def main(opt):
if torch.cuda.is_available():
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
train_params = {
"batch_size": opt.batch_size,
"shuffle": True,
"drop_last": False,
"num_workers": opt.num_workers,
"collate_fn": collate_fn
}
eval_params = {
"batch_size": opt.batch_size,
"shuffle": True,
"drop_last": False,
"num_workers": opt.num_workers,
"collate_fn": collate_fn
}
dboxes = generate_dboxes()
model = SSD()
train_set = OIDataset(SimpleTransformer(dboxes), train=True)
train_loader = DataLoader(train_set, **train_params)
val_set = OIDataset(SimpleTransformer(dboxes, eval=True), validation=True)
val_loader = DataLoader(val_set, **eval_params)
encoder = Encoder(dboxes)
opt.lr = opt.lr * (opt.batch_size / 32)
criterion = Loss(dboxes)
optimizer = torch.optim.SGD(model.parameters(),
lr=opt.lr,
momentum=opt.momentum,
weight_decay=opt.weight_decay,
nesterov=True)
scheduler = MultiStepLR(optimizer=optimizer,
milestones=opt.multistep,
gamma=0.1)
if torch.cuda.is_available():
model.cuda()
criterion.cuda()
model = torch.nn.DataParallel(model)
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
if not os.path.isdir(opt.save_folder):
os.makedirs(opt.save_folder)
checkpoint_path = os.path.join(opt.save_folder, "SSD.pth")
writer = SummaryWriter(opt.log_path)
if os.path.isfile(checkpoint_path):
checkpoint = torch.load(checkpoint_path)
first_epoch = checkpoint["epoch"] + 1
model.module.load_state_dict(checkpoint["model_state_dict"])
scheduler.load_state_dict(checkpoint["scheduler"])
optimizer.load_state_dict(checkpoint["optimizer"])
else:
first_epoch = 0
for epoch in range(first_epoch, opt.epochs):
train(model, train_loader, epoch, writer, criterion, optimizer,
scheduler)
evaluate(model, val_loader, encoder, opt.nms_threshold)
checkpoint = {
"epoch": epoch,
"model_state_dict": model.module.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict()
}
torch.save(checkpoint, checkpoint_path)
def main(opt):
if torch.cuda.is_available():
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
num_gpus = torch.distributed.get_world_size()
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
num_gpus = 1
train_params = {
"batch_size": opt.batch_size * num_gpus,
"shuffle": True,
"drop_last": False,
"num_workers": opt.num_workers,
"collate_fn": collate_fn
}
test_params = {
"batch_size": opt.batch_size * num_gpus,
"shuffle": False,
"drop_last": False,
"num_workers": opt.num_workers,
"collate_fn": collate_fn
}
if opt.model == "ssd":
dboxes = generate_dboxes(model="ssd")
model = SSD(backbone=ResNet(), num_classes=len(coco_classes))
else:
dboxes = generate_dboxes(model="ssdlite")
model = SSDLite(backbone=MobileNetV2(), num_classes=len(coco_classes))
train_set = CocoDataset(opt.data_path, 2017, "train",
SSDTransformer(dboxes, (300, 300), val=False))
train_loader = DataLoader(train_set, **train_params)
test_set = CocoDataset(opt.data_path, 2017, "val",
SSDTransformer(dboxes, (300, 300), val=True))
test_loader = DataLoader(test_set, **test_params)
encoder = Encoder(dboxes)
opt.lr = opt.lr * num_gpus * (opt.batch_size / 32)
criterion = Loss(dboxes)
optimizer = torch.optim.SGD(model.parameters(),
lr=opt.lr,
momentum=opt.momentum,
weight_decay=opt.weight_decay,
nesterov=True)
scheduler = MultiStepLR(optimizer=optimizer,
milestones=opt.multistep,
gamma=0.1)
if torch.cuda.is_available():
model.cuda()
criterion.cuda()
if opt.amp:
from apex import amp
from apex.parallel import DistributedDataParallel as DDP
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
else:
from torch.nn.parallel import DistributedDataParallel as DDP
# It is recommended to use DistributedDataParallel, instead of DataParallel
# to do multi-GPU training, even if there is only a single node.
model = DDP(model)
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
if not os.path.isdir(opt.save_folder):
os.makedirs(opt.save_folder)
checkpoint_path = os.path.join(opt.save_folder, "SSD.pth")
writer = SummaryWriter(opt.log_path)
if os.path.isfile(checkpoint_path):
checkpoint = torch.load(checkpoint_path)
first_epoch = checkpoint["epoch"] + 1
model.module.load_state_dict(checkpoint["model_state_dict"])
scheduler.load_state_dict(checkpoint["scheduler"])
optimizer.load_state_dict(checkpoint["optimizer"])
else:
first_epoch = 0
for epoch in range(first_epoch, opt.epochs):
train(model, train_loader, epoch, writer, criterion, optimizer,
scheduler, opt.amp)
evaluate(model, test_loader, epoch, writer, encoder, opt.nms_threshold)
checkpoint = {
"epoch": epoch,
"model_state_dict": model.module.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict()
}
torch.save(checkpoint, checkpoint_path)
def main(agrs):
start_epoch = 0
# Initialize model or load trained checkpoint
if args.resume:
start_epoch, model, optimizer = load_checkpoint(args.trained_model)
else:
model = SSD('train', args)
optimizer = init_optimizer(model, args)
# Move to default device and set 'train' mode
model = model.cuda()
model.train()
# Create multibox loss
criterion = MultiBoxLoss(PriorBox().forward().cuda(),
args.overlap_threshold, args.negpos_ratio,
args.alpha)
# VOC dataloaders
train_dataset = VOCxx('train',
args.dataroot,
args.datayears,
args.datanames,
discard_difficult=args.discard_difficult,
use_augment=args.use_augment)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
collate_fn=train_dataset.collate_fn,
num_workers=1,
pin_memory=True)
# Loop and decay params
epochs = args.iterations // (len(train_dataset) // args.batch_size)
decay_iters = [int(it) for it in args.lr_decay.split(',')]
decay_lr_at = [
it // (len(train_dataset) // args.batch_size) for it in decay_iters
]
print('total length of dataset : ', len(train_dataset))
print('total epochs : ', epochs)
print('decay lr at : ', decay_lr_at)
# Epochs
loc_losses, conf_losses = [], []
for epoch in range(start_epoch, epochs):
# Decay learning rate at particular epochs
if epoch in decay_lr_at:
optimizer = adjust_lr(optimizer)
for i, (images, targets) in enumerate(train_loader):
# Move to default device
images = images.cuda()
targets = [t.cuda() for t in targets]
# Forward prop
preds = model(images)
# Loss
loc_loss, conf_loss = criterion(preds, targets)
loss = loc_loss + conf_loss
# Backward prop
optimizer.zero_grad()
loss.backward()
# Clip gradients if necessary
if args.clip_grad:
clip_gradient(model.parameters(), args.clip_grad)
# Update model
optimizer.step()
# Print status
if i % 200 == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Loss : {loss:.4f}\t'.format(epoch,
i,
len(train_loader),
loss=loss.item()))
loc_losses.append(loc_loss.item())
conf_losses.append(conf_loss.item())
# Plot losses
plot_losses(loc_losses, 'regression', args.model_save_name)
plot_losses(conf_losses, 'classification', args.model_save_name)
# Save checkpoint
save_checkpoint(epoch, model, optimizer, args.model_save_name)