def infer(valid_queue, model, criterion):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.eval()
for step, (input, target) in enumerate(valid_queue):
input = input.cuda()
target = target.cuda(non_blocking=True)
with torch.no_grad():
logits, _ = model(input)
loss = criterion(logits, target)
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
objs.update(loss.data.item(), n)
top1.update(prec1.data.item(), n)
top5.update(prec5.data.item(), n)
if step % args.report_freq == 0:
end_time = time.time()
if step == 0:
duration = 0
start_time = time.time()
else:
duration = end_time - start_time
start_time = time.time()
logging.info('VALID Step: %03d Objs: %e R1: %f R5: %f Duration: %ds', step, objs.avg, top1.avg, top5.avg, duration)
return top1.avg, top5.avg, objs.avg
def train(train_queue, model, criterion, optimizer):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
model.train()
for step, (inp, target) in enumerate(train_queue):
global global_step
global_step += 1
inp = inp.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
model.zero_grad()
logits = model(inp)
loss = criterion(logits, target)
loss.backward()
if args.grad_clip > 0:
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
prec1, _ = utils.accuracy(logits, target, topk=(1, 5))
n = inp.size(0)
objs.update(loss.clone().item(), n)
top1.update(prec1.clone().item(), n)
writer_tr.add_scalar('loss', loss.item(), global_step)
writer_tr.add_scalar('acc', prec1.item(), global_step)
decomp_loss_1 = torch.mean(torch.log(torch.sum(torch.exp(logits - torch.max(logits, 1)[0].view(-1, 1)), 1))).item()
writer_tr.add_scalar('decomp_loss_lgsmex', decomp_loss_1, global_step)
decomp_loss_2 = torch.mean(torch.max(logits, 1)[0] - logits[range(len(inp)), target.cpu().numpy()]).item()
writer_tr.add_scalar('decomp_loss_max', decomp_loss_2, global_step)
if (step + 1) % args.report_freq == 0:
logging.info('Train Step: %03d Objs: %e Acc: %.2f', step + 1, objs.avg, top1.avg)
return top1.avg, objs.avg
def train(train_queue, model, criterion, optimizer):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
model.train()
for step, (inp, target) in enumerate(train_queue):
inp = inp.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
model.zero_grad()
logits = model(inp)
loss = criterion(logits, target)
loss.backward()
optimizer.step()
prec1, _ = utils.accuracy(logits, target, topk=(1, 5))
n = inp.size(0)
objs.update(loss.clone().item(), n)
top1.update(prec1.clone().item(), n)
step += 1
if step % args.report_freq == 0:
print('Train Step: {:3d} Objs: {:.4f} Acc: {:.2f}'.format(step, objs.avg, top1.avg))
return top1.avg, objs.avg
def infer(test_queue, model, criterion, depth_norm):
objs = utils.AverageMeter()
objs_ = utils.AverageMeter()
model.eval()
targets = []
predicteds = []
for step, (input, target) in enumerate(test_queue):
input = Variable(input.float()).cuda()
target = Variable(target.float()).cuda(non_blocking=True)
logits, _ = model(input)
loss = criterion(torch.squeeze(logits), target)
loss_ = criterion(torch.squeeze(logits) * 10, target * 10)
predicteds.extend(torch.squeeze(logits / depth_norm).cpu().tolist())
targets.extend((target / depth_norm).cpu().tolist())
n = input.size(0)
objs.update(loss.item(), n)
objs_.update(loss_.item(), n)
if step % args.report_freq == 0:
logging.info('test %03d %e %e', step, objs.avg, objs_.avg)
return objs.avg, targets, predicteds
def train(train_queue, model, criterion, optimizer):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.train()
for step, (x, target) in enumerate(train_queue):
x = x.cuda()
target = target.cuda(non_blocking=True)
optimizer.zero_grad()
logits, logits_aux = model(x)
loss = criterion(logits, target)
if args.auxiliary:
loss_aux = criterion(logits_aux, target)
loss += args.auxiliary_weight * loss_aux
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = x.size(0)
objs.update(loss.item(), n)
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
if step % args.report_freq == 0:
logging.info('train %03d %e %f %f', step, objs.avg, top1.avg,
top5.avg)
return top1.avg, objs.avg
def train(train_queue, model, criterion, optimizer):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
model.train()
for step, (input, target) in enumerate(train_queue):
input = input.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
model.zero_grad()
model.generate_share_alphas()
logits = model(input)
loss = criterion(logits, target)
loss.backward()
if args.grad_clip > 0:
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
prec1, _ = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
objs.update(loss.clone().item(), n)
top1.update(prec1.clone().item(), n)
if (step + 1) % args.report_freq == 0:
logging.info('Train Step: %03d Objs: %e Acc: %.2f', step + 1,
objs.avg, top1.avg)
return top1.avg, objs.avg
def infer(valid_queue, model, criterion):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.eval()
with torch.no_grad():
for step, (data, target) in enumerate(valid_queue):
data = data.cuda()
target = target.cuda()
logits, _ = model(data)
loss = criterion(logits, target)
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = data.size(0)
objs.update(loss.item(), n)
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
if step % args.report_freq == 0:
logging.info('valid %03d %e %f %f', step, objs.avg, top1.avg,
top5.avg)
return top1.avg, top5.avg, objs.avg
def infer(valid_queue, model, alpha, criterion):
"""Run model in eval only mode."""
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
# model.eval()
with torch.no_grad():
for step, (data, target) in enumerate(valid_queue):
n = data.size(0)
data = data.cuda()
target = target.cuda()
weights = alpha(data.size(0))
logits = model(data, weights)
loss = criterion(logits, target)
# Calculate the accuracy.
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
objs.update(loss.data, n)
top1.update(prec1.data, n)
top5.update(prec5.data, n)
if step % args.report_freq == 0 or step == len(valid_queue) - 1:
objs_avg = utils.reduce_tensor(objs.avg, args.world_size)
top1_avg = utils.reduce_tensor(top1.avg, args.world_size)
top5_avg = utils.reduce_tensor(top5.avg, args.world_size)
logging.info('valid %03d %e %f %f', step, objs_avg, top1_avg,
top5_avg)
return top1_avg, objs_avg
def infer(valid_queue, model, criterion):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
model.eval()
with torch.no_grad():
for step, (input, target) in enumerate(valid_queue):
input = Variable(input)
target = Variable(target)
if args.gpu != -1:
input = input.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
logits = model(input)
loss = criterion(logits, target)
prec1 = utils.accuracy(logits, target, topk=(1, ))
n = input.size(0)
objs.update(loss.detach().item(), n)
top1.update(prec1[0].item(), n)
if step % args.report_freq == 0:
logging.info('valid %03d %e %f', step, objs.avg, top1.avg)
return top1.avg, objs.avg
def validate(epoch, val_loader, model, criterion, args):
batch_time = utils.AverageMeter('Time', ':6.3f')
losses = utils.AverageMeter('Loss', ':.4e')
top1 = utils.AverageMeter('Acc@1', ':6.2f')
top5 = utils.AverageMeter('Acc@5', ':6.2f')
# switch to evaluation mode
model.eval()
with torch.no_grad():
end = time.time()
for i, (images, target) in enumerate(val_loader):
images = images.cuda()
target = target.cuda()
# compute output
logits = model(images)
loss = criterion(logits, target)
# measure accuracy and record loss
pred1, pred5 = utils.accuracy(logits, target, topk=(1, 5))
n = images.size(0)
losses.update(loss.item(), n)
top1.update(pred1[0], n)
top5.update(pred5[0], n)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
logger.info(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}'
.format(top1=top1, top5=top5))
return losses.avg, top1.avg, top5.avg
def eval(self, valid_queue):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
# top5 = utils.AverageMeter()
self.model.eval()
for step, (x, y) in enumerate(valid_queue):
x, y = x.to(self.device), y.to(self.device)
logits = self.model(x)
loss = self.criterion(logits, y)
prec1 = utils.accuracy(logits, y, topk=1)
n = x.size(0)
objs.update(loss.item(), n)
top1.update(prec1.item(), n)
# top5.update(prec5.item(), n)
if step % 100 == 0:
logging.info('valid %03d %e %f %f', step, objs.avg, top1.avg)
print('valid: {} loss: {} acc: {}'.format(
step, objs.avg, top1.avg))
return top1.avg, objs.avg
def train(epoch, train_loader, model, criterion, optimizer, scheduler):
batch_time = utils.AverageMeter('Time', ':6.3f')
data_time = utils.AverageMeter('Data', ':6.3f')
losses = utils.AverageMeter('Loss', ':.4e')
top1 = utils.AverageMeter('Acc@1', ':6.2f')
top5 = utils.AverageMeter('Acc@5', ':6.2f')
'''progress = utils.ProgressMeter(
len(train_loader),
[batch_time, data_time, losses, top1, top5],
prefix="Epoch: [{}]".format(epoch))#'''
model.train()
end = time.time()
scheduler.step()
for param_group in optimizer.param_groups:
cur_lr = param_group['lr']
logger.info('learning_rate: ' + str(cur_lr))
num_iter = len(train_loader)
for i, (images, target) in enumerate(train_loader):
data_time.update(time.time() - end)
images = images.cuda()
target = target.cuda()
# compute outputy
logits = model(images)
loss = criterion(logits, target)
# measure accuracy and record loss
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = images.size(0)
losses.update(loss.item(), n) #accumulated loss
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
# 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:
logger.info('Epoch[{0}]({1}/{2}): '
'Loss {loss.avg:.4f} '
'Prec@1(1,5) {top1.avg:.2f}, {top5.avg:.2f}'.format(
epoch,
i,
num_iter,
loss=losses,
top1=top1,
top5=top5))
#progress.display(i)
return losses.avg, top1.avg, top5.avg
def train(train_queue, model, criterion, optimizer, epoch, init_lr,
warmup_epochs, global_step):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.train()
for step, (data, target) in enumerate(train_queue):
n = data.size(0)
data = data.cuda()
target = target.cuda()
# Change lr.
if epoch < warmup_epochs:
len_epoch = len(train_queue)
scale = float(1 + step + epoch * len_epoch) / \
(warmup_epochs * len_epoch)
lr = init_lr * scale
for param_group in optimizer.param_groups:
param_group['lr'] = lr
# Forward.
optimizer.zero_grad()
logits, logits_aux = model(data)
loss = criterion(logits, target)
if args.auxiliary:
loss_aux = criterion(logits_aux, target)
loss += args.auxiliary_weight * loss_aux
# Backward and step.
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
# Calculate the accuracy.
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
reduced_loss = utils.reduce_tensor(loss.data, args.world_size)
prec1 = utils.reduce_tensor(prec1, args.world_size)
prec5 = utils.reduce_tensor(prec5, args.world_size)
objs.update(to_python_float(reduced_loss), n)
top1.update(to_python_float(prec1), n)
top5.update(to_python_float(prec5), n)
if step % args.report_freq == 0:
current_lr = list(optimizer.param_groups)[0]['lr']
logging.info('train %03d %e %f %f lr: %e', step, objs.avg,
top1.avg, top5.avg, current_lr)
writer.add_scalar('train/loss', objs.avg, global_step)
writer.add_scalar('train/acc_top1', top1.avg, global_step)
writer.add_scalar('train/acc_top5', top5.avg, global_step)
writer.add_scalar('train/lr',
optimizer.state_dict()['param_groups'][0]['lr'],
global_step)
global_step += 1
return top1.avg, objs.avg, global_step
def validate(epoch, val_loader, model, criterion, args):
batch_time = utils.AverageMeter('Time', ':6.3f')
losses = utils.AverageMeter('Loss', ':.4e')
top1 = utils.AverageMeter('Acc@1', ':6.2f')
top5 = utils.AverageMeter('Acc@5', ':6.2f')
if args.use_dali:
num_iter = val_loader._size // args.batch_size
else:
num_iter = len(val_loader)
model.eval()
with torch.no_grad():
end = time.time()
if args.use_dali:
for batch_idx, batch_data in enumerate(val_loader):
images = batch_data[0]['data'].cuda()
targets = batch_data[0]['label'].squeeze().long().cuda()
# compute output
logits = model(images)
loss = criterion(logits, targets)
# measure accuracy and record loss
pred1, pred5 = utils.accuracy(logits, targets, topk=(1, 5))
n = images.size(0)
losses.update(loss.item(), n)
top1.update(pred1[0], n)
top5.update(pred5[0], n)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
else:
for batch_idx, (images, targets) in enumerate(val_loader):
images = images.cuda()
targets = targets.cuda()
# compute output
logits = model(images)
loss = criterion(logits, targets)
# measure accuracy and record loss
pred1, pred5 = utils.accuracy(logits, targets, topk=(1, 5))
n = images.size(0)
losses.update(loss.item(), n)
top1.update(pred1[0], n)
top5.update(pred5[0], n)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
logger.info(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}'.format(
top1=top1, top5=top5))
return losses.avg, top1.avg, top5.avg
def train(train_queue, valid_iter, model, architect, criterion, optimizer, lr,
loggers):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
batches = len(train_queue)
for step, (input, target) in enumerate(train_queue):
model.train()
model.tick(1 / batches)
n = input.size(0)
input = Variable(input, requires_grad=False).cuda(non_blocking=True)
target = Variable(target, requires_grad=False).cuda(non_blocking=True)
# get a random minibatch from the search queue without replacement
input_search, target_search = next(valid_iter)
input_search = Variable(input_search,
requires_grad=False).cuda(non_blocking=True)
target_search = Variable(target_search,
requires_grad=False).cuda(non_blocking=True)
valid_loss = architect.step(input,
target,
input_search,
target_search,
lr,
optimizer,
unrolled=args.unrolled)
utils.log_loss(loggers["val"], valid_loss, None, model.clock)
optimizer.zero_grad()
logits = model(input)
loss = criterion(logits, target)
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
model.mask_alphas()
model.track_FI()
model.update_history()
prec1 = utils.accuracy(logits, target, topk=(1, ))
objs.update(loss.item(), n)
top1.update(prec1[0].item(), n)
utils.log_loss(loggers["train"], loss.item(), prec1[0].item(),
model.clock)
if step % args.report_freq == 0:
logging.info('train %03d %e %f', step, objs.avg, top1.avg)
if (step + 1) % args.admm_freq == 0:
model.update_Z()
model.update_U()
return top1.avg, objs.avg
def train(train_queue, valid_queue, model, architect, criterion, optimizer,
loggers):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
valid_iter = iter(valid_queue)
batches = len(train_queue)
for step, (input, target) in enumerate(train_queue):
model.train()
n = input.size(0)
model.tick(1 / batches)
input = Variable(input, requires_grad=False).cuda(non_blocking=True)
target = Variable(target, requires_grad=False).cuda(non_blocking=True)
# get a random minibatch from the search queue with replacement
input_search, target_search = next(valid_iter)
input_search = Variable(input_search,
requires_grad=False).cuda(non_blocking=True)
target_search = Variable(target_search,
requires_grad=False).cuda(non_blocking=True)
architect.step(input_search, target_search)
optimizer.zero_grad()
logits = model(input)
loss = criterion(logits, target)
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
model.FI_hist.append(
torch.norm(
torch.stack([
torch.norm(p.grad.detach(), 2.0).cuda()
for p in model.parameters() if p.grad is not None
]), 2.0)**2)
if len(model.batchstep) > 0:
model.batchstep.append(model.batchstep[-1] + 1 / batches)
else:
model.batchstep.append(0.0)
optimizer.step()
prec1 = utils.accuracy(logits, target, topk=(1, ))
objs.update(loss.item(), n)
top1.update(prec1[0].item(), n)
utils.log_loss(loggers["train"], loss.item(), prec1[0].item(),
model.clock)
model.update_history()
if step % args.report_freq == 0:
logging.info('train %03d %e %f', step, objs.avg, top1.avg)
return top1.avg, objs.avg
def train(epoch, train_loader, model, criterion, optimizer):
batch_time = utils.AverageMeter('Time', ':6.3f')
data_time = utils.AverageMeter('Data', ':6.3f')
losses = utils.AverageMeter('Loss', ':.4e')
top1 = utils.AverageMeter('Acc@1', ':6.2f')
top5 = utils.AverageMeter('Acc@5', ':6.2f')
model.train()
end = time.time()
#scheduler.step()
num_iter = train_loader._size // args.batch_size
for batch_idx, batch_data in enumerate(train_loader):
images = batch_data[0]['data'].cuda()
targets = batch_data[0]['label'].squeeze().long().cuda()
data_time.update(time.time() - end)
adjust_learning_rate(optimizer, epoch, batch_idx, num_iter)
# compute output
logits = model(images)
loss = criterion(logits, targets)
# measure accuracy and record loss
prec1, prec5 = utils.accuracy(logits, targets, topk=(1, 5))
n = images.size(0)
losses.update(loss.item(), n) #accumulated loss
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
# 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 batch_idx % print_freq == 0:
logger.info('Epoch[{0}]({1}/{2}): '
'Loss {loss.avg:.4f} '
'Prec@1(1,5) {top1.avg:.2f}, {top5.avg:.2f}'.format(
epoch,
batch_idx,
num_iter,
loss=losses,
top1=top1,
top5=top5))
return losses.avg, top1.avg, top5.avg
def train(train_queue, valid_queue, model, architect, criterion, optimizer, lr, loggers):
objs = utils.AverageMeter()
valid_iter = iter(valid_queue)
batches = len(train_queue)
for step, (input, target) in enumerate(train_queue):
model.train()
n = input.size(0)
model.tick(1 / batches)
input = Variable(input.float(), requires_grad=False).cuda(non_blocking=True)
target = Variable(target.float(), requires_grad=False).cuda(non_blocking=True)
input_search, target_search = next(valid_iter)
input_search = Variable(input_search.float(), requires_grad=False).cuda(non_blocking=True)
target_search = Variable(target_search.float(), requires_grad=False).cuda(non_blocking=True)
valid_loss = architect.step(input, target, input_search, target_search, lr, optimizer, unrolled=args.unrolled)
utils.log_loss(loggers["val"], valid_loss.item(), None, model.clock)
optimizer.zero_grad()
logits = model(input)
loss = criterion(logits, target)
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
objs.update(loss.item(), n)
utils.log_loss(loggers["train"], loss.item(), None, model.clock)
if step % args.report_freq == 0:
logging.info('train %03d %e', step, objs.avg)
return objs.avg
def train(train_queue, model, criterion, optimizer, train_logger):
objs = utils.AverageMeter()
model.train()
batches = len(train_queue)
for step, (input, target) in enumerate(train_queue):
input = Variable(input.float()).cuda()
target = Variable(target.float()).cuda(non_blocking=True)
optimizer.zero_grad()
logits, logits_aux = model(input)
loss = criterion(torch.squeeze(logits), target)
if args.auxiliary:
loss_aux = criterion(logits_aux, target)
loss += args.auxiliary_weight * loss_aux
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
n = input.size(0)
objs.update(loss.item(), n)
utils.log_loss(train_logger, loss.item(), None, 1 / batches)
if step % args.report_freq == 0:
logging.info('train %03d %e', step, objs.avg)
return objs.avg
def infer(valid_queue, model, criterion):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.eval()
if TORCH_VERSION.startswith('1'):
with torch.no_grad():
for step, (input, target) in enumerate(valid_queue):
input = input.to(device)
target = target.to(device)
logits, _ = model(input)
loss = criterion(logits, target)
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
objs.update(loss.item(), n)
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
if step % args.report_freq == 0:
logging.info('valid %03d %e %f %f', step, objs.avg,
top1.avg, top5.avg)
if args.debug:
break
else:
for step, (input, target) in enumerate(valid_queue):
input = Variable(input, volatile=True).cuda()
target = Variable(target, volatile=True).cuda(async=True)
logits, _ = model(input)
loss = criterion(logits, target)
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
objs.update(loss.data[0], n)
top1.update(prec1.data[0], n)
top5.update(prec5.data[0], n)
if step % args.report_freq == 0:
logging.info('valid %03d %e %f %f', step, objs.avg, top1.avg,
top5.avg)
if args.debug:
break
return top1.avg, objs.avg
def train(train_queue, model, criterion, optimizer):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.train()
for step, (input, target) in enumerate(train_queue):
if TORCH_VERSION in ['1.0.1', '1.1.0']:
input = input.to(device)
target = target.to(device)
else:
input = Variable(input).cuda()
target = Variable(target).cuda(async=True)
optimizer.zero_grad()
logits, logits_aux = model(input)
loss = criterion(logits, target)
if args.auxiliary:
loss_aux = criterion(logits_aux, target)
loss += args.auxiliary_weight * loss_aux
loss.backward()
if TORCH_VERSION.startswith('1'):
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
else:
nn.utils.clip_grad_norm(model.parameters(), args.grad_clip)
optimizer.step()
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
if TORCH_VERSION.startswith('1'):
objs.update(loss.item(), n)
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
else:
objs.update(loss.data[0], n)
top1.update(prec1.data[0], n)
top5.update(prec5.data[0], n)
if step % args.report_freq == 0:
logging.info('train %03d %e %f %f', step, objs.avg, top1.avg,
top5.avg)
if args.debug:
break
return top1.avg, objs.avg
def infer(valid_queue, model, criterion):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
model.eval()
for step, (inp, target) in enumerate(valid_queue):
inp = inp.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
with torch.no_grad():
logits = model(inp)
loss = criterion(logits, target)
prec1, _ = utils.accuracy(logits, target, topk=(1, 5))
n = inp.size(0)
objs.update(loss.clone().item(), n)
top1.update(prec1.clone().item(), n)
return top1.avg, objs.avg
def train(self, train_queue, valid_queue, lr):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
# top5 = utils.AverageMeter()
for step, (x, y) in enumerate(train_queue):
self.model.train()
n = x.size(0)
x, y = x.to(self.device), y.to(self.device)
# get a random mini batch from the search queue with replacement
x_s, y_s = next(iter(valid_queue))
x_s, y_s = x_s.to(self.device), y_s.to(self.device)
self.architect.step(x,
y,
x_s,
y_s,
lr,
self.optimizer,
unrolled=self.unrolled)
self.optimizer.zero_grad()
logits = self.model(x)
loss = self.criterion(logits, y)
loss.backward()
nn.utils.clip_grad_norm_(self.model.parameters(), self.grad_clip)
self.optimizer.step()
prec1 = utils.accuracy(logits, y, topk=1)
objs.update(loss.item(), n)
top1.update(prec1.item(), n)
# top5.update(prec5.item(), n)
if step % 100 == 0:
logging.info('train %03d %e %f %f', step, objs.avg, top1.avg)
print('train: {} loss: {} acc: {}'.format(
step, objs.avg, top1.avg))
return top1.avg, objs.avg
def train_init(train_queue, model, alpha, criterion, optimizer, weight_params):
"""Update network weights on train set and architecture on val set."""
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
for step, (data, target) in enumerate(train_queue):
model.train()
n = data.size(0)
# Update network weights using the train set.
data = data.cuda()
target = target.cuda()
weights = alpha(data.size(0))
weights_no_grad = alpha.module.clone_weights(weights)
optimizer.zero_grad()
logits = model(data, weights_no_grad)
loss = criterion(logits, target)
dummy = sum([torch.sum(param) for param in model.parameters()])
loss += dummy * 0.
loss.backward()
nn.utils.clip_grad_norm_(weight_params, args.grad_clip)
optimizer.step()
# Calculate the accuracy.
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
objs.update(loss.data, n)
top1.update(prec1.data, n)
top5.update(prec5.data, n)
if (step + 1) % args.report_freq == 0 or step == len(train_queue) - 1:
objs_avg = utils.reduce_tensor(objs.avg, args.world_size)
top1_avg = utils.reduce_tensor(top1.avg, args.world_size)
top5_avg = utils.reduce_tensor(top5.avg, args.world_size)
logging.info('train_init %03d %e %f %f', step, objs_avg, top1_avg,
top5_avg)
return top1_avg, objs_avg
def infer(test_queue, model, criterion):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
model.eval()
for step, (input, target) in enumerate(test_queue):
input = Variable(input).cuda()
target = Variable(target).cuda(non_blocking=True)
logits, _ = model(input)
loss = criterion(logits, target)
prec1 = utils.accuracy(logits, target, topk=(1,))
n = input.size(0)
objs.update(loss.item(), n)
top1.update(prec1[0].item(), n)
if step % args.report_freq == 0:
logging.info('test %03d %e %f', step, objs.avg, top1.avg)
return top1.avg, objs.avg
def train(train_queue, model, criterion, optimizer):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
batch_time = utils.AverageMeter()
model.train()
for step, (input, target) in enumerate(train_queue):
target = target.cuda(non_blocking=True)
input = input.cuda(non_blocking=True)
b_start = time.time()
optimizer.zero_grad()
logits, logits_aux = model(input)
loss = criterion(logits, target)
if args.auxiliary:
loss_aux = criterion(logits_aux, target)
loss += args.auxiliary_weight*loss_aux
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
batch_time.update(time.time() - b_start)
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
objs.update(loss.data.item(), n)
top1.update(prec1.data.item(), n)
top5.update(prec5.data.item(), n)
if step % args.report_freq == 0:
end_time = time.time()
if step == 0:
duration = 0
start_time = time.time()
else:
duration = end_time - start_time
start_time = time.time()
logging.info('TRAIN Step: %03d Objs: %e R1: %f R5: %f Duration: %ds BTime: %.3fs',
step, objs.avg, top1.avg, top5.avg, duration, batch_time.avg)
return top1.avg, objs.avg
def infer(valid_queue, model, criterion):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.eval()
def valid_generator():
for x, t in valid_queue:
yield x, t
valid_gen = valid_generator()
# for step, (input, target) in enumerate(valid_queue):
step = 0
for input, target in valid_gen:
step += 1
input = Variable(input)
target = Variable(target)
if not args.disable_cuda:
input = input.cuda()
target = target.cuda(async=True)
logits = model(input, 0)
loss = criterion(logits, target)
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
objs.update(loss.item(), n)
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
if step % args.report_freq == 0:
logging.info('valid %03d %e %f %f', step, objs.avg, top1.avg,
top5.avg)
if args.debug:
break
return top1.avg, objs.avg
def train(train_queue, model, criterion, optimizer, global_step):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.train()
for step, (data, target) in enumerate(train_queue):
n = data.size(0)
data = data.cuda()
target = target.cuda()
# Forward.
optimizer.zero_grad()
logits = model(data)
loss = criterion(logits, target)
# Backward and step.
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
# Calculate the accuracy.
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
reduced_loss = utils.reduce_tensor(loss.data, args.world_size)
prec1 = utils.reduce_tensor(prec1, args.world_size)
prec5 = utils.reduce_tensor(prec5, args.world_size)
objs.update(to_python_float(reduced_loss), n)
top1.update(to_python_float(prec1), n)
top5.update(to_python_float(prec5), n)
if (step + 1) % args.report_freq == 0:
current_lr = list(optimizer.param_groups)[0]['lr']
logging.info('train %03d %e %f %f lr: %e', step, objs.avg,
top1.avg, top5.avg, current_lr)
global_step += 1
return top1.avg, top5.avg, objs.avg, global_step
def infer(valid_queue, model, criterion):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
top5 = utils.AverageMeter()
model.eval()
for step, (input, target) in enumerate(valid_queue):
input = Variable(input, volatile=True).cuda()
target = Variable(target, volatile=True).cuda(non_blocking=True)
logits, _ = model(input)
loss = criterion(logits, target)
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
objs.update(loss.data[0], n)
top1.update(prec1.data[0], n)
top5.update(prec5.data[0], n)
if step % args.report_freq == 0:
logging.info('valid %03d %e %f %f', step, objs.avg, top1.avg, top5.avg)
return top1.avg, top5.avg, objs.avg
def infer(valid_queue, model, criterion, use_fly_bn=True):
objs = utils.AverageMeter()
top1 = utils.AverageMeter()
if not use_fly_bn:
model.eval()
else:
model.train()
for step, (input, target) in enumerate(valid_queue):
input = input.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
with torch.no_grad():
logits = model(input)
loss = criterion(logits, target)
prec1, _ = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
objs.update(loss.clone().item(), n)
top1.update(prec1.clone().item(), n)
logging.info('Valid Stats --- Objs: %e Acc: %f', objs.avg, top1.avg)
return top1.avg, objs.avg
