check_trace=True,
optimize=args.optimize)
if args.action == 'inspect':
framework.inspect(network,
val_loader,
args.artifacts,
early_stop=args.early_stop)
elif args.action == 'evaluate':
callbacks = [PrecRec(), IsicIoU()]
framework.test(
network,
val_loader,
loss_fn,
callbacks,
start_epoch,
writer=writer,
early_stop=args.early_stop,
)
for callback in callbacks:
results = callback.get_dict()
print(results)
for key in results:
writer.add_scalar(f'Evaluation/{key}', results[key],
start_epoch)
elif args.action == 'train':
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optim,
'min',
example = next(iter(train_loader))[0].cuda()
network = torch.jit.trace(network,
example,
check_trace=True,
optimize=args.optimize)
if args.action == 'inspect':
framework.inspect(network,
val_loader,
args.artifacts,
early_stop=args.early_stop)
elif args.action == 'evaluate':
prec_rec = PrecRec()
framework.test(network,
val_loader,
loss_fn, [],
logger=logger,
callbacks=[prec_rec])
elif args.action == 'train':
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optim,
'min',
patience=2,
verbose=True,
cooldown=1)
for epoch in range(start_epoch, args.epochs):
train_loss = framework.train(network,
train_loader,
loss_fn,
optim,
if not args.load:
print('Set start epoch and best score to 0')
best_score = 0.
start_epoch = 0
if args.action == 'inspect':
framework.inspect(network,
val_loader,
args.artifacts,
criteria=criteria,
early_stop=args.early_stop)
elif args.action == 'evaluate':
prec_rec = PrecRec()
framework.test(network,
val_loader,
criteria,
logger=logger,
callbacks=[prec_rec])
f1, thres = prec_rec.best_f1()
print('F1', f1, 'at', thres)
elif args.action == 'train':
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optim, 'min', patience=2, verbose=True, cooldown=1)
for epoch in range(start_epoch, args.epochs):
train_loss = framework.train(network,
train_loader,
loss_fn,
optim,
epoch,