def train_mnist_classification_model(args, train_loader, val_loader):
# short names
seq_len = args[Synthetic.seq_len]
n_features, hidden_size = args[AutoEncoder.n_features], args[
AutoEncoder.hidden_size]
optimizer_name, lr, momentum = args[AutoEncoder.optimizer], args[
AutoEncoder.lr], args[AutoEncoder.momentum]
# init auto encoder
ae_lstm = LstmAutoencoderMnistClassification(seq_len=seq_len,
n_features=n_features,
hidden_size=hidden_size)
# init optim, crtitertion and loaders
optimizer = auxlariy.get_optim(optim_name=optimizer_name,
ae=ae_lstm,
lr=lr,
momentum=momentum)
criterion_ce = auxlariy.get_criterion(args[AutoEncoder.criterion])
criterion_mse = auxlariy.get_criterion(Criterion.mse)
# training
train(train_loader, val_loader, criterion_ce, criterion_mse, optimizer,
ae_lstm, args)
torch.save(ae_lstm.state_dict(),
r'..\models\MNIST\lstm_ae_mnist_classification_model.pt'
) # save the best model
return ae_lstm
def test_mnist_classification_model(args, test_loader, model=None):
# short names
seq_len = args[Synthetic.seq_len]
n_features, hidden_size = args[AutoEncoder.n_features], args[
AutoEncoder.hidden_size]
# load trained model
if model is None:
model = LstmAutoencoderMnistClassification(seq_len=seq_len,
n_features=n_features,
hidden_size=hidden_size)
model.load_state_dict(
torch.load(
r'..\models\MNIST\lstm_ae_mnist_classification_model.pt'))
# get criterion
criterion_ce = auxlariy.get_criterion(args[AutoEncoder.criterion])
criterion_mse = auxlariy.get_criterion(Criterion.mse)
# test
loss = test(criterion_ce=criterion_ce,
criterion_mse=criterion_mse,
test_loader=test_loader,
model=model)
# print('test loss: %.3f' % loss)
return loss
def train_toy_model(args, train_loader, val_loader):
# short names
seq_range, seq_len, num_samples = args[Synthetic.seq_range], args[
Synthetic.seq_len], args[Synthetic.num_samples]
n_features, hidden_size = args[AutoEncoder.n_features], args[
AutoEncoder.hidden_size]
optimizer_name, lr, momentum = args[AutoEncoder.optimizer], args[
AutoEncoder.lr], args[AutoEncoder.momentum]
# init auto encoder
ae_lstm = LstmAutoencoderToy(seq_len=seq_len,
n_features=n_features,
hidden_size=hidden_size)
# init optim, crtitertion and loaders
optimizer = auxlariy.get_optim(optim_name=optimizer_name,
ae=ae_lstm,
lr=lr,
momentum=momentum)
criterion = auxlariy.get_criterion(args[AutoEncoder.criterion])
# training
train(train_loader, val_loader, criterion, optimizer, ae_lstm, args)
return ae_lstm
def train_snp500_model(args, train_loader, val_loader, ae):
# short names
optimizer_name, lr, momentum = args[AutoEncoder.optimizer], args[
AutoEncoder.lr], args[AutoEncoder.momentum]
# init optim, crtitertion and loaders
optimizer = auxlariy.get_optim(optim_name=optimizer_name,
ae=ae,
lr=lr,
momentum=momentum)
criterion = auxlariy.get_criterion(args[AutoEncoder.criterion])
# training
train(train_loader, val_loader, criterion, optimizer, ae, args)
return ae
def test_snp500_model(args, test_loader, model=None):
# short names
seq_len = args[Synthetic.seq_len]
n_features, hidden_size = args[AutoEncoder.n_features], args[
AutoEncoder.hidden_size]
# # load trained model
if model is None:
model = LstmAutoencoderSnp500FC(seq_len=seq_len,
n_features=n_features,
hidden_size=hidden_size)
model.load_state_dict(torch.load(r'..\models\SP500\lstm_ae_snp500.pt'))
# get criterion
criterion = auxlariy.get_criterion(args[AutoEncoder.criterion])
# test
loss = test(criterion=criterion, test_loader=test_loader, model=model)
# print('test loss: %.3f' % loss)
return loss