def load_model(checkpoints_folder,device):
model =ResNetSimCLR(**config['model'])
model.eval()
state_dict = torch.load(os.path.join(checkpoints_folder, 'model.pth'), map_location=torch.device('cpu'))
model.load_state_dict(state_dict)
model = model.to(device)
return model
def encode(save_root, model_file, data_folder, model_name='ca', dataset_name='celeba', batch_size=64, device='cuda:0', out_dim=256):
os.makedirs(save_root, exist_ok=True)
os.makedirs(data_folder, exist_ok=True)
if dataset_name == 'celeba':
train_loader = DataLoader(datasets.CelebA(data_folder, split='train', download=True, transform=transforms.ToTensor()),
batch_size=batch_size, shuffle=False)
valid_loader = DataLoader(datasets.CelebA(data_folder, split='valid', download=True, transform=transforms.ToTensor()),
batch_size=batch_size, shuffle=False)
elif dataset_name == 'stanfordCars':
t = transforms.Compose([
transforms.Resize(512),
transforms.CenterCrop(512),
transforms.ToTensor(),
transforms.Lambda(lambda x: x.repeat(3,1,1) if x.shape[0] == 1 else x)
])
train_data_dir = os.path.join(data_folder, 'cars_train/')
train_annos = os.path.join(data_folder, 'devkit/cars_train_annos.mat')
train_loader = DataLoader(CarsDataset(train_annos, train_data_dir, t), batch_size=batch_size, shuffle=False)
valid_data_dir = os.path.join(data_folder, 'cars_test/')
valid_annos = os.path.join(data_folder, 'devkit/cars_test_annos_withlabels.mat')
valid_loader = DataLoader(CarsDataset(valid_annos, valid_data_dir, t), batch_size=batch_size, shuffle=False)
elif dataset_name == 'compCars':
t = transforms.Compose([
transforms.Resize(512),
transforms.CenterCrop(512),
transforms.ToTensor()
])
train_loader = DataLoader(CompCars(data_folder, True, t), batch_size=batch_size, shuffle=False)
valid_loader = DataLoader(CompCars(data_folder, False, t), batch_size=batch_size, shuffle=False)
model = ResNetSimCLR('resnet50', out_dim)
model.load_state_dict(torch.load(model_file, map_location=device))
model = model.to(device)
model.eval()
print('Starting on training data')
train_encodings = []
for x, _ in train_loader:
x = x.to(device)
h, _ = model(x)
train_encodings.append(h.cpu().detach())
torch.save(torch.cat(train_encodings, dim=0), os.path.join(save_root, f'{dataset_name}-{model_name}model-train_encodings.pt'))
print('Starting on validation data')
valid_encodings = []
for x, _ in valid_loader:
x = x.to(device)
h, _ = model(x)
if len(h.shape) == 1:
h = h.unsqueeze(0)
valid_encodings.append(h.cpu().detach())
torch.save(torch.cat(valid_encodings, dim=0), os.path.join(save_root, f'{dataset_name}-{model_name}model-valid_encodings.pt'))
def evaluation(checkpoints_folder, config, device):
model = ResNetSimCLR(**config['model'])
model.eval()
model.load_state_dict(
torch.load(os.path.join(checkpoints_folder, 'model.pth')))
model = model.to(device)
train_set = torchvision.datasets.CIFAR10(
root='../data/CIFAR10',
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
train=True,
download=True)
test_set = torchvision.datasets.CIFAR10(root='../data/CIFAR10',
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
train=False,
download=True)
# num_train = len(train_dataset)
# indices = list(range(num_train))
# np.random.shuffle(indices)
#
# split = int(np.floor(0.05 * num_train))
# train_idx, test_idx = indices[split:], indices[:split]
#
# # define samplers for obtaining training and validation batches
# train_sampler = torch.utils.data.sampler.SubsetRandomSampler(train_idx)
# test_sampler = torch.utils.data.sampler.SubsetRandomSampler(test_idx) # ?????sampler????????????
# train_loader = torch.utils.data.DataLoader(train_set, batch_size=config['batch_size'], drop_last=True, shuffle=True)
#
# test_loader = torch.utils.data.DataLoader(test_set, batch_size=config['batch_size'], drop_last=True, shuffle=True)
train_loader = torch.utils.data.DataLoader(train_set,
batch_size=48,
drop_last=True,
shuffle=True)
test_loader = torch.utils.data.DataLoader(test_set,
batch_size=48,
drop_last=True,
shuffle=True)
X_train_feature = []
label_train = []
for data in train_loader:
x, y = data
x = x.to(device)
features, _ = model(x)
X_train_feature.extend(features.cpu().detach().numpy())
label_train.extend(y.cpu().detach().numpy())
X_train_feature = np.array(X_train_feature)
label_train = np.array(label_train)
X_test_feature = []
label_test = []
for data in test_loader:
x, y = data
x = x.to(device)
features, _ = model(x)
X_test_feature.extend(features.cpu().detach().numpy())
label_test.extend(y.cpu().detach().numpy())
X_test_feature = np.array(X_test_feature)
label_test = np.array(label_test)
scaler = preprocessing.StandardScaler()
print('ok')
scaler.fit(X_train_feature)
# print(X_test_feature.shape)
# print(y_test.shape)
linear_model_eval(scaler.transform(X_train_feature), label_train,
scaler.transform(X_test_feature), label_test)
def eval_frozen(self):
train_loader, val_loader, num_classes = self.dataset.get_dataset_eval()
model = ResNetSimCLR(**self.config["model"]).to(self.device)
model = self._load_pre_trained_weights(model)
model.to(self.device)
model.eval()
lineal_classifier = nn.Linear(model.l1.in_features, num_classes)
lineal_classifier.to(self.device)
optimizer = torch.optim.SGD(lineal_classifier.parameters(),
1e-3,
weight_decay=eval(
self.config['weight_decay']))
epochs = self.config['epochs']
scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=[int(0.5 * epochs),
int(0.8 * epochs)],
gamma=0.1,
last_epoch=-1)
criterion = nn.CrossEntropyLoss()
best_acc = 0
for epoch in range(epochs):
print('=' * 20)
print(f'Epoch: {epoch + 1} / {epochs}')
top1 = 0
running_loss = 0
n = 0
lineal_classifier.train()
for idx, (img, lab) in enumerate(train_loader):
B = img.size(0)
img = img.to(self.device)
lab = lab.to(self.device)
optimizer.zero_grad()
loss, top1_batch = self._step_eval_train(
model, lineal_classifier, img, lab)
loss.backward()
optimizer.step()
top1 += (x.argmax(dim=1) == lab).sum().item()
running_loss += loss.item() * B
n += B
print('Training {}/{} - Loss: {:.2f} - top 1: {:.2f}'.format(
idx + 1, len(train_loader), running_loss / n,
100 * top1 / n),
end='\r')
print('\n')
top1 = 0
running_loss = 0
n = 0
lineal_classifier.eval()
for idx, (img, lab) in enumerate(val_loader):
B = img.size(0)
img = img.to(self.device)
lab = lab.to(self.device)
loss, top1_batch = self._step_eval_eval(
model, lineal_classifier, img, lab)
top1 += top1_batch
running_loss += loss.item() * B
n += B
print('Val {}/{} - Loss: {:.2f} - top 1: {:.2f}'.format(
idx + 1, len(val_loader), running_loss / n,
100 * top1 / n),
end='\r')
print('\n')
if best_acc < top1:
best_acc = top1
print(f'Best ACC: {best_acc * 100 / n}')
def main():
args = parser.parse_args()
assert args.n_views == 2, "Only two view training is supported. Please use --n-views 2."
# check if gpu training is available
if not args.disable_cuda and torch.cuda.is_available():
args.device = torch.device('cuda')
cudnn.deterministic = True
cudnn.benchmark = True
else:
args.device = torch.device('cpu')
args.gpu_index = -1
if args.mode == 'simclr':
dataset = ContrastiveLearningDataset(args.data)
train_dataset = dataset.get_dataset(args.dataset_name,
args.n_views,
train=True)
model = ResNetSimCLR(base_model=args.arch, out_dim=args.out_dim)
trainer_class = SimCLRTrainer
elif args.mode == 'supervised':
dataset = SupervisedLearningDataset(args.data)
train_dataset = dataset.get_dataset(args.dataset_name,
args.supervised_augments,
train=True)
model = ResNetSimCLR(base_model=args.arch,
out_dim=len(train_dataset.classes))
trainer_class = SupervisedTrainer
else:
raise InvalidTrainingMode()
if args.target_shuffle is not None:
random.seed(args.target_shuffle)
random.shuffle(train_dataset.targets)
checkpoints = []
for root, dirs, files in os.walk(
os.path.join('experiments', args.experiment_group, 'wandb')):
for file in files:
if file == args.estimate_checkpoint:
checkpoints += [os.path.join(root, file)]
set_random_seed(args.seed)
sample_indices = torch.randint(len(train_dataset),
size=(args.batch_size *
args.estimate_batches, ))
# It’s a no-op if the 'gpu_index' argument is a negative integer or None.
estimated_prob, estimated_argmax = [], []
with torch.cuda.device(args.gpu_index):
for file in checkpoints:
state = torch.load(file)
model.load_state_dict(state['model'])
model.eval()
trainer = trainer_class(model=model,
optimizer=None,
scheduler=None,
args=args)
checkpoint_prob, checkpoint_argmax = [], []
for i in range(args.estimate_batches):
if args.fixed_augments:
set_random_seed(args.seed)
if args.mode == 'simclr':
images = [[], []]
for index in sample_indices[i:i + args.batch_size]:
example = train_dataset[index][0]
images[0] += [example[0]]
images[1] += [example[1]]
images[0] = torch.stack(images[0], dim=0)
images[1] = torch.stack(images[1], dim=0)
labels = None
elif args.mode == 'supervised':
images, labels = [], []
for index in sample_indices[i:i + args.batch_size]:
example = train_dataset[index]
images += [example[0]]
labels += [example[1]]
images = torch.stack(images, dim=0)
labels = torch.tensor(labels, dtype=torch.long)
with torch.no_grad():
logits, labels = trainer.calculate_logits(images, labels)
prob = torch.softmax(logits,
dim=1)[torch.arange(labels.shape[0]),
labels]
checkpoint_prob += [prob.detach().cpu()]
argmax = (torch.argmax(logits,
dim=1) == labels).to(torch.int)
checkpoint_argmax += [argmax.detach().cpu()]
checkpoint_prob = torch.cat(checkpoint_prob, dim=0)
estimated_prob += [checkpoint_prob]
checkpoint_argmax = torch.cat(checkpoint_argmax, dim=0)
estimated_argmax += [checkpoint_argmax]
estimated_prob = torch.stack(estimated_prob, dim=0)
estimated_argmax = torch.stack(estimated_argmax, dim=0)
torch.save(
{
'indices': sample_indices,
'prob': estimated_prob,
'argmax': estimated_argmax
}, os.path.join('experiments', args.experiment_group, args.out_file))
if train_gpu:
xis = xis.cuda()
xjs = xjs.cuda()
loss = step(xis, xjs)
train_writer.add_scalar('train_loss', loss, global_step=n_iter)
loss.backward()
optimizer.step()
n_iter += 1
if epoch_counter % config['eval_every_n_epochs'] == 0:
# validation steps
with torch.no_grad():
model.eval()
valid_loss = 0.0
for counter, ((xis, xjs), _) in enumerate(valid_loader):
if train_gpu:
xis = xis.cuda()
xjs = xjs.cuda()
loss = (step(xis, xjs))
valid_loss += loss.item()
valid_loss /= counter
if valid_loss < best_valid_loss:
# save the model weights
best_valid_loss = valid_loss
def main():
args = parser.parse_args()
assert args.n_views == 2, "Only two view training is supported. Please use --n-views 2."
# check if gpu training is available
if not args.disable_cuda and torch.cuda.is_available():
print("Using GPU")
args.device = torch.device('cuda')
cudnn.deterministic = True
cudnn.benchmark = True
else:
args.device = torch.device('cpu')
args.gpu_index = -1
dataset = ContrastiveLearningDataset(args.data)
train_dataset = dataset.get_dataset(args.dataset_name, args.n_views)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True,
drop_last=True,
worker_init_fn=worker_init_fn)
if args.dataset_name == "mnist":
in_channels = 1
else:
in_channels = 3
model = ResNetSimCLR(base_model=args.arch,
out_dim=args.out_dim,
in_channels=in_channels)
if args.model_path is not None:
checkpoint = torch.load(args.model_path, map_location=args.device)
model.load_state_dict(checkpoint['state_dict'])
optimizer = torch.optim.Adam(model.parameters(),
args.lr,
weight_decay=args.weight_decay)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=len(train_loader), eta_min=0, last_epoch=-1)
head_dataset = HeadDataset(args.data)
train_head_dataset = head_dataset.get_dataset(args.dataset_name,
train=True,
split="train")
test_head_dataset = head_dataset.get_dataset(args.dataset_name,
train=False,
split="test")
args.num_classes = head_dataset.get_num_classes(args.dataset_name)
train_head_loader = torch.utils.data.DataLoader(train_head_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True,
drop_last=True)
test_head_loader = torch.utils.data.DataLoader(test_head_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True,
drop_last=True)
# It’s a no-op if the 'gpu_index' argument is a negative integer or None.
with torch.cuda.device(args.gpu_index):
if not (args.head_only or args.tsne_only):
simclr = SimCLR(model=model,
optimizer=optimizer,
scheduler=scheduler,
args=args)
model = simclr.train(train_loader)
model.eval()
if not args.tsne_only:
headsimclr = SimCLRHead(model=model, args=args)
headsimclr.train(train_head_loader, test_head_loader)
tsne_plot = TSNE_project(model, test_head_loader, args)
