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
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)
model = ResNetSimCLR(base_model=args.arch, out_dim=args.out_dim)
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)
# It’s a no-op if the 'gpu_index' argument is a negative integer or None.
with torch.cuda.device(args.gpu_index):
simclr = SimCLR(model=model, optimizer=optimizer, scheduler=scheduler, args=args)
simclr.train(train_loader)
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
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)
model = ResNetSimCLR(base_model=args.arch,
out_dim=args.out_dim).to(args.device)
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)
if args.resume:
if os.path.isfile(args.resume):
print("=> loading resumed checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume,
map_location=torch.device('cpu'))
args.start_epoch = checkpoint['epoch']
args.start_iter = checkpoint['iteration']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
print("=> loaded resumed checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("[Warning] no checkpoint found at '{}'".format(args.resume))
else:
args.start_iter = 0
args.start_epoch = 0
# It’s a no-op if the 'gpu_index' argument is a negative integer or None.
with torch.cuda.device(args.gpu_index):
simclr = SimCLR(model=model,
optimizer=optimizer,
scheduler=scheduler,
args=args)
simclr.train(train_loader)
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
set_random_seed(args.seed)
train_loader, valid_loader = get_dataloaders(args)
if args.mode == 'simclr':
model = ResNetSimCLR(base_model=args.arch, out_dim=args.out_dim)
trainer_class = SimCLRTrainer
elif args.mode == 'supervised':
model = ResNetSimCLR(base_model=args.arch, out_dim=len(train_loader.dataset.classes))
trainer_class = SupervisedTrainer
else:
raise InvalidTrainingMode()
if args.optimizer_mode == 'simclr':
optimizer = torch.optim.Adam(model.parameters(), lr=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)
else:
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr,
momentum=0.9, weight_decay=args.weight_decay)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=200)
# It’s a no-op if the 'gpu_index' argument is a negative integer or None.
with torch.cuda.device(args.gpu_index):
trainer = trainer_class(model=model, optimizer=optimizer, scheduler=scheduler, args=args)
trainer.train(train_loader, valid_loader)
def run(config):
model = ResNetSimCLR('resnet50', config.out_dim)
model.load_state_dict(
torch.load(config.model_file, map_location=config.device))
model = model.to(config.device)
clf = nn.Linear(2048, 196)
full = FineTuner(model, clf)
optim = torch.optim.Adam(list(model.parameters()) + list(clf.parameters()))
objective = nn.CrossEntropyLoss()
t = T.Compose([
T.Resize(512),
T.CenterCrop(512),
T.ToTensor(),
T.Lambda(lambda x: x.repeat(3, 1, 1) if x.shape[0] == 1 else x)
])
train_data_dir = os.path.join(config.data_root, 'cars_train/')
train_annos = os.path.join(config.data_root, 'devkit/cars_train_annos.mat')
train_dataset = StanfordCarsMini(train_annos, train_data_dir, t)
train_loader = DataLoader(train_dataset,
batch_size=config.batch_size,
shuffle=True)
valid_data_dir = os.path.join(config.data_root, 'cars_test/')
valid_annos = os.path.join(config.data_root,
'devkit/cars_test_annos_withlabels.mat')
valid_dataset = CarsDataset(valid_annos, valid_data_dir, t)
valid_loader = DataLoader(valid_dataset, batch_size=config.batch_size)
solver = CESolver(full,
train_loader,
valid_loader,
config.save_root,
name=config.name,
device=config.device)
solver.train(config.num_epochs)
def train(self):
train_loader, valid_loader = self.dataset.get_data_loaders()
model = ResNetSimCLR(**self.config["model"]).to(self.device)
model = self._load_pre_trained_weights(model)
optimizer = torch.optim.Adam(model.parameters(),
3e-4,
weight_decay=eval(
self.config['weight_decay']))
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=self.config['epochs'], eta_min=0, last_epoch=-1)
if apex_support and self.config['fp16_precision']:
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O2',
keep_batchnorm_fp32=True)
model_checkpoints_folder = os.path.join(self.writer.log_dir,
'checkpoints')
# save config file
_save_config_file(model_checkpoints_folder)
n_iter = 0
valid_n_iter = 0
best_valid_loss = np.inf
for epoch_counter in range(self.config['epochs']):
for (xis, xjs), _ in train_loader:
optimizer.zero_grad()
xis = xis.to(self.device)
xjs = xjs.to(self.device)
loss = self._step(model, xis, xjs, n_iter)
if n_iter % self.config['log_every_n_steps'] == 0:
self.writer.add_scalar('train_loss',
loss,
global_step=n_iter)
if apex_support and self.config['fp16_precision']:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
n_iter += 1
# validate the model if requested
if epoch_counter % self.config['eval_every_n_epochs'] == 0:
valid_loss = self._validate(model, valid_loader)
if valid_loss < best_valid_loss:
# save the model weights
best_valid_loss = valid_loss
torch.save(
model.state_dict(),
os.path.join(model_checkpoints_folder, 'model.pth'))
self.writer.add_scalar('validation_loss',
valid_loss,
global_step=valid_n_iter)
valid_n_iter += 1
# warmup for the first 10 epochs
if epoch_counter >= 10:
scheduler.step()
self.writer.add_scalar('cosine_lr_decay',
scheduler.get_lr()[0],
global_step=n_iter)
def train(self):
train_loader, valid_loader = self.dataset.get_data_loaders()
model = ResNetSimCLR(**self.config["model"]).to(self.device)
model = self._load_pre_trained_weights(model)
if self.augmentor_type == "cnn":
if self.config["normalization_type"] == "original":
augmentor = LpAugmentor(
clip=self.config["augmentor_clip_output"])
augmentor.to(self.device)
elif self.config["normalization_type"] == "spectral":
augmentor = LpAugmentorSpecNorm(
clip=self.config["augmentor_clip_output"])
augmentor.to(self.device)
else:
raise ValueError("Unregonized normalization type: {}".format(
self.config["normalization_type"]))
elif self.augmentor_type == "style_transfer":
augmentor = LpAugmentorStyleTransfer(
clip=self.config["augmentor_clip_output"])
augmentor.to(self.device)
elif self.augmentor_type == "transformer":
augmentor = LpAugmentorTransformer(
clip=self.config["augmentor_clip_output"])
augmentor.to(self.device)
else:
raise ValueError("Unrecognized augmentor type: {}".format(
self.augmentor_type))
augmentor_optimizer = torch.optim.Adam(augmentor.parameters(), 3e-4)
augmentor_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
augmentor_optimizer,
T_max=len(train_loader),
eta_min=0,
last_epoch=-1)
optimizer = torch.optim.Adam(
list(model.parameters()),
3e-4,
weight_decay=eval(self.config["weight_decay"]),
)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=len(train_loader), eta_min=0, last_epoch=-1)
if apex_support and self.config["fp16_precision"]:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
keep_batchnorm_fp32=True)
model_checkpoints_folder = os.path.join(self.writer.log_dir,
"checkpoints")
# save config file
_save_config_file(model_checkpoints_folder)
n_iter = 0
valid_n_iter = 0
best_valid_loss = np.inf
for epoch_counter in range(self.config["epochs"]):
print("====== Epoch {} =======".format(epoch_counter))
for (xis, xjs), _ in train_loader:
optimizer.zero_grad()
xis = xis.to(self.device)
xjs = xjs.to(self.device)
loss = self._adv_step(model, augmentor, xis, xjs, n_iter)
if n_iter % self.config["log_every_n_steps"] == 0:
self.writer.add_scalar("train_loss",
loss,
global_step=n_iter)
if apex_support and self.config["fp16_precision"]:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# for p in augmentor.parameters():
# # print(p.name)
# p.grad *= -1.0
optimizer.step()
# Update augmentor
augmentor_optimizer.zero_grad()
loss = self._adv_step(model, augmentor, xis, xjs, n_iter)
if self.augmentor_loss_type == "hinge":
loss = torch.clamp(loss, 0.0, 5.4)
loss *= -1.0
loss.backward()
augmentor_optimizer.step()
n_iter += 1
# validate the model if requested
if epoch_counter % self.config["eval_every_n_epochs"] == 0:
valid_loss = self._validate(model, augmentor, valid_loader)
if valid_loss < best_valid_loss:
# save the model weights
best_valid_loss = valid_loss
torch.save(
model.state_dict(),
os.path.join(model_checkpoints_folder, "model.pth"),
)
print("validation loss: ", valid_loss)
self.writer.add_scalar("validation_loss",
valid_loss,
global_step=valid_n_iter)
valid_n_iter += 1
# warmup for the first 10 epochs
if epoch_counter >= 10:
scheduler.step()
augmentor_scheduler.step()
self.writer.add_scalar("cosine_lr_decay",
scheduler.get_lr()[0],
global_step=n_iter)
def main_worker(gpu, ngpus_per_node, args):
args.gpu = gpu
# suppress printing if not master
if args.multiprocessing_distributed and args.gpu != 0:
def print_pass(*args):
pass
builtins.print = print_pass
if args.gpu is not None:
print(args.gpu)
print("Use GPU: {} for training".format(args.gpu))
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
# create model
print("=> creating model '{}'".format(args.arch))
model = ResNetSimCLR(base_model=args.arch,
out_dim=args.out_dim).to(args.gpu)
if args.distributed:
# For multiprocessing distributed, DistributedDataParallel constructor
# should always set the single device scope, otherwise,
# DistributedDataParallel will use all available devices.
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
# When using a single GPU per process and per
# DistributedDataParallel, we need to divide the batch size
# ourselves based on the total number of GPUs we have
args.batch_size = int(args.batch_size / ngpus_per_node)
args.num_workers = int(
(args.num_workers + ngpus_per_node - 1) / ngpus_per_node)
model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
else:
model.cuda()
# DistributedDataParallel will divide and allocate batch_size to all
# available GPUs if device_ids are not set
model = torch.nn.parallel.DistributedDataParallel(model)
elif args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
# comment out the following line for debugging
#raise NotImplementedError("Only DistributedDataParallel is supported.")
#else:
# AllGather implementation (batch shuffle, queue update, etc.) in
# this code only supports DistributedDataParallel.
#raise NotImplementedError("Only DistributedDataParallel is supported.")
# Data loader
train_loader, train_sampler = data_loader(args.dataset,
args.data_path,
args.batch_size,
args.num_workers,
download=args.download,
distributed=args.distributed,
supervised=False)
#optimizer = torch.optim.Adam(model.parameters(), 3e-4, weight_decay=args.weight_decay)
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,
T_max=args.epochs,
eta_min=0,
last_epoch=-1)
criterion = NTXentLoss(args.gpu, args.batch_size, args.temperature,
True).cuda(args.gpu)
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
if args.gpu is None:
checkpoint = torch.load(args.resume)
else:
loc = 'cuda:{}'.format(args.gpu)
checkpoint = torch.load(args.resume, map_location=loc)
args.start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
if apex_support and args.fp16_precision:
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O2',
keep_batchnorm_fp32=True)
cudnn.benchmark = True
train(model, train_loader, train_sampler, criterion, optimizer, scheduler,
args, ngpus_per_node)
num_workers=config['num_workers'],
drop_last=True,
shuffle=True)
# model = Encoder(out_dim=out_dim)
model = ResNetSimCLR(base_model=config["base_convnet"], out_dim=out_dim)
train_gpu = torch.cuda.is_available()
print("Is gpu available:", train_gpu)
# moves the model parameters to gpu
if train_gpu:
model.cuda()
criterion = torch.nn.CrossEntropyLoss(reduction='sum')
optimizer = optim.Adam(model.parameters(), 3e-4)
train_writer = SummaryWriter()
sim_func_dim1, sim_func_dim2 = get_similarity_function(use_cosine_similarity)
# Mask to remove positive examples from the batch of negative samples
negative_mask = get_negative_mask(batch_size)
n_iter = 0
for e in range(config['epochs']):
for step, ((xis, xjs), _) in enumerate(train_loader):
optimizer.zero_grad()
if train_gpu:
def train(self):
train_loader, valid_loader = self.dataset.get_data_loaders()
model = ResNetSimCLR(**self.config["model"]).to(self.device)
model = self._load_pre_trained_weights(model)
criterion = nn.CrossEntropyLoss() # loss function
optimizer = torch.optim.Adam(model.parameters(),
3e-4,
weight_decay=eval(
self.config['weight_decay']))
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=self.config['epochs'], eta_min=0, last_epoch=-1)
if apex_support and self.config['fp16_precision']:
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O2',
keep_batchnorm_fp32=True)
model_checkpoints_folder = os.path.join(
'/home/zhangchunhui/MedicalAI/USCL/checkpoints_multi_aug',
'checkpoint_' + str(self.Checkpoint_Num))
# save config file
_save_config_file(model_checkpoints_folder)
start_time = time.time()
end_time = time.time()
valid_n_iter = 0
best_valid_loss = np.inf
for epoch in range(self.config['epochs']):
for i, data in enumerate(train_loader, 1):
# forward
# mixupimg1, label1, mixupimg2, label2, original img1, original img2
xis, labelis, xjs, labeljs, imgis, imgjs = data # N samples of left branch, N samples of right branch
xis = xis.to(self.device)
xjs = xjs.to(self.device)
####### 1-Semi-supervised
hi, xi, outputis = model(xis)
hj, xj, outputjs = model(xjs)
labelindexi, labelindexj = FindNotX(
labelis.tolist(), 9999), FindNotX(labeljs.tolist(),
9999) # X=9999=no label
lossi = criterion(outputis[labelindexi],
labelis.to(self.device)[labelindexi])
lossj = criterion(outputjs[labelindexj],
labeljs.to(self.device)[labelindexj])
# lumbda1=lumbda2 # small value is better
lumbda1, lumbda2 = self.lumbda1, self.lumbda2 # small value is better
loss = self._step(model, xis,
xjs) + lumbda1 * lossi + lumbda2 * lossj
########################################################################################################
####### 2-Self-supervised
# loss = self._step(model, xis, xjs)
########################################################################################################
# backward
optimizer.zero_grad()
if apex_support and self.config['fp16_precision']:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# update weights
optimizer.step()
if i % self.config['log_every_n_steps'] == 0:
# self.writer.add_scalar('train_loss', loss, global_step=i)
start_time, end_time = end_time, time.time()
print(
"\nTraining:Epoch[{:0>3}/{:0>3}] Iteration[{:0>3}/{:0>3}] Loss: {:.4f} Time: {:.2f}s"
.format(epoch + 1, self.config['epochs'], i,
len(train_loader), loss,
end_time - start_time))
# validate the model if requested
if epoch % self.config['eval_every_n_epochs'] == 0:
start_time = time.time()
valid_loss = self._validate(model, valid_loader)
end_time = time.time()
if valid_loss < best_valid_loss:
# save the model weights
best_valid_loss = valid_loss
torch.save(
model.state_dict(),
os.path.join(model_checkpoints_folder,
'best_model.pth'))
print(
"Valid:\t Epoch[{:0>3}/{:0>3}] Iteration[{:0>3}/{:0>3}] Loss: {:.4f} Time: {:.2f}s"
.format(epoch + 1, self.config['epochs'],
len(valid_loader), len(valid_loader), valid_loss,
end_time - start_time))
# self.writer.add_scalar('validation_loss', valid_loss, global_step=valid_n_iter)
valid_n_iter += 1
print('Learning rate this epoch:',
scheduler.get_last_lr()[0]) # python >=3.7
# print('Learning rate this epoch:', scheduler.base_lrs[0]) # python 3.6
# warmup for the first 10 epochs
if epoch >= 10:
scheduler.step()
def train(self):
train_loader, valid_loader = self.dataset.get_data_loaders()
model = ResNetSimCLR(**self.config["model"]).to(self.device) #just a resnet backbone
model = self._load_pre_trained_weights(model) #checkpoints (shall we convert TF checkpoint in to Torch and train or train from the scratch since we have f*****g lot?)
optimizer = torch.optim.Adam(model.parameters(), 3e-4, weight_decay=eval(self.config['weight_decay']))
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=len(train_loader), eta_min=0,
last_epoch=-1) #learning rate shedulers (let's use as it is)
if apex_support and self.config['fp16_precision']:
model, optimizer = amp.initialize(model, optimizer,
opt_level='O2',
keep_batchnorm_fp32=True)
model_checkpoints_folder = os.path.join(self.writer.log_dir, 'checkpoints')
# save config file
_save_config_file(model_checkpoints_folder)
n_iter = 0
valid_n_iter = 0
best_valid_loss = np.inf
for epoch_counter in range(self.config['epochs']): #start training
for (x, y) in train_loader: #dataset
optimizer.zero_grad()
x = x.to(self.device) #in SimCLR we calculate the loss with two augmentation versions
y = y.to(self.device)
loss = self._step(model, x, y)
if n_iter % self.config['log_every_n_steps'] == 0:
self.writer.add_scalar('train_loss', loss, global_step=n_iter)
if apex_support and self.config['fp16_precision']:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
n_iter += 1
# validate the model if requested
if epoch_counter % self.config['eval_every_n_epochs'] == 0:
valid_loss = self._validate(model, valid_loader)
print('Epoch:',epoch_counter,' ---',' validation_loss:',valid_loss)
if valid_loss < best_valid_loss:
# save the model weights
best_valid_loss = valid_loss
torch.save(model.state_dict(), os.path.join(model_checkpoints_folder, 'model.pth'))
self.writer.add_scalar('validation_loss', valid_loss, global_step=valid_n_iter)
valid_n_iter += 1
# warmup for the first 10 epochs
if epoch_counter >= 10:
scheduler.step()
self.writer.add_scalar('cosine_lr_decay', scheduler.get_lr()[0], global_step=n_iter)
def train(self):
#Data
train_loader, valid_loader = self.dataset.get_data_loaders()
#Model
model = ResNetSimCLR(**self.config["model"])
if self.device == 'cuda':
model = nn.DataParallel(model, device_ids=[i for i in range(self.config['gpu']['gpunum'])])
#model = model.to(self.device)
model = model.cuda()
print(model)
model = self._load_pre_trained_weights(model)
each_epoch_steps = len(train_loader)
total_steps = each_epoch_steps * self.config['train']['epochs']
warmup_steps = each_epoch_steps * self.config['train']['warmup_epochs']
scaled_lr = eval(self.config['train']['lr']) * self.batch_size / 256.
optimizer = torch.optim.Adam(
model.parameters(),
scaled_lr,
weight_decay=eval(self.config['train']['weight_decay']))
'''
optimizer = LARS(params=model.parameters(),
lr=eval(self.config['train']['lr']),
momentum=self.config['train']['momentum'],
weight_decay=eval(self.config['train']['weight_decay'],
eta=0.001,
max_epoch=self.config['train']['epochs'])
'''
# scheduler during warmup stage
lambda1 = lambda epoch:epoch*1.0 / int(warmup_steps)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1)
if apex_support and self.config['train']['fp16_precision']:
model, optimizer = amp.initialize(model, optimizer,
opt_level='O2',
keep_batchnorm_fp32=True)
model_checkpoints_folder = os.path.join(self.writer.log_dir, 'checkpoints')
# save config file
_save_config_file(model_checkpoints_folder)
n_iter = 0
valid_n_iter = 0
best_valid_loss = np.inf
lr = eval(self.config['train']['lr'])
end = time.time()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
for epoch_counter in range(self.config['train']['epochs']):
model.train()
for i, ((xis, xjs), _) in enumerate(train_loader):
data_time.update(time.time() - end)
optimizer.zero_grad()
xis = xis.cuda()
xjs = xjs.cuda()
loss = self._step(model, xis, xjs, n_iter)
#print("Loss: ",loss.data.cpu())
losses.update(loss.item(), 2 * xis.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
print('Epoch: [{epoch}][{step}/{each_epoch_steps}] Loss {loss.val:.4f} Avg Loss {loss.avg:.4f} DataTime {datatime.val:.4f} BatchTime {batchtime.val:.4f} LR {lr})'.format(epoch=epoch_counter, step=i, each_epoch_steps=each_epoch_steps, loss=losses, datatime=data_time, batchtime=batch_time, lr=lr))
if n_iter % self.config['train']['log_every_n_steps'] == 0:
self.writer.add_scalar('train_loss', loss, global_step=n_iter)
if apex_support and self.config['train']['fp16_precision']:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
n_iter += 1
#adjust lr
if n_iter == warmup_steps:
# scheduler after warmup stage
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=total_steps-warmup_steps, eta_min=0, last_epoch=-1)
scheduler.step()
lr = scheduler.get_lr()[0]
self.writer.add_scalar('cosine_lr_decay', scheduler.get_lr()[0], global_step=n_iter)
sys.stdout.flush()
# validate the model if requested
if epoch_counter % self.config['train']['eval_every_n_epochs'] == 0:
valid_loss = self._validate(model, valid_loader)
if valid_loss < best_valid_loss:
# save the model weights
best_valid_loss = valid_loss
torch.save(model.state_dict(), os.path.join(model_checkpoints_folder, 'model.pth'))
self.writer.add_scalar('validation_loss', valid_loss, global_step=valid_n_iter)
valid_n_iter += 1
def train(self, callback=lambda m, e, l: None):
train_loader, valid_loader = self.dataset.get_data_loaders()
model = ResNetSimCLR(**self.config["model"]).to(self.device)
model = self._load_pre_trained_weights(model)
optimizer = torch.optim.Adam(model.parameters(), 3e-4, weight_decay=eval(self.config['weight_decay']))
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=len(train_loader), eta_min=0,
last_epoch=-1)
if apex_support and self.config['fp16_precision']:
model, optimizer = amp.initialize(model, optimizer,
opt_level='O2',
keep_batchnorm_fp32=True)
else:
print("No apex_support or config not fp16 precision")
model_checkpoints_folder = os.path.join(self.writer.log_dir, 'checkpoints')
# save config file
_save_config_file(model_checkpoints_folder)
n_iter = 0
valid_n_iter = 0
best_valid_loss = np.inf
eval_freq = self.config['eval_every_n_epochs']
num_epochs = self.config["epochs"]
train_len = len(train_loader)
valid_len = len(valid_loader)
loop = tqdm(total=num_epochs * train_len, position=0)
for epoch_counter in range(num_epochs):
for it, ((xis, xjs), _) in enumerate(train_loader):
optimizer.zero_grad()
xis = xis.to(self.device)
xjs = xjs.to(self.device)
loss = self._step(model, xis, xjs, n_iter)
if n_iter % self.config['log_every_n_steps'] == 0:
self.writer.add_scalar('train_loss', loss, global_step=n_iter)
if apex_support and self.config['fp16_precision']:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
n_iter += 1
loop.update(1)
loop.set_description(f"E {epoch_counter}/{num_epochs}, it: {it}/{train_len}, Loss: {loss.item()}")
# validate the model if requested
if epoch_counter % self.config['eval_every_n_epochs'] == 0:
valid_loss = self._validate(model, valid_loader)
callback(model, epoch_counter, valid_loss)
if valid_loss < best_valid_loss:
# save the model weights
best_valid_loss = valid_loss
torch.save(model.state_dict(), os.path.join(model_checkpoints_folder,
f'{self.dataset.name}-model-{epoch_counter}.pth'))
self.writer.add_scalar('validation_loss', valid_loss, global_step=valid_n_iter)
valid_n_iter += 1
# warmup for the first 10 epochs
if epoch_counter >= 10:
scheduler.step()
self.writer.add_scalar('cosine_lr_decay', scheduler.get_lr()[0], global_step=n_iter)
def train(self):
train_loader, valid_loader = self.dataset.get_data_loaders()
print(
f'The current dataset has {self.dataset.get_train_length()} items')
model = ResNetSimCLR(**self.config["model"]).to(self.device)
model = self._load_pre_trained_weights(model)
if self.device == self.cuda_name and self.config['allow_multiple_gpu']:
gpu_count = torch.cuda.device_count()
if gpu_count > 1:
print(
f'There are {gpu_count} GPUs with the current setup, so we will run on all the GPUs'
)
model = torch.nn.DataParallel(model)
optimizer = torch.optim.Adam(model.parameters(),
3e-4,
weight_decay=eval(
self.config['weight_decay']))
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, T_max=len(train_loader), eta_min=0, last_epoch=-1)
if apex_support and self.config['fp16_precision']:
model, optimizer = amp.initialize(model,
optimizer,
opt_level='O2',
keep_batchnorm_fp32=True)
model_checkpoints_folder = os.path.join(self.writer.log_dir,
'checkpoints')
# save config file
_save_config_file(model_checkpoints_folder)
n_iter = 0
valid_n_iter = 0
best_valid_loss = np.inf
for epoch_counter in range(self.config['epochs']):
t1 = time.time()
for (xis, xjs), _ in train_loader:
optimizer.zero_grad()
xis = xis.to(self.device)
xjs = xjs.to(self.device)
loss = self._step(model, xis, xjs, n_iter)
if n_iter % self.config['log_every_n_steps'] == 0:
print(
f"Epoch {epoch_counter}. Loss = {loss}. Time: {time.strftime('%c', time.localtime())}."
)
self.writer.add_scalar('train_loss',
loss,
global_step=n_iter)
if apex_support and self.config['fp16_precision']:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
n_iter += 1
# validate the model if requested
if epoch_counter % self.config['eval_every_n_epochs'] == 0:
valid_loss = self._validate(model, valid_loader)
if valid_loss < best_valid_loss:
# save the model weights
best_valid_loss = valid_loss
torch.save(
model.state_dict(),
os.path.join(model_checkpoints_folder, 'model.pth'))
time_for_epoch = int(time.time() - t1)
print(f"===\n \
Epoch {epoch_counter}. Time for previous epoch: {time_for_epoch} seconds. Time to go: {((self.config['epochs'] - epoch_counter)*time_for_epoch)/60} minutes. Validation loss: {valid_loss}. Best valid loss: {best_valid_loss}\
\n===")
self.writer.add_scalar('validation_loss',
valid_loss,
global_step=valid_n_iter)
valid_n_iter += 1
# warmup for the first 10 epochs
if epoch_counter >= 10:
scheduler.step()
self.writer.add_scalar('cosine_lr_decay',
scheduler.get_lr()[0],
global_step=n_iter)
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)