X = X.to(device)
Y = Y.to(device)
logits = model(X)
preds = torch.max(logits, 1)[1].view(Y.size())
loss = criterion(logits, Y)
total_epoch_loss += loss.item()
total_epoch_correct += (preds == Y).float().sum().item()
total_count += len(Y)
return total_epoch_loss / len(_iter), total_epoch_correct / total_count
if __name__ == "__main__":
embedding, train_iter, valid_iter, test_iter = get_dataset(
params.dataset, rate=params.noise_ratio, batch_size=params.batch_size)
model_params = {}
if params.dataset == "IMDB":
model_params["output_size"] = 2
model_params["hidden_size"] = 256
if params.model == "LSTM" or params.model == "LSTMATT":
model_params["weights"] = embedding
if params.dataset == "CIFAR100":
model_params["num_classes"] = 100
model = get_model(params.model, **model_params)
if os.path.exists(save_path) and params.load:
print("model loaded")
Y = Y.reshape(-1, 1)
top1_num_corrects = (preds1 == Y).float().sum()
top5_num_corrects = (preds5 == Y.expand_as(preds5)).float().sum()
total_epoch_loss += loss.item()
total_epoch_acc1 += top1_num_corrects.item()
total_epoch_acc5 += top5_num_corrects.item()
total_counter += len(Y)
return (total_epoch_loss / len(_iter),
total_epoch_acc1 / total_counter * 100,
total_epoch_acc5 / total_counter * 100)
if __name__ == "__main__":
train_iter, test_iter = get_dataset("TinyImageNet", rate=params.noise_ratio, batch_size=params.batch_size)
model = get_model(params.model)
if params.load_from:
model.load_state_dict(torch.load(params.load_from))
model.to(device)
optim = Vadam2(model.parameters(), lr=params.lr * params.batch_size / 128, eps=1e-15,
v=params.v, alpha=params.alpha, auto_v=params.auto,
weight_decay=params.weight_decay)
imgsize = EfficientNet.get_image_size(params.model)
top1acc, top5acc = [], []
for epoch in range(params.num_epochs):
train_loss, train_acc = train_model(model, train_iter, size=imgsize)
test_loss, test_acc1, test_acc5 = eval_model(model, test_iter, size=imgsize)
top1acc.append(test_acc1)
top5acc.append(test_acc5)