in eval_linear.py [0:0]
def eval_linear(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
cudnn.benchmark = True
# ============ building network ... ============
# if the network is a Vision Transformer (i.e. vit_tiny, vit_small, vit_base)
if args.arch in vits.__dict__.keys():
model = vits.__dict__[args.arch](patch_size=args.patch_size, num_classes=0)
embed_dim = model.embed_dim * (args.n_last_blocks + int(args.avgpool_patchtokens))
# if the network is a XCiT
elif "xcit" in args.arch:
model = torch.hub.load('facebookresearch/xcit:main', args.arch, num_classes=0)
embed_dim = model.embed_dim
# otherwise, we check if the architecture is in torchvision models
elif args.arch in torchvision_models.__dict__.keys():
model = torchvision_models.__dict__[args.arch]()
embed_dim = model.fc.weight.shape[1]
model.fc = nn.Identity()
else:
print(f"Unknow architecture: {args.arch}")
sys.exit(1)
model.cuda()
model.eval()
# load weights to evaluate
utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
print(f"Model {args.arch} built.")
linear_classifier = LinearClassifier(embed_dim, num_labels=args.num_labels)
linear_classifier = linear_classifier.cuda()
linear_classifier = nn.parallel.DistributedDataParallel(linear_classifier, device_ids=[args.gpu])
# ============ preparing data ... ============
val_transform = pth_transforms.Compose([
pth_transforms.Resize(256, interpolation=3),
pth_transforms.CenterCrop(224),
pth_transforms.ToTensor(),
pth_transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
dataset_val = datasets.ImageFolder(os.path.join(args.data_path, "val"), transform=val_transform)
val_loader = torch.utils.data.DataLoader(
dataset_val,
batch_size=args.batch_size_per_gpu,
num_workers=args.num_workers,
pin_memory=True,
)
if args.evaluate:
utils.load_pretrained_linear_weights(linear_classifier, args.arch, args.patch_size)
test_stats = validate_network(val_loader, model, linear_classifier, args.n_last_blocks, args.avgpool_patchtokens)
print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
return
train_transform = pth_transforms.Compose([
pth_transforms.RandomResizedCrop(224),
pth_transforms.RandomHorizontalFlip(),
pth_transforms.ToTensor(),
pth_transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
dataset_train = datasets.ImageFolder(os.path.join(args.data_path, "train"), transform=train_transform)
sampler = torch.utils.data.distributed.DistributedSampler(dataset_train)
train_loader = torch.utils.data.DataLoader(
dataset_train,
sampler=sampler,
batch_size=args.batch_size_per_gpu,
num_workers=args.num_workers,
pin_memory=True,
)
print(f"Data loaded with {len(dataset_train)} train and {len(dataset_val)} val imgs.")
# set optimizer
optimizer = torch.optim.SGD(
linear_classifier.parameters(),
args.lr * (args.batch_size_per_gpu * utils.get_world_size()) / 256., # linear scaling rule
momentum=0.9,
weight_decay=0, # we do not apply weight decay
)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs, eta_min=0)
# Optionally resume from a checkpoint
to_restore = {"epoch": 0, "best_acc": 0.}
utils.restart_from_checkpoint(
os.path.join(args.output_dir, "checkpoint.pth.tar"),
run_variables=to_restore,
state_dict=linear_classifier,
optimizer=optimizer,
scheduler=scheduler,
)
start_epoch = to_restore["epoch"]
best_acc = to_restore["best_acc"]
for epoch in range(start_epoch, args.epochs):
train_loader.sampler.set_epoch(epoch)
train_stats = train(model, linear_classifier, optimizer, train_loader, epoch, args.n_last_blocks, args.avgpool_patchtokens)
scheduler.step()
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
'epoch': epoch}
if epoch % args.val_freq == 0 or epoch == args.epochs - 1:
test_stats = validate_network(val_loader, model, linear_classifier, args.n_last_blocks, args.avgpool_patchtokens)
print(f"Accuracy at epoch {epoch} of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
best_acc = max(best_acc, test_stats["acc1"])
print(f'Max accuracy so far: {best_acc:.2f}%')
log_stats = {**{k: v for k, v in log_stats.items()},
**{f'test_{k}': v for k, v in test_stats.items()}}
if utils.is_main_process():
with (Path(args.output_dir) / "log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
save_dict = {
"epoch": epoch + 1,
"state_dict": linear_classifier.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
"best_acc": best_acc,
}
torch.save(save_dict, os.path.join(args.output_dir, "checkpoint.pth.tar"))
print("Training of the supervised linear classifier on frozen features completed.\n"
"Top-1 test accuracy: {acc:.1f}".format(acc=best_acc))