def train(train_queue, model, criterion, optimizer):
from taowei.torch2.utils.classif import ProgressMeter
progress = ProgressMeter(iters_per_epoch=len(train_queue),
epoch=args.epoch,
epochs=args.epochs,
split='train',
writer=args.writer,
batch_size=args.batch_size)
# args.epoch = epoch # keep a record of current epoch for evaluate. TODO: a more elegant way
# objs = utils.AvgrageMeter()
# top1 = utils.AvgrageMeter()
# top5 = utils.AvgrageMeter()
model.train()
timer = Timer()
timer.tic()
for step, (input, target) in enumerate(train_queue):
# measure data loading time
progress.update('data_time', timer.toc(from_last_toc=True))
target = target.cuda(non_blocking=True)
input = input.cuda()
input = Variable(input)
target = Variable(target)
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
progress.update('forward_time', timer.toc(from_last_toc=True))
loss.backward()
progress.update('backward_time', timer.toc(from_last_toc=True))
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
optimizer.step()
progress.update('update_time', timer.toc(from_last_toc=True))
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
progress.update('loss', loss.item(), n)
progress.update('top1', prec1.item(), n)
progress.update('top5', prec5.item(), n)
# objs.update(loss.item(), n)
# top1.update(prec1.item(), n)
# top5.update(prec5.item(), n)
# measure elapsed time
progress.update('batch_time', timer.toctic())
if step % args.report_freq == 0:
progress.log_iter_stats(iter=step,
batch_size=n,
lr=optimizer.param_groups[0]['lr'])
progress.log_epoch_stats(lr=optimizer.param_groups[0]['lr'])
return progress.stats['top1'].avg, progress.stats['loss'].avg
def infer(valid_queue, model, criterion):
from taowei.torch2.utils.classif import ProgressMeter
# epoch = args.start_epoch - 1 if 'epoch' not in args else args.epoch
progress = ProgressMeter(iters_per_epoch=len(valid_queue),
epoch=args.epoch, split='val', writer=args.writer, batch_size=args.batch_size)
model.eval()
timer = Timer()
timer.tic()
for step, (input, target) in enumerate(valid_queue):
# measure data loading time
progress.update('data_time', timer.toc(from_last_toc=True))
input = input.cuda()
target = target.cuda(non_blocking=True)
with torch.no_grad():
logits, _ = model(input)
loss = criterion(logits, target)
progress.update('forward_time', timer.toc(from_last_toc=True))
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
n = input.size(0)
progress.update('loss', loss.item(), n)
progress.update('top1', prec1.item(), n)
progress.update('top5', prec5.item(), n)
progress.update('batch_time', timer.toctic())
progress.update('batch_time', timer.toctic())
if step % args.report_freq == 0:
progress.log_iter_stats(iter=step, batch_size=n)
progress.log_epoch_stats()
return progress.stats['top1'].avg, progress.stats['top5'].avg, progress.stats['loss'].avg
def train(train_queue, valid_queue, model, architect, criterion, optimizer,
lr):
from taowei.torch2.utils.classif import ProgressMeter
progress = ProgressMeter(iters_per_epoch=len(train_queue),
epoch=args.epoch,
epochs=args.epochs,
split='train',
writer=args.writer,
batch_size=args.batch_size)
# objs = utils.AvgrageMeter()
# top1 = utils.AvgrageMeter()
# top5 = utils.AvgrageMeter()
timer = Timer()
timer.tic()
for step, (input, target) in enumerate(train_queue):
# measure data loading time
progress.update('data_time', timer.toc(from_last_toc=True))
model.train()
n = input.size(0)
input = Variable(input, requires_grad=False).cuda()
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(iter(valid_queue))
input_search = Variable(input_search, requires_grad=False).cuda()
target_search = Variable(target_search,
requires_grad=False).cuda(non_blocking=True)
architect.step(input,
target,
input_search,
target_search,
lr,
optimizer,
unrolled=args.unrolled)
progress.update('arch_step_time', timer.toc(from_last_toc=True))
optimizer.zero_grad()
logits = model(input)
loss = criterion(logits, target)
progress.update('forward_time', timer.toc(from_last_toc=True))
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
progress.update('backward_time', timer.toc(from_last_toc=True))
optimizer.step()
progress.update('update_time', timer.toc(from_last_toc=True))
prec1, prec5 = utils.accuracy(logits, target, topk=(1, 5))
progress.update('loss', loss.item(), n)
progress.update('top1', prec1.item(), n)
progress.update('top5', prec5.item(), n)
# objs.update(loss.item(), n)
# top1.update(prec1.item(), n)
# top5.update(prec5.item(), n)
# measure elapsed time
progress.update('batch_time', timer.toctic())
if step % args.report_freq == 0:
progress.log_iter_stats(iter=step,
batch_size=n,
lr=optimizer.param_groups[0]['lr'])
progress.log_epoch_stats(lr=optimizer.param_groups[0]['lr'])
return progress.stats['top1'].avg, progress.stats['loss'].avg