def eval_net(net,args):
top1 = AverageMeter()
top5 = AverageMeter()
losses = AverageMeter()
if args.validating:
print('Validating at epoch {}'.format(args.epoch + 1))
if args.testing:
print('Testing at epoch {}'.format(args.epoch + 1))
if not args.__contains__('validating'):
args.validating = False
if not args.__contains__('testing'):
args.testing = False
net.eval()
total = 1e-3
total_time = 0
end_time = time.time()
for batch_idx, (inputs, targets) in enumerate(args.data_loader):
with torch.no_grad():
if args.use_gpu:
targets = targets.cuda()
outputs = net(inputs)
if type(outputs) is list:
outputs = outputs[-1]
if args.loss == 'CE':
loss = args.criterion(outputs, targets) # .mean()
elif args.loss == 'L2':
from util import targets_to_one_hot
targets_one_hot = targets_to_one_hot(targets, args.num_outputs)
loss = args.criterion(outputs, targets_one_hot)*args.num_outputs*0.5
losses.update(loss.item(), inputs.size(0))
prec1, prec5 = accuracy(outputs, targets, topk=(1, 5))
top1.update(prec1[0].item(), inputs.size(0))
top5.update(prec5[0].item(), inputs.size(0))
total_time += (time.time() - end_time)
end_time = time.time()
if args.msg:
print('Loss: %.3f | top1: %.3f%% ,top5: %.3f%%'
% (losses.avg, top1.avg, top5.avg))
if args.testing:
args.test_losses.append(losses.avg)
args.test_accuracies.append(top1.avg)
args.test_epoch_logger.log({
'epoch': (args.epoch + 1),
'loss': losses.avg,
'top1': top1.avg,
'top5': top5.avg,
'time': total_time,
})
if args.validating:
args.valid_losses.append(losses.avg)
args.valid_accuracies.append(top1.avg)
args.valid_epoch_logger.log({
'epoch': (args.epoch + 1),
'loss': losses.avg,
'top1': top1.avg,
'top5': top5.avg,
'time': total_time,
})
# Save checkpoint.
is_best=(top1.avg > args.best_acc)
if is_best:
args.best_acc = top1.avg
states = {
'state_dict': net.module.state_dict() if hasattr(net,'module') else net.state_dict(),
'epoch': args.epoch+1,
'arch': args.arch,
'best_acc': args.best_acc,
'train_losses': args.train_losses,
'optimizer': args.current_optimizer.state_dict()
}
if args.__contains__('acc'):
states['acc']=top1.avg,
if args.__contains__('valid_losses'):
states['valid_losses']=args.valid_losses
if args.__contains__('test_losses'):
states['test_losses'] = args.test_losses
if (args.checkpoint_epoch > 0):
if not os.path.isdir(args.checkpoint_path):
os.mkdir(args.checkpoint_path)
save_file_path = os.path.join(args.checkpoint_path, 'checkpoint.pth.tar')
torch.save(states, save_file_path)
if is_best:
shutil.copyfile(save_file_path, os.path.join(args.checkpoint_path,'model_best.pth.tar'))
print('Loss: %.3f | top1: %.3f%%, top5: %.3f%%, elasped time: %3.f seconds. Best Acc: %.3f%%'
% (losses.avg , top1.avg, top5.avg, total_time, args.best_acc))
def train_net(net,args):
top1 = AverageMeter()
top5 = AverageMeter()
losses = AverageMeter()
print('training at epoch {}'.format(args.epoch+1))
net.train()
total_time=0
data_time=0
total=1e-3
correct=0
extra=0.
optimizer=args.current_optimizer
end_time = time.time()
for batch_idx, (inputs, targets) in enumerate(args.data_loader):
#ff
if args.use_gpu:
targets = targets.cuda()
data_time += (time.time() - end_time)#loading time
outputs = net(inputs)
maps=None
if type(outputs) is list:
maps=outputs
outputs=outputs[-1]
#loss = args.criterion(outputs, targets).mean()
if args.loss=='CE':
loss = args.criterion(outputs, targets)#.mean()
elif args.loss=='L2':
from util import targets_to_one_hot
targets_one_hot=targets_to_one_hot(targets,args.num_outputs)
loss = args.criterion(outputs, targets_one_hot)*args.num_outputs*0.5
losses.update(loss.item(), inputs.size(0))
prec1, prec5 = accuracy(outputs.data, targets, topk=(1, 5))
top1.update(prec1[0].item(), inputs.size(0))
top5.update(prec5[0].item(), inputs.size(0))
#bp
loss.backward()
optimizer.step()
if args.lr_scheduler == 'cosine':
args.current_scheduler.step()
if args.tensorboard:
if args.logger_n_iter%args.print_freq==0:
args.writer.add_scalar('loss', loss.item(), args.logger_n_iter )
args.logger_n_iter += 1
optimizer.zero_grad() # flush
total_time += (time.time() - end_time)
end_time = time.time()
if args.msg:
print('Loss: %.3f | top1: %.3f%% ,top5: %.3f%%'
% (losses.avg, top1.avg, top5.avg))
args.train_batch_logger.log({
'epoch': (args.epoch+1),
'batch': batch_idx + 1,
'loss': losses.avg,
'top1': top1.avg,
'top5': top5.avg,
'time': total_time,
})
args.train_epoch_logger.log({
'epoch': (args.epoch+1),
'loss': losses.avg,
'top1': top1.avg,
'top5':top5.avg,
'time': total_time,
})
print('Loss: %.3f | top1: %.3f%%, top5: %.3f%% elasped time: %3.f seconds.'
% (losses.avg, top1.avg, top5.avg, total_time))
args.train_accuracies.append(top1.avg)
args.train_losses.append(losses.avg)
def train_net(net, opts):
top1 = AverageMeter()
top5 = AverageMeter()
losses = AverageMeter()
print('training at epoch {}'.format(opts.epoch + 1))
net.train()
total_time = 0
data_time = 0
total = 1e-3
correct = 0
extra = 0.
optimizer = opts.current_optimizer
end_time = time.time_ns()
for batch_idx, (inputs, targets) in enumerate(opts.data_loader):
#ff
if opts.use_gpu:
targets = targets.cuda()
data_time += (time.time() - end_time) #loading time
outputs = net(inputs)
maps = None
if type(outputs) is list:
maps = outputs
outputs = outputs[-1]
#loss = opts.criterion(outputs, targets).mean()
if opts.loss == 'CE':
loss = opts.criterion(outputs, targets) #.mean()
elif opts.loss == 'L2':
from util import targets_to_one_hot
targets_one_hot = targets_to_one_hot(targets, opts.num_outputs)
loss = opts.criterion(outputs, targets_one_hot) #.mean()
losses.update(loss.item(), inputs.size(0))
prec1, prec5 = accuracy(outputs.data, targets, topk=(1, 5))
top1.update(prec1[0].item(), inputs.size(0))
top5.update(prec5[0].item(), inputs.size(0))
#bp
loss.backward()
if opts.lr_scheduler == 'cosine':
opts.current_scheduler.step()
optimizer.step()
if opts.tensorboard and (maps is not None):
if batch_idx % opts.viz_T == 0:
# Log images (image summary)
for l in range(len(maps) - 1):
c, h, w = maps[l][0].shape
tmp = maps[l][0].detach().permute([1, 0,
2]).contiguous().view(
h, w * c).cpu()
opts.writer.image_summary(
'train maps {}'.format(l),
[tensor2array(tmp, max_value=None, colormap='bone')],
opts.logger_n_iter)
# Log values and gradients of the parameters (histogram summary)
for tag, value in net.named_parameters():
#print(tag)
tag = tag.replace('.', '/')
opts.writer.histo_summary(tag,
value.data.cpu().numpy(),
opts.logger_n_iter)
if hasattr(value.grad, 'data'):
opts.writer.histo_summary(
tag + '/grad',
value.grad.data.cpu().numpy(), opts.logger_n_iter)
if opts.tensorboardX:
if opts.logger_n_iter % opts.viz_T == 0:
opts.writer.add_scalar('loss', loss.item(), opts.logger_n_iter)
opts.logger_n_iter += 1
optimizer.zero_grad() # flush
total_time += (time.time_ns() - end_time) / (10**9)
end_time = time.time_ns()
if opts.msg:
print('Loss: %.3f | top1: %.3f%% ,top5: %.3f%%' %
(losses.avg, top1.avg, top5.avg))
opts.train_batch_logger.log({
'epoch': (opts.epoch + 1),
'batch': batch_idx + 1,
'loss': losses.avg,
'top1': top1.avg,
'top5': top5.avg,
'time': total_time,
})
opts.train_epoch_logger.log({
'epoch': (opts.epoch + 1),
'loss': losses.avg,
'top1': top1.avg,
'top5': top5.avg,
'time': total_time,
})
print(
'Loss: %.3f | top1: %.3f%%, top5: %.3f%% elasped time: %3.f seconds.' %
(losses.avg, top1.avg, top5.avg, total_time))
opts.train_accuracies.append(top1.avg)
opts.train_losses.append(losses.avg)