if (epoch + 1) % args.print_freq == 0:
for j in range(len(best_acc)):
print(
"epoch:{} model:{} train loss:{:.2f} acc:{:.2f} val loss{:.2f} acc:{:.2f}"
.format(epoch + 1, args.model_names[j], train_losses[j],
train_acces[j], val_losses[j], val_acces[j]))
for k in range(len(best_acc)):
print("model:{} best acc:{:.2f}".format(args.model_names[k],
best_acc[k]))
if __name__ == '__main__':
train_loader, val_loader = get_cifar100_dataloaders(
root=args.root,
batch_size=args.batch_size,
num_workers=args.num_workers,
is_instance=False)
model_list = []
optimizer_list = []
scheduler_list = []
for name in args.model_names:
lr = 0.01 if name in ['MobileNetV2', 'ShuffleV1', 'ShuffleV2'
] else args.lr
model = model_dict[name](num_classes=100)
if torch.cuda.is_available(): model = model.cuda()
optimizer = optim.SGD(model.parameters(),
lr=lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_id)
if args.s_arch in ['MobileNetV2', 'ShuffleV1', 'ShuffleV2']:
args.lr = 0.01
exp_name = f'CAM/nodistill_mag_S_{args.s_arch}'
exp_path = './experiments/{}'.format(exp_name)
os.makedirs(exp_path, exist_ok=True)
print(f"CAM no distillation MAG Student:{args.s_arch} [{exp_path}]")
logger = SummaryWriter(osp.join(exp_path, 'events'), flush_secs=10)
train_loader, val_loader, n_data = get_cifar100_dataloaders(
root=args.root,
batch_size=args.batch_size,
num_workers=4,
is_instance=True)
args.n_data = n_data
# student model definition
s_model = model_dict[args.s_arch](num_classes=100)
s_model = wrapper(module=s_model, cfg=args).cuda()
# ---------------- start distillation ! -------------------
print("-------------start distillation ! -------------")
# construct kd loss and optimizer
s_optimizer = optim.SGD(chain(s_model.parameters()),
lr=args.lr,