if val_accuracy > best_accuracy:
best_accuracy = val_accuracy
best_epoch = epoch
if model_path is not None:
model.save_weights(model_path + '.npz')
cPickle.dump(model, open(model_path + '.pkl', 'w'))
print(
'epoch={epoch:05d}, iteration={iteration:05d}, loss={loss:.04f}, val_loss={val_loss:.04f}, val_acc={val_acc:.04f} best=[accuracy={best_accuracy:.04f} epoch={best_epoch:05d}]'
.format(epoch=epoch,
iteration=iteration,
loss=train_loss,
val_loss=val_loss,
val_acc=val_accuracy,
best_accuracy=best_accuracy,
best_epoch=best_epoch))
iteration += 1
if iteration % len(train_files) == 0:
epoch += 1
x_train, y_train = load_model_data(next(train_files_iter),
args.data_name,
args.target_name,
n=args.n_train)
if __name__ == '__main__':
parser = modeling.parser.build_lasagne()
sys.exit(main(parser.parse_args()))
x_validation, y_validation_one_hot = preprocessor.transform(
x_validation, y_validation_one_hot)
if isinstance(net, keras.models.Graph):
train_data = marshaller.marshal(
x_train, y_train_one_hot)
validation_data = marshaller.marshal(
x_validation, y_validation_one_hot)
net.fit(train_data,
shuffle=args.shuffle,
nb_epoch=args.n_epochs,
batch_size=model_cfg.batch_size,
validation_data=validation_data,
callbacks=callbacks,
class_weight=class_weight,
verbose=2 if args.log else 1)
else:
net.fit(x_train, y_train_one_hot,
shuffle=args.shuffle,
nb_epoch=args.n_epochs,
batch_size=model_cfg.batch_size,
show_accuracy=True,
validation_data=(x_validation, y_validation_one_hot),
callbacks=callbacks,
class_weight=class_weight,
verbose=2 if args.log else 1)
if __name__ == '__main__':
parser = modeling.parser.build_keras()
sys.exit(main(parser.parse_args()))