def validate(val_loader, model, criterion):
""" Evaluate model using criterion on validation set """
losses = Meter(ptag='Loss')
acc = Meter(ptag='Accuracy')
# top1 = Meter(ptag='Prec@1')
# top5 = Meter(ptag='Prec@5')
# switch to evaluate mode
model.eval()
with torch.no_grad():
for i, (features, target) in enumerate(val_loader):
target = target.cuda(non_blocking=True)
# create one-hot vector from target
kl_target = torch.zeros(target.shape[0], 1000,
device='cuda').scatter_(
1, target.view(-1, 1), 1)
# compute output
output = model(features)
loss = criterion(output, kl_target)
# measure accuracy and record loss
# prec1, prec5 = accuracy(output, target, topk=())
acc_val = accuracy(output, target)
losses.update(loss.item(), features.size(0))
acc.update(acc_val, features.size(0))
# top1.update(prec1.item(), features.size(0))
# top5.update(prec5.item(), features.size(0))
log.info(' * Accuracy {acc.avg:.3f}'.format(acc=acc))
return acc.avg
def validate(val_loader, model, criterion):
""" Evaluate model using criterion on validation set """
losses = Meter(ptag='Loss')
top1 = Meter(ptag='Prec@1')
top5 = Meter(ptag='Prec@5')
# switch to evaluate mode
model.eval()
model.disable_gossip()
with torch.no_grad():
for i, (features, target) in enumerate(val_loader):
target = target.cuda(non_blocking=True)
# compute output
output = model(features)
loss = criterion(output, target)
# measure accuracy and record loss
prec1, prec5 = accuracy(output, target, topk=(1, 5))
losses.update(loss.item(), features.size(0))
top1.update(prec1.item(), features.size(0))
top5.update(prec5.item(), features.size(0))
log.info(' * Prec@1 {top1.avg:.3f} Prec@5 {top5.avg:.3f}'.format(
top1=top1, top5=top5))
return top1.avg
def validate(val_loader, model, criterion, log):
""" Evaluate model using criterion on validation set """
losses = Meter(ptag='Loss')
top1 = Meter(ptag='Prec@1')
top5 = Meter(ptag='Prec@5')
# switch to evaluate mode
model.eval()
with torch.no_grad():
for i, (features, target) in enumerate(val_loader):
# if args.fp16:
# features = features.cuda(non_blocking=True).half()
# This is not needed but let it be since there is no harm
target = target.cuda(non_blocking=True)
# create one-hot vector from target
kl_target = torch.zeros(
target.shape[0], 1000, device='cuda').scatter_(
1, target.view(-1, 1), 1)
# compute output
output = model(features)
loss = criterion(output, kl_target)
# measure accuracy and record loss
prec1, prec5 = accuracy(output, target, topk=(1, 5))
losses.update(loss.item(), features.size(0))
top1.update(prec1.item(), features.size(0))
top5.update(prec5.item(), features.size(0))
# log.info(' * Prec@1 {top1.avg:.3f} Prec@5 {top5.avg:.3f}'
# .format(top1=top1, top5=top5))
log.info(' * Prec@1 {top1.avg:.3f} Prec@5 {top5.avg:.3f} Loss {losses.avg:.3f}'
.format(top1=top1, top5=top5, losses = losses))
# print('pp * Prec@1 {top1.avg:.3f} Prec@5 {top5.avg:.3f} Loss {losses.avg:.3f}'
# .format(top1=top1, top5=top5, losses = losses))
return losses.avg, top1.avg, top5.avg
def train(model, criterion, optimizer, batch_meter, data_meter, nn_meter,
loader, epoch, itr, begin_time, num_itr_ignore):
losses = Meter(ptag='Loss')
acc = Meter(ptag="Accuracy")
# top1 = Meter(ptag='Prec@1')
# top5 = Meter(ptag='Prec@5')
# switch to train mode
model.train()
# spoof sampler to continue from checkpoint w/o loading data all over again
_train_loader = loader.__iter__()
for i in range(itr):
try:
next(_train_loader.sample_iter)
except Exception:
# finished epoch but prempted before state was updated
log.info('Loader spoof error attempt {}/{}'.format(i, len(loader)))
return
log.debug('Training (epoch {})'.format(epoch))
batch_time = time.time()
for i, (batch, target) in enumerate(_train_loader, start=itr):
target = target.cuda(non_blocking=True)
# create one-hot vector from target
# kl_target = torch.zeros(target.shape[0], 1000, device='cuda').scatter_(
# 1, target.view(-1, 1), 1)
if num_itr_ignore == 0:
data_meter.update(time.time() - batch_time)
# ----------------------------------------------------------- #
# Forward/Backward pass
# ----------------------------------------------------------- #
nn_time = time.time()
output = model(batch)
loss = criterion(output, target)
loss.backward()
if i % 100 == 0:
update_learning_rate(optimizer,
epoch,
itr=i,
itr_per_epoch=len(loader))
optimizer.step() # optimization update
optimizer.zero_grad()
if not args.overlap and not args.all_reduce:
log.debug('Transferring params')
model.transfer_params()
if num_itr_ignore == 0:
nn_meter.update(time.time() - nn_time)
# ----------------------------------------------------------- #
if num_itr_ignore == 0:
batch_meter.update(time.time() - batch_time)
batch_time = time.time()
log_time = time.time()
# measure accuracy and record loss
acc_val = accuracy(output, target)
losses.update(loss.item(), batch.size(0))
acc.update(acc_val, batch.size(0))
# top1.update(prec1.item(), batch.size(0))
# top5.update(prec5.item(), batch.size(0))
if i % args.print_freq == 0:
with open(args.out_fname, '+a') as f:
print('{ep},{itr},{bt},{nt},{dt},'
'{loss.val:.4f},{loss.avg:.4f},'
'{acc.val:.3f},{acc.avg:.3f},'
'-1'.format(ep=epoch,
itr=i,
bt=batch_meter,
dt=data_meter,
nt=nn_meter,
loss=losses,
acc=acc),
file=f)
if num_itr_ignore > 0:
num_itr_ignore -= 1
log_time = time.time() - log_time
log.debug(log_time)
if (args.num_iterations_per_training_epoch != -1
and i + 1 == args.num_iterations_per_training_epoch):
break
with open(args.out_fname, '+a') as f:
print('{ep},{itr},{bt},{nt},{dt},'
'{loss.val:.4f},{loss.avg:.4f},'
'{acc.val:.3f},{acc.avg:.3f},'
'-1'.format(ep=epoch,
itr=i,
bt=batch_meter,
dt=data_meter,
nt=nn_meter,
loss=losses,
acc=acc),
file=f)
def train(model, criterion, optimizer, batch_meter, data_meter, nn_meter,
loader, epoch, itr, begin_time):
losses = Meter(ptag='Loss')
top1 = Meter(ptag='Prec@1')
top5 = Meter(ptag='Prec@5')
# switch to train mode
model.train()
# spoof sampler to continue from checkpoint w/o loading data all over again
_train_loader = loader.__iter__()
for i in range(itr):
try:
next(_train_loader.sample_iter)
except Exception:
# finished epoch but prempted before state was updated
log.info('Loader spoof error attempt {}/{}'.format(i, len(loader)))
return
log.debug('Training (epoch {})'.format(epoch))
model.enable_gossip()
batch_time = time.time()
for i, (batch, target) in enumerate(_train_loader, start=itr):
target = target.cuda(non_blocking=True)
data_meter.update(time.time() - batch_time)
# ----------------------------------------------------------- #
# Forward/Backward pass
# ----------------------------------------------------------- #
nn_time = time.time()
output = model(batch)
loss = criterion(output, target)
bilat_freq = 100
if i == 0:
update_global_iteration_counter(itr=1, itr_per_epoch=len(loader))
update_bilat_learning_rate(model, itr_per_epoch=len(loader))
elif (i + args.rank) % (bilat_freq) == 0:
update_global_iteration_counter(itr=bilat_freq,
itr_per_epoch=len(loader))
update_bilat_learning_rate(model, itr_per_epoch=len(loader))
loss.backward()
update_learning_rate(optimizer,
epoch,
itr=i,
itr_per_epoch=len(loader))
optimizer.step() # optimization update
optimizer.zero_grad()
nn_meter.update(time.time() - nn_time)
# ----------------------------------------------------------- #
batch_meter.update(time.time() - batch_time)
batch_time = time.time()
log_time = time.time()
# measure accuracy and record loss
prec1, prec5 = accuracy(output, target, topk=(1, 5))
losses.update(loss.item(), batch.size(0))
top1.update(prec1.item(), batch.size(0))
top5.update(prec5.item(), batch.size(0))
if i % args.print_freq == 0:
ep = args.global_epoch
itr = args.global_itr % (len(loader) * args.world_size)
with open(args.out_fname, '+a') as f:
print('{ep},{itr},{bt},{nt},{dt},'
'{loss.val:.4f},{loss.avg:.4f},'
'{top1.val:.3f},{top1.avg:.3f},'
'{top5.val:.3f},{top5.avg:.3f},-1'.format(ep=ep,
itr=itr,
bt=batch_meter,
dt=data_meter,
nt=nn_meter,
loss=losses,
top1=top1,
top5=top5),
file=f)
log_time = time.time() - log_time
log.debug(log_time)
with open(args.out_fname, '+a') as f:
print('{ep},{itr},{bt},{nt},{dt},'
'{loss.val:.4f},{loss.avg:.4f},'
'{top1.val:.3f},{top1.avg:.3f},'
'{top5.val:.3f},{top5.avg:.3f},-1'.format(ep=epoch,
itr=i,
bt=batch_meter,
dt=data_meter,
nt=nn_meter,
loss=losses,
top1=top1,
top5=top5),
file=f)
def train(config, model, criterion, optimizer, batch_meter, data_meter, nn_meter,
loader, epoch, itr, begin_time, num_itr_ignore, log):
losses = Meter(ptag='Loss')
top1 = Meter(ptag='Prec@1')
top5 = Meter(ptag='Prec@5')
# switch to train mode
model.train()
# spoof sampler to continue from checkpoint w/o loading data all over again
_train_loader = loader.__iter__()
for i in range(itr):
try:
next(_train_loader.sample_iter)
except Exception:
# finished epoch but prempted before state was updated
log.info('Loader spoof error attempt {}/{}'.format(i, len(loader)))
return
log.debug('Training (epoch {})'.format(epoch))
batch_time = time.time()
for i, (batch, target) in enumerate(_train_loader, start=itr):
# if args.fp16:
# batch = batch.cuda(non_blocking=True).half()
target = target.cuda(non_blocking=True)
# create one-hot vector from target
kl_target = torch.zeros(target.shape[0], 1000, device='cuda').scatter_(
1, target.view(-1, 1), 1)
if num_itr_ignore == 0:
data_meter.update(time.time() - batch_time)
# ----------------------------------------------------------- #
# Forward/Backward pass
# ----------------------------------------------------------- #
nn_time = time.time()
output = model(batch)
loss = criterion(output, kl_target)
# if args.fp16:
# if args.amp:
# with amp_handle.scale_loss(loss, optimizer) as scaled_loss:
# scaled_loss.backward()
# else:
# optimizer.backward(loss)
# else:
# loss.backward()
loss.backward()
if i % 100 == 0:
update_learning_rate(config, optimizer, epoch, log, itr=i,
itr_per_epoch=len(loader))
optimizer.step() # optimization update
optimizer.zero_grad()
if not config['overlap'] and not config['all_reduce']:
log.debug('Transferring params')
model.transfer_params()
if num_itr_ignore == 0:
nn_meter.update(time.time() - nn_time)
# ----------------------------------------------------------- #
if num_itr_ignore == 0:
batch_meter.update(time.time() - batch_time)
batch_time = time.time()
log_time = time.time()
# measure accuracy and record loss
prec1, prec5 = accuracy(output, target, topk=(1, 5))
losses.update(loss.item(), batch.size(0))
top1.update(prec1.item(), batch.size(0))
top5.update(prec5.item(), batch.size(0))
if i % config['print_freq'] == 0:
with open(config['out_fname'], '+a') as f:
print('{ep},{itr},{bt},{nt},{dt},'
'{loss.val:.4f},{loss.avg:.4f},'
'{top1.val:.3f},{top1.avg:.3f},'
'{top5.val:.3f},{top5.avg:.3f},-1'
.format(ep=epoch, itr=i,
bt=batch_meter,
dt=data_meter, nt=nn_meter,
loss=losses, top1=top1,
top5=top5), file=f)
if num_itr_ignore > 0:
num_itr_ignore -= 1
log_time = time.time() - log_time
log.debug(log_time)
if (config['num_iterations_per_training_epoch'] != -1 and
i+1 == config['num_iterations_per_training_epoch']):
break
with open(config['out_fname'], '+a') as f:
print('{ep},{itr},{bt},{nt},{dt},'
'{loss.val:.4f},{loss.avg:.4f},'
'{top1.val:.3f},{top1.avg:.3f},'
'{top5.val:.3f},{top5.avg:.3f},-1'
.format(ep=epoch, itr=i,
bt=batch_meter,
dt=data_meter, nt=nn_meter,
loss=losses, top1=top1,
top5=top5), file=f)
return losses.avg, top1.avg, top5.avg