def predict_cnn(config):
# load tokenizer and torchtext Field
pickle_tokenizer = open('pickles/tokenizer.pickle', 'rb')
cohesion_scores = pickle.load(pickle_tokenizer)
tokenizer = LTokenizer(scores=cohesion_scores)
pickle_vocab = open('pickles/text.pickle', 'rb')
text = pickle.load(pickle_vocab)
model = CNN(config)
model.load_state_dict(torch.load(config.save_model))
model.to(device)
model.eval()
tokenized = tokenizer.tokenize(config.input)
min_len = config.filter_sizes[-1]
# if user's input sentence is shorter than the largest filter size, add pad tokens to input sentence
if len(tokenized) < min_len:
tokenized += ['<pad>'] * (min_len - len(tokenized))
indexed = [text.vocab.stoi[token] for token in tokenized]
length = [len(indexed)]
tensor = torch.LongTensor(indexed).to(device)
tensor = tensor.unsqueeze(1)
length_tensor = torch.LongTensor(length)
prediction = torch.sigmoid(model(tensor, length_tensor))
label = torch.round(prediction)
if label == 1:
label = 'Positive'
else:
label = 'Negative'
sentiment_percent = prediction.item()
print(f'[in] >> {config.input}')
print(f'[out] >> {sentiment_percent*100:.2f} % : {label}')
# for CNN model
if model_type == 'CNN':
hidden_size = 128
pool_size = 2
n_filters = 128
filter_sizes = [3, 8]
n_epochs = 30
cnn = CNN(vocab_size, embedding_size, n_filters, filter_sizes,
pool_size, hidden_size, n_classes, dropout_rate)
optimizer = optim.Adam(cnn.parameters(), lr=lr)
train_loss, train_acc, val_loss, val_acc = train_val(
n_epochs, cnn, train_iter, val_iter, optimizer, lfunc, model_type,
'two_layer', optim_type)
export_graph('cnn24_loss.png', n_epochs, train_loss, val_loss,
'Loss Across Epochs (CNN)', 'Loss')
export_graph('cnn24_acc.png', n_epochs, train_acc, val_acc,
'Network Accuracy Across Epochs (CNN)', 'Accuracy')
# test model
cnn.load_state_dict(
torch.load(
os.path.join(
path, 'CNN_' + 'two_layer' + '_' + optim_type +
'_saved_state.pt')))
test_loss, test_acc = validate(cnn, test_iter, lfunc)
print(f'Test Loss: {test_loss:.3f} | Test Acc: {test_acc * 100:.2f}%')
def main():
args = options()
use_cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
device = torch.device("cuda" if use_cuda else "cpu")
train_set = preload.datasets.MNISTDataset('../data',
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, ),
(0.5, ))
]))
train_loader = preload.dataloader.DataLoader(train_set,
batch_size=args.batch_size)
test_loader = preload.dataloader.DataLoader(
preload.datasets.MNISTDataset('../data',
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, ),
(0.5, ))
])),
batch_size=args.test_batch_size)
if args.model == 'cnn':
model = CNN().to(device)
elif args.model == 'cnn_leaky_relu':
model = CNNLeakyReLU.to(device)
else:
print("model error")
exit()
start_point = copy.deepcopy(model.state_dict())
# print("\nNormal training:")
# if args.load_model:
# model.load_state_dict(torch.load("mnist_cnn.pt"))
# else:
# model.load_state_dict(start_point)
# optimizer = optim.SGD(model.parameters(), lr=args.lr)
# scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
# normal_method = NormalTrain(model, device, train_loader, optimizer)
# model_training(args, model, normal_method, device, test_loader, scheduler)
# if args.save_model:
# torch.save(model.state_dict(), "mnist_cnn.pt")
# evaluation(args, model, device, test_loader)
# print("\nNormal training with L2 regularization:")
# if args.load_model:
# model.load_state_dict(torch.load("mnist_cnn_l2_regular.pt"))
# else:
# model.load_state_dict(start_point)
# optimizer = optim.SGD(model.parameters(), lr=args.lr)
# scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
# l2_method = L2RegularTrain(model,
# device,
# train_loader,
# optimizer,
# weight_decay=args.weight_decay)
# model_training(args, model, l2_method, device, test_loader, scheduler)
# if args.save_model:
# torch.save(model.state_dict(), "mnist_cnn_l2_regular.pt")
# evaluation(args, model, device, test_loader)
# print("\nTraining with adversarial gradient regularization:")
# if args.load_model:
# model.load_state_dict(torch.load("mnist_cnn_adv_grad_regular.pt"))
# else:
# model.load_state_dict(start_point)
# optimizer = optim.SGD(model.parameters(), lr=args.lr)
# scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
# adv_grad_reg_method = AdversarialGradientRegularTrain(model,
# device,
# train_loader,
# optimizer,
# gradient_decay=args.gradient_decay)
# model_training(args, model, adv_grad_reg_method, device, test_loader, scheduler)
# if args.save_model:
# torch.save(model.state_dict(), "mnist_cnn_adv_grad_regular.pt")
# evaluation(args, model, device, test_loader)
# print("\nAdversarial training (FGSM):")
# if args.load_model:
# model.load_state_dict(torch.load("mnist_cnn_fgsm.pt"))
# else:
# model.load_state_dict(start_point)
# optimizer = optim.SGD(model.parameters(), lr=args.lr)
# scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
# fgsm = FastGradientSignMethod(lf=F.nll_loss, eps=args.eps)
# adv_method = AdversarialTrain(model, device, train_loader, optimizer, attack=fgsm)
# model_training(args, model, adv_method, device, test_loader, scheduler)
# if args.save_model:
# torch.save(model.state_dict(), "mnist_cnn_{}.pt".format(fgsm.name))
# evaluation(args, model, device, test_loader)
# print("\nAdversarial training (BIM):")
# if args.load_model:
# model.load_state_dict(torch.load("mnist_cnn_bim.pt"))
# else:
# model.load_state_dict(start_point)
# optimizer = optim.SGD(model.parameters(), lr=args.lr)
# scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
# bim = BasicIterativeMethod(lf=F.nll_loss, eps=args.eps, alpha=args.alpha, iter_max=args.iter_max)
# adv_method = AdversarialTrain(model, device, train_loader, optimizer, attack=bim)
# model_training(args, model, adv_method, device, test_loader, scheduler)
# if args.save_model:
# torch.save(model.state_dict(), "mnist_cnn_{}.pt".format(bim.name))
# evaluation(args, model, device, test_loader)
# print("\nAdversarial training (PGD):")
# if args.load_model:
# model.load_state_dict(torch.load("mnist_cnn_pgd.pt"))
# else:
# model.load_state_dict(start_point)
# optimizer = optim.SGD(model.parameters(), lr=args.lr)
# scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
# pgd = ProjectedGradientDescent(lf=F.nll_loss, eps=args.eps, alpha=args.alpha, iter_max=args.iter_max)
# adv_method = AdversarialTrain(model, device, train_loader, optimizer, attack=pgd)
# model_training(args, model, adv_method, device, test_loader, scheduler)
# if args.save_model:
# torch.save(model.state_dict(), "mnist_cnn_{}.pt".format(pgd.name))
# evaluation(args, model, device, test_loader)
print("\nAdversarial guided training (FGSM):")
if args.load_model:
model.load_state_dict(torch.load("mnist_cnn_adv_guided.pt"))
else:
model.load_state_dict(start_point)
optimizer = optim.SGD(model.parameters(), lr=args.lr)
scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
guide_sets = make_guide_set(train_set, size=1000)
adv_guided_method = AdversarialGuidedTrain(
model,
device,
train_loader,
optimizer,
guide_sets=guide_sets,
epsilon=args.eps,
beta=args.beta,
weight_decay=args.weight_decay,
gradient_decay=args.gradient_decay)
model_training(args, model, adv_guided_method, device, test_loader,
scheduler)
if args.save_model:
torch.save(model.state_dict(), "mnist_cnn_adv_guided.pt")
evaluation(args, model, device, test_loader)
