print(data_loaders)
print(data_sizes)
res_loss = dict()
res_top1_acc = dict()
res_top5_acc = dict()
num_classes = 100
num_epochs = 50
for name in ['3e-4 -> 1e-4', '3e-4 -> 3e-5', '3e-4 -> 0']:
model = SqueezeNet(num_classes=num_classes)
model.eval()
# print(model)
model = model.to(device)
criterion = SmoothLabelCritierion(label_smoothing=0.1)
optimizer = optim.Adam(model.parameters(), lr=3e-4, weight_decay=3e-5)
if name == '3e-4 -> 1e-4':
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, num_epochs - 5, eta_min=1e-4)
elif name == '3e-4 -> 3e-5':
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, num_epochs - 5, eta_min=3e-5)
else:
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, num_epochs - 5, eta_min=0)
warmup_scheduler = GradualWarmupScheduler(optimizer, multiplier=1, total_epoch=5, after_scheduler=lr_scheduler)
optimizer.zero_grad()
optimizer.step()
warmup_scheduler.step()
util.check_dir('../data/models/')
best_model, loss_dict, top1_acc_dict, top5_acc_dict = train_model(
data_loaders, data_sizes, name, model, criterion, optimizer, warmup_scheduler,
num_epochs=num_epochs, device=device)
optimizer.step()
return losses
if __name__ == '__main__':
device = util.get_device()
# device = torch.device('cpu')
data_loader = load_data()
num_classes = 100
res_dict = dict()
for weight_decay in [0, 1e-3, 1e-4, 1e-5]:
# for weight_decay in [3e-5, 1e-4, 3e-4]:
model = SqueezeNet(num_classes=num_classes)
model.eval()
# print(model)
model = model.to(device)
criterion = SmoothLabelCritierion(label_smoothing=0.1)
optimizer = optim.Adam(model.parameters(),
lr=3e-4,
weight_decay=weight_decay)
losses = find_wd(data_loader, model, criterion, optimizer, device)
res_dict[str(weight_decay)] = {'loss': losses}
print('{} done'.format(weight_decay))
util.plot_loss_lr(res_dict)
print('done')
# 加载为c, 存储于CPU
c = torch.load('a.pth', map_location=lambda storage, loc: storage)
# 加载为d, 存储于GPU0上
d = torch.load('a.pth', map_location={'cuda:1': 'cuda:0'})
# In[40]:
torch.set_default_tensor_type('torch.FloatTensor')
from torchvision.models import SqueezeNet
model = SqueezeNet()
# module的state_dict是一个字典
model.state_dict().keys()
# Module对象的保存与加载
torch.save(model.state_dict(), 'squeezenet.pth')
model.load_state_dict(torch.load('squeezenet.pth'))
optimizer = torch.optim.Adam(model.parameters(), lr=0.1)
torch.save(optimizer.state_dict(), 'optimizer.pth')
optimizer.load_state_dict(torch.load('optimizer.pth'))
all_data = dict(optimizer=optimizer.state_dict(),
model=model.state_dict(),
info=u'模型和优化器的所有参数')
torch.save(all_data, 'all.pth')
all_data = torch.load('all.pth')
all_data.keys()
print(data_sizes)
res_loss = dict()
res_top1_acc = dict()
res_top5_acc = dict()
num_classes = 100
num_epochs = 10
for name in ['lr=1e-3', 'lr=3e-4']:
model = SqueezeNet(num_classes=num_classes)
model.eval()
# print(model)
model = model.to(device)
criterion = SmoothLabelCritierion(label_smoothing=0.1)
if name == 'lr=1e-3':
optimizer = optim.Adam(model.parameters(), lr=1e-3)
else:
optimizer = optim.Adam(model.parameters(), lr=3e-4)
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(
optimizer, num_epochs)
util.check_dir('../data/models/')
best_model, loss_dict, top1_acc_dict, top5_acc_dict = train_model(
data_loaders,
data_sizes,
name,
model,
criterion,
optimizer,
lr_scheduler,
num_epochs=num_epochs,