val_loss.avg, epoch)
writer.add_scalar('acc', acc, epoch)
writer.add_scalar('acc_cls', acc_cls, epoch)
writer.add_scalar('mean_iu', mean_iu, epoch)
writer.add_scalar('fwavacc', fwavacc, epoch)
writer.add_scalar('lr', optimizer.param_groups[1]['lr'], epoch)
net.train()
return val_loss.avg
def train(segmentation_module, iterator, optimizers, history, epoch, args):
batch_time = AverageMeter()
data_time = AverageMeter()
names = ['object', 'part', 'scene', 'material']
ave_losses = {n: AverageMeter() for n in names}
ave_metric = {n: AverageMeter() for n in names}
ave_losses['total'] = AverageMeter()
segmentation_module.train(not args.fix_bn)
# main loop
tic = time.time()
for i in range(args.epoch_iters):
batch_data, src_idx = next(iterator)
data_time.update(time.time() - tic)
segmentation_module.zero_grad()
# forward pass
ret = segmentation_module(batch_data)
# Backward
loss = ret['loss']['total'].mean()
loss.backward()
for optimizer in optimizers:
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - tic)
tic = time.time()
# measure losses
for name in ret['loss'].keys():
ave_losses[name].update(ret['loss'][name].mean().item())
# measure metrics
# NOTE: scene metric will be much lower than benchmark
for name in ret['metric'].keys():
ave_metric[name].update(ret['metric'][name].mean().item())
# calculate accuracy, and display
if i % args.disp_iter == 0:
loss_info = "Loss: total {:.4f}, ".format(ave_losses['total'].average())
loss_info += ", ".join(["{} {:.2f}".format(
n[0], ave_losses[n].average()
if ave_losses[n].average() is not None else 0) for n in names])
acc_info = "Accuracy: " + ", ".join(["{} {:4.2f}".format(
n[0], ave_metric[n].average()
if ave_metric[n].average() is not None else 0) for n in names])
print('Epoch: [{}][{}/{}], Time: {:.2f}, Data: {:.2f}, '
'LR: encoder {:.6f}, decoder {:.6f}, {}, {}'
.format(epoch, i, args.epoch_iters,
batch_time.average(), data_time.average(),
args.running_lr_encoder, args.running_lr_decoder,
acc_info, loss_info))
fractional_epoch = epoch - 1 + 1. * i / args.epoch_iters
history['train']['epoch'].append(fractional_epoch)
history['train']['loss'].append(loss.item())
# adjust learning rate
cur_iter = i + (epoch - 1) * args.epoch_iters
adjust_learning_rate(optimizers, cur_iter, args)
def train(self, epoch):
batch_time = AverageMeter()
losses = AverageMeter()
acc = AverageMeter()
self.scheduler.step()
self.model.train()
end = time.time()
lr = self.scheduler.get_lr()[0]
# for batch, (softmax_data, triplet_data) in enumerate(itertools.zip_longest(self.softmax_train_loader, self.triplet_train_loader)):
for batch, (softmax_data, triplet_data, gallery_data) in enumerate(zip(self.softmax_train_loader, self.triplet_train_loader, self.gallery_loader)):
loss = 0
# 1st
softmax_inputs, softmax_labels = softmax_data
# 转cuda
softmax_inputs = softmax_inputs.to(self.device) if torch.cuda.device_count() >= 1 else softmax_inputs
softmax_labels = softmax_labels.to(self.device) if torch.cuda.device_count() >= 1 else softmax_labels
softmax_score, softmax_outputs = self.model(softmax_inputs)
traditional_loss = self.softmax_loss(softmax_score, softmax_outputs, softmax_labels)
loss += traditional_loss
# total
losses.update(loss.item(), softmax_inputs.size(0))
prec = (softmax_score.max(1)[1] == softmax_labels).float().mean()
acc.update(prec, softmax_inputs.size(0))
# 2nd
triplet_inputs, triplet_labels = triplet_data
# 转cuda
triplet_inputs = triplet_inputs.to(self.device) if torch.cuda.device_count() >= 1 else triplet_inputs
triplet_labels = triplet_labels.to(self.device) if torch.cuda.device_count() >= 1 else triplet_labels
triplet_score, triplet_outputs = self.model(triplet_inputs)
triplet_loss = self.triplet_loss(triplet_score, triplet_outputs, triplet_labels)
loss += triplet_loss
# 3rd
gallery_inputs, gallery_labels = gallery_data
gallery_inputs = gallery_inputs.to(self.device) if torch.cuda.device_count() >= 1 else gallery_inputs
gallery_score, gallery_outputs = self.model(gallery_inputs)
query_feats = []
for query_inputs, query_labels in data.query_loader:
query_inputs = query_inputs.cuda()
query_score, query_outputs = model(query_inputs)
query_feats.append(query_outputs)
logger.debug('query_outputs: {}'.format(query_outputs.shape))
query_feats = torch.cat(query_feats, dim=0)
logger.debug('query_feats: {}'.format(query_feats.shape))
entropy = self.entropy_loss(gallery_outputs, query_feats)
loss += entropy
self.optimizer.zero_grad()
if opt.fp16: # we use optimier to backward loss
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# loss.backward(retain_graph=True)
self.optimizer.step()
# 评估训练耗时
batch_time.update(time.time() - end)
end = time.time()
# 打印耗时与结果
if (batch+1) % 10 == 0:
logger.debug('Epoch: [{}][{}/{}]\t'
'Base_lr: [{:.2e}]\t'
'Time: ({batch_time.avg:.3f})\t'
'Loss_val: {loss.val:.4f} (Loss_avg: {loss.avg:.4f})\t'
'Accuray_val: {acc.val:.4f} (Accuray_avg: {acc.avg:.4f})'.format(
epoch, batch+1, len(self.softmax_train_loader), lr, batch_time=batch_time, loss=losses, acc=acc))
# 每个epoch的结果
log_text = 'Epoch[{}]\tBase_lr {:.2e}\tAccuray {acc.avg:.4f}\tLoss {loss.avg:.4f}'.format(epoch, lr, acc=acc, loss=losses)
logger.info(log_text)
with open(log_file, 'a') as f:
f.write(log_text + '\n')
f.flush()
def main():
cnt_errors = {
'mae': AverageMeter(),
'mse': AverageMeter(),
'nae': AverageMeter(),
}
metrics_s = {
'tp': AverageMeter(),
'fp': AverageMeter(),
'fn': AverageMeter(),
'tp_c': AverageCategoryMeter(num_classes),
'fn_c': AverageCategoryMeter(num_classes)
}
metrics_l = {
'tp': AverageMeter(),
'fp': AverageMeter(),
'fn': AverageMeter(),
'tp_c': AverageCategoryMeter(num_classes),
'fn_c': AverageCategoryMeter(num_classes)
}
pred_data, gt_data = read_pred_and_gt(pred_file, gt_file)
for i_sample in id_std:
print(i_sample)
# init
gt_p,pred_p,fn_gt_index,tp_pred_index,fp_pred_index= [],[],[],[],[]
tp_s, fp_s, fn_s, tp_l, fp_l, fn_l = [0, 0, 0, 0, 0, 0]
tp_c_s = np.zeros([num_classes])
fn_c_s = np.zeros([num_classes])
tp_c_l = np.zeros([num_classes])
fn_c_l = np.zeros([num_classes])
if gt_data[i_sample]['num'] == 0 and pred_data[i_sample]['num'] != 0:
pred_p = pred_data[i_sample]['points']
fp_pred_index = np.array(range(pred_p.shape[0]))
fp_s = fp_pred_index.shape[0]
fp_l = fp_pred_index.shape[0]
if pred_data[i_sample]['num'] == 0 and gt_data[i_sample]['num'] != 0:
gt_p = gt_data[i_sample]['points']
level = gt_data[i_sample]['level']
fn_gt_index = np.array(range(gt_p.shape[0]))
fn_s = fn_gt_index.shape[0]
fn_l = fn_gt_index.shape[0]
for i_class in range(num_classes):
fn_c_s[i_class] = (level[fn_gt_index] == i_class).sum()
fn_c_l[i_class] = (level[fn_gt_index] == i_class).sum()
if gt_data[i_sample]['num'] != 0 and pred_data[i_sample]['num'] != 0:
pred_p = pred_data[i_sample]['points']
gt_p = gt_data[i_sample]['points']
sigma_s = gt_data[i_sample]['sigma'][:, 0]
sigma_l = gt_data[i_sample]['sigma'][:, 1]
level = gt_data[i_sample]['level']
# dist
dist_matrix = ss.distance_matrix(pred_p, gt_p, p=2)
match_matrix = np.zeros(dist_matrix.shape, dtype=bool)
# sigma_s and sigma_l
tp_s, fp_s, fn_s, tp_c_s, fn_c_s = compute_metrics(
dist_matrix, match_matrix, pred_p.shape[0], gt_p.shape[0],
sigma_s, level)
tp_l, fp_l, fn_l, tp_c_l, fn_c_l = compute_metrics(
dist_matrix, match_matrix, pred_p.shape[0], gt_p.shape[0],
sigma_l, level)
metrics_s['tp'].update(tp_s)
metrics_s['fp'].update(fp_s)
metrics_s['fn'].update(fn_s)
metrics_s['tp_c'].update(tp_c_s)
metrics_s['fn_c'].update(fn_c_s)
metrics_l['tp'].update(tp_l)
metrics_l['fp'].update(fp_l)
metrics_l['fn'].update(fn_l)
metrics_l['tp_c'].update(tp_c_l)
metrics_l['fn_c'].update(fn_c_l)
gt_count, pred_cnt = gt_data[i_sample]['num'], pred_data[i_sample][
'num']
s_mae = abs(gt_count - pred_cnt)
s_mse = (gt_count - pred_cnt) * (gt_count - pred_cnt)
cnt_errors['mae'].update(s_mae)
cnt_errors['mse'].update(s_mse)
if gt_count != 0:
s_nae = abs(gt_count - pred_cnt) / gt_count
cnt_errors['nae'].update(s_nae)
ap_s = metrics_s['tp'].sum / (metrics_s['tp'].sum + metrics_s['fp'].sum +
1e-20)
ar_s = metrics_s['tp'].sum / (metrics_s['tp'].sum + metrics_s['fn'].sum +
1e-20)
f1m_s = 2 * ap_s * ar_s / (ap_s + ar_s)
ar_c_s = metrics_s['tp_c'].sum / (metrics_s['tp_c'].sum +
metrics_s['fn_c'].sum + 1e-20)
ap_l = metrics_l['tp'].sum / (metrics_l['tp'].sum + metrics_l['fp'].sum +
1e-20)
ar_l = metrics_l['tp'].sum / (metrics_l['tp'].sum + metrics_l['fn'].sum +
1e-20)
f1m_l = 2 * ap_l * ar_l / (ap_l + ar_l)
ar_c_l = metrics_l['tp_c'].sum / (metrics_l['tp_c'].sum +
metrics_l['fn_c'].sum + 1e-20)
print('-----Localization performance-----')
print('AP_small: ' + str(ap_s))
print('AR_small: ' + str(ar_s))
print('F1m_small: ' + str(f1m_s))
print('AR_small_category: ' + str(ar_c_s))
print(' avg: ' + str(ar_c_s.mean()))
print('AP_large: ' + str(ap_l))
print('AR_large: ' + str(ar_l))
print('F1m_large: ' + str(f1m_l))
print('AR_large_category: ' + str(ar_c_l))
print(' avg: ' + str(ar_c_l.mean()))
mae = cnt_errors['mae'].avg
mse = np.sqrt(cnt_errors['mse'].avg)
nae = cnt_errors['nae'].avg
print('-----Counting performance-----')
print('MAE: ' + str(mae))
print('MSE: ' + str(mse))
print('NAE: ' + str(nae))
def valid_epoch(model,
valid_dataloader,
cost,
print_freq=40,
batch_num=None,
random_layer=None):
batch_time = AverageMeter()
losses = AverageMeter()
error = AverageMeter()
end = time.time()
model.eval()
for i, (images, labels, noises) in enumerate(valid_dataloader):
images, labels, noises = images.cuda(), labels.cuda(), noises.cuda()
noises = noises * args.phi
if random_layer:
images = random_layer(images)
outputs = model(images, noises)
loss = cost(outputs, labels)
batch_size = labels.size(0)
outputs = outputs.max(1)[1]
error.update(
torch.ne(outputs.cpu(), labels.cpu()).float().sum().item() /
batch_size, batch_size)
losses.update(loss.item(), batch_size)
batch_time.update(time.time() - end)
end = time.time()
if i % print_freq == 0:
res = '\t'.join([
'Valid',
'Iter: [%d/%d]' % (i + 1, batch_num),
'Time %.3f (%.3f)' % (batch_time.val, batch_time.avg),
'Loss %.4f (%.4f)' % (losses.val, losses.avg),
'Error %.4f (%.4f)' % (error.val, error.avg),
])
print(res)
return batch_time.avg, losses.avg, error.avg
def test(testloader, model, criterion, epoch, use_cuda):
global best_acc
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to evaluate mode
model.eval()
end = time.time()
start_index = 0
for batch_idx, (inputs, targets) in enumerate(testloader):
# measure data loading time
data_time.update(time.time() - end)
if use_cuda:
inputs, targets = inputs.cuda(), targets.cuda()
with torch.no_grad():
inputs, targets = torch.autograd.Variable(
inputs), torch.autograd.Variable(targets)
# compute output
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))
losses.update(loss.item(), inputs.size(0))
top1.update(prec1.item(), inputs.size(0))
top5.update(prec5.item(), inputs.size(0))
progress_bar(
batch_idx, len(testloader),
'Loss: %.2f | Top1: %.2f | Top5: %.2f' %
(losses.avg, top1.avg, top5.avg))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
return (losses.avg, top1.avg, top5.avg)
def train(train_loader, model, criterions, optimizer, epoch):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accs = AverageMeter()
# switch to train mode
model.train()
end = time.time()
for i, (inputs, targets, infos) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - end)
# construct targets
target_clsf, target_unc, target_ex = targets
target_conf = (target_clsf + target_ex) > 0
rhem_wt = torch.zeros_like(target_clsf).cuda()
rhem_wt[target_conf] = 1.
target_clsf = target_clsf.cuda()
target_clsf_wt = 1-target_unc.cuda()
# run model
inputs = [input.cuda() for input in inputs]
out = model(inputs)
# compute losses
emb = out['emb']
A, P, N = select_triplets_multilabel(emb, target_clsf)
loss_metric = criterions['metric'](A, P, N)
prob1 = out['class_prob1']
loss_ce1 = criterions['wce'](prob1, target_clsf, infos, wt=target_clsf_wt)
loss_rhem = criterions['rhem'](prob1, target_clsf, infos, wt=rhem_wt)
if config.SCORE_PROPAGATION:
prob2 = out['class_prob2']
loss_ce2 = criterions['wce'](prob2, target_clsf, infos, wt=target_clsf_wt)
sub_losses = [loss_ce1, loss_rhem, loss_metric, loss_ce2]
wts_names = ['CE_LOSS_WT_1', 'RHEM_LOSS_WT', 'TRIPLET_LOSS_WT', 'CE_LOSS_WT_2']
else:
sub_losses = [loss_ce1, loss_rhem, loss_metric]
wts_names = ['CE_LOSS_WT_1', 'RHEM_LOSS_WT', 'TRIPLET_LOSS_WT']
loss = 0
wts = [eval('config.TRAIN.' + name1) for name1 in wts_names]
for wt1, loss1 in zip(wts, sub_losses):
loss += wt1 * loss1
losses.update(loss.item())
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
clip_gradient(model, default.clip_gradient)
optimizer.step()
# measure accuracy
if config.SCORE_PROPAGATION:
prob_np = prob2.detach().cpu().numpy()
else:
prob_np = prob1.detach().cpu().numpy()
pred_labels = score2label(prob_np, config.TEST.SCORE_PARAM)
targets_np = target_clsf.detach().cpu().numpy()
target_unc = target_unc.numpy()
acc = compute_all_acc_wt(targets_np > 0, pred_labels, prob_np, target_unc == 0)[config.TEST.CRITERION]
accs.update(acc)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if i % default.frequent == 0:
crit = 'mean_pcF1' if config.TEST.CRITERION == 'mean_perclass_f1' else config.TEST.CRITERION
msg = 'Epoch: [{0}][{1}/{2}] Time {batch_time.val:.1f} ' \
'({batch_time.avg:.1f}, {data_time.val:.1f})\t' \
.format(epoch, i, len(train_loader), batch_time=batch_time, data_time=data_time)
msg += 'Loss {loss.val:.3f} ({loss.avg:.3f}){{'.format(loss=losses)
for wt1, loss1 in zip(wts, sub_losses):
msg += '%.3f*%.1f, ' % (loss1, wt1)
msg += '}}\t{crit} {accs.val:.3f} ({accs.avg:.3f})'.format(
crit=crit, accs=accs, ms=prob_np.max())
logger.info(msg)
def train_epoch(model,
train_dataloader,
optimizer,
cost,
epoch,
n_epochs,
print_freq=40,
batch_num=None,
random_layer=None):
print(random_layer)
batch_time = AverageMeter()
losses = AverageMeter()
error = AverageMeter()
end = time.time()
model.train()
for i, (images, labels, noises) in enumerate(train_dataloader):
#if i>30:
# return batch_time.avg, losses.avg, error.avg
#print(type(images),type(labels))
#print(images.shape,labels.shape)
images, labels, noises = images.cuda(), labels.cuda(), noises.cuda()
noises = noises * args.phi
if random_layer:
images = random_layer(images)
#print(images.shape)
#print(images.shape,labels.shape)
optimizer.zero_grad()
outputs = model(images, noises)
#print("outpits.size={},labels.size={}",outputs.size(),labels.size())
loss = cost(outputs, labels)
loss.backward()
#optimizer.step()
batch_size = labels.size(0)
outputs = outputs.max(1)[1]
error.update(
torch.ne(outputs.cpu(), labels.cpu()).float().sum().item() /
batch_size, batch_size)
losses.update(loss.item(), batch_size)
batch_time.update(time.time() - end)
end = time.time()
if i % print_freq == 0:
res = '\t'.join([
'Epoch: [%d/%d]' % (epoch + 1, n_epochs),
'Iter: [%d/%d]' % (i + 1, batch_num),
'Time %.3f (%.3f)' % (batch_time.val, batch_time.avg),
'Loss %.4f (%.4f)' % (losses.val, losses.avg),
'Error %.4f (%.4f)' % (error.val, error.avg),
])
print(res)
return batch_time.avg, losses.avg, error.avg
def test(val_loader, model, criterion, epoch, use_cuda):
global best_acc
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to evaluate mode
model.eval()
end = time.time()
bar = Bar('Processing', max=len(val_loader))
for batch_idx, (inputs, targets) in enumerate(val_loader):
# measure data loading time
data_time.update(time.time() - end)
if use_cuda:
inputs, targets = inputs.cuda(), targets.cuda()
inputs, targets = torch.autograd.Variable(inputs, volatile=True), torch.autograd.Variable(targets)
# compute output
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))
losses.update(loss.data[0], inputs.size(0))
top1.update(prec1[0], inputs.size(0))
top5.update(prec5[0], inputs.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# plot progress
bar.suffix = '({batch}/{size}) Data: {data:.3f}s | Batch: {bt:.3f}s | Total: {total:} | ETA: {eta:} | Loss: {loss:.4f} | top1: {top1: .4f} | top5: {top5: .4f}'.format(
batch=batch_idx + 1,
size=len(val_loader),
data=data_time.avg,
bt=batch_time.avg,
total=bar.elapsed_td,
eta=bar.eta_td,
loss=losses.avg,
top1=top1.avg,
top5=top5.avg,
)
bar.next()
bar.finish()
return (losses.avg, top1.avg)
def train(train_loader, model, criterion, optimizer, epoch, use_cuda):
# switch to train mode
model.train()
torch.set_grad_enabled(True)
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
end = time.time()
bar = Bar('Processing', max=len(train_loader))
show_step = len(train_loader) // 10
prefetcher = data_prefetcher(train_loader)
inputs, targets = prefetcher.next()
batch_idx = -1
while inputs is not None:
# for batch_idx, (inputs, targets) in enumerate(train_loader):
batch_idx += 1
batch_size = inputs.size(0)
if batch_size < args.train_batch:
break
# measure data loading time
data_time.update(time.time() - end)
#if use_cuda:
# inputs, targets = inputs.cuda(), targets.cuda(async=True)
#inputs, targets = torch.autograd.Variable(inputs), torch.autograd.Variable(targets)
# compute output
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))
losses.update(loss.data, inputs.size(0))
top1.update(prec1, inputs.size(0))
top5.update(prec5, inputs.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# plot progress
bar.suffix = '({batch}/{size}) Data: {data:.3f}s | Batch: {bt:.3f}s | Total: {total:} | ETA: {eta:} | Loss: {loss:.4f} | top1: {top1: .4f} | top5: {top5: .4f}'.format(
batch=batch_idx + 1,
size=len(train_loader),
data=data_time.val,
bt=batch_time.val,
total=bar.elapsed_td,
eta=bar.eta_td,
loss=losses.avg,
top1=top1.avg,
top5=top5.avg,
)
if (batch_idx) % show_step == 0:
print(bar.suffix)
bar.next()
inputs, targets = prefetcher.next()
bar.finish()
return (losses.avg, top1.avg)
def validate(val_loader, net, criterion, optimizer, epoch, train_args, visualize):
# the following code is written assuming that batch size is 1
net.eval()
val_loss = AverageMeter()
gts_all = np.zeros((len(val_loader), args['shorter_size'], 2 * args['shorter_size']), dtype=int)
predictions_all = np.zeros((len(val_loader), args['shorter_size'], 2 * args['shorter_size']), dtype=int)
for vi, data in enumerate(val_loader):
input, gt, slices_info = data
assert len(input.size()) == 5 and len(gt.size()) == 4 and len(slices_info.size()) == 3
input.transpose_(0, 1)
gt.transpose_(0, 1)
slices_info.squeeze_(0)
assert input.size()[3:] == gt.size()[2:]
count = torch.zeros(args['shorter_size'], 2 * args['shorter_size']).cuda()
output = torch.zeros(voc.num_classes, args['shorter_size'], 2 * args['shorter_size']).cuda()
slice_batch_pixel_size = input.size(1) * input.size(3) * input.size(4)
for input_slice, gt_slice, info in zip(input, gt, slices_info):
input_slice = Variable(input_slice).cuda()
gt_slice = Variable(gt_slice).cuda()
output_slice = net(input_slice)
assert output_slice.size()[2:] == gt_slice.size()[1:]
assert output_slice.size()[1] == voc.num_classes
output[:, info[0]: info[1], info[2]: info[3]] += output_slice[0, :, :info[4], :info[5]].data
gts_all[vi, info[0]: info[1], info[2]: info[3]] += gt_slice[0, :info[4], :info[5]].data.cpu().numpy()
count[info[0]: info[1], info[2]: info[3]] += 1
val_loss.update(criterion(output_slice, gt_slice).data[0], slice_batch_pixel_size)
output /= count
gts_all[vi, :, :] /= count.cpu().numpy().astype(int)
predictions_all[vi, :, :] = output.max(0)[1].squeeze_(0).cpu().numpy()
print('validating: %d / %d' % (vi + 1, len(val_loader)))
acc, acc_cls, mean_iu, fwavacc = evaluate(predictions_all, gts_all, voc.num_classes)
train_args['best_record']['val_loss'] = val_loss.avg
train_args['best_record']['epoch'] = epoch
train_args['best_record']['acc'] = acc
train_args['best_record']['acc_cls'] = acc_cls
train_args['best_record']['mean_iu'] = mean_iu
train_args['best_record']['fwavacc'] = fwavacc
snapshot_name = 'epoch_%d_loss_%.5f_acc_%.5f_acc-cls_%.5f_mean-iu_%.5f_fwavacc_%.5f_lr_%.10f' % (
epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc, optimizer.param_groups[1]['lr'])
torch.save(net.state_dict(), os.path.join(ckpt_path, exp_name, snapshot_name + '.pth'))
torch.save(optimizer.state_dict(), os.path.join(ckpt_path, exp_name, 'opt_' + snapshot_name + '.pth'))
if train_args['val_save_to_img_file']:
to_save_dir = os.path.join(ckpt_path, exp_name, str(epoch))
check_mkdir(to_save_dir)
val_visual = []
for idx, data in enumerate(zip(gts_all, predictions_all)):
gt_pil = voc.colorize_mask(data[0])
predictions_pil = voc.colorize_mask(data[1])
if train_args['val_save_to_img_file']:
predictions_pil.save(os.path.join(to_save_dir, '%d_prediction.png' % idx))
gt_pil.save(os.path.join(to_save_dir, '%d_gt.png' % idx))
val_visual.extend([visualize(gt_pil.convert('RGB')),
visualize(predictions_pil.convert('RGB'))])
val_visual = torch.stack(val_visual, 0)
val_visual = vutils.make_grid(val_visual, nrow=2, padding=5)
writer.add_image(snapshot_name, val_visual)
print('-----------------------------------------------------------------------------------------------------------')
print('[epoch %d], [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f]' % (
epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc))
print('best record: [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f], [epoch %d]' % (
train_args['best_record']['val_loss'], train_args['best_record']['acc'], train_args['best_record']['acc_cls'],
train_args['best_record']['mean_iu'], train_args['best_record']['fwavacc'], train_args['best_record']['epoch']))
print('-----------------------------------------------------------------------------------------------------------')
writer.add_scalar('val_loss', val_loss.avg, epoch)
writer.add_scalar('acc', acc, epoch)
writer.add_scalar('acc_cls', acc_cls, epoch)
writer.add_scalar('mean_iu', mean_iu, epoch)
writer.add_scalar('fwavacc', fwavacc, epoch)
net.train()
return val_loss.avg
def test(val_loader, model, criterion, epoch, use_cuda):
global best_acc
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
# switch to evaluate mode
model.eval()
# torch.set_grad_enabled(False)
end = time.time()
bar = Bar('Processing', max=len(val_loader))
prefetcher = data_prefetcher(val_loader)
inputs, targets = prefetcher.next()
batch_idx = -1
while inputs is not None:
# for batch_idx, (inputs, targets) in enumerate(val_loader):
batch_idx += 1
# measure data loading time
data_time.update(time.time() - end)
#if use_cuda:
# inputs, targets = inputs.cuda(), targets.cuda()
#inputs, targets = torch.autograd.Variable(inputs, volatile=True), torch.autograd.Variable(targets)
# compute output
with torch.no_grad():
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
prec1, prec5 = accuracy(outputs.data, targets.data, topk=(1, 5))
# losses.update(loss.data[0], inputs.size(0))
losses.update(loss.data, inputs.size(0))
#top1.update(prec1[0], inputs.size(0))
top1.update(prec1, inputs.size(0))
#top5.update(prec5[0], inputs.size(0))
top5.update(prec5, inputs.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# plot progress
bar.suffix = '({batch}/{size}) Data: {data:.3f}s | Batch: {bt:.3f}s | Total: {total:} | ETA: {eta:} | Loss: {loss:.4f} | top1: {top1: .4f} | top5: {top5: .4f}'.format(
batch=batch_idx + 1,
size=len(val_loader),
data=data_time.avg,
bt=batch_time.avg,
total=bar.elapsed_td,
eta=bar.eta_td,
loss=losses.avg,
top1=top1.avg,
top5=top5.avg,
)
bar.next()
inputs, targets = prefetcher.next()
print(bar.suffix)
bar.finish()
return (losses.avg, top1.avg)
