train_net.py 文件源码

def main(resume=None):
    l = 300
    dataset = './data/ubiquitous_train.hkl'
    print('Loading dataset {}...'.format(dataset))
    X_train, y_train = hkl.load(dataset)
    X_train = X_train.reshape(-1, 4, 1, l).astype(floatX)
    y_train = np.array(y_train, dtype='int32')
    indice = np.arange(X_train.shape[0])
    np.random.shuffle(indice)
    X_train = X_train[indice]
    y_train = y_train[indice]
    print('X_train shape: {}, y_train shape: {}'.format(X_train.shape, y_train.shape))

    layers = [
            (InputLayer, {'shape': (None, 4, 1, l)}),
            (Conv2DLayer, {'num_filters': 64, 'filter_size': (1, 4)}),
            (Conv2DLayer, {'num_filters': 64, 'filter_size': (1, 3)}),
            (Conv2DLayer, {'num_filters': 64, 'filter_size': (1, 3)}),
            (MaxPool2DLayer, {'pool_size': (1, 2)}),
            (Conv2DLayer, {'num_filters': 64, 'filter_size': (1, 2)}),
            (Conv2DLayer, {'num_filters': 64, 'filter_size': (1, 2)}),
            (Conv2DLayer, {'num_filters': 64, 'filter_size': (1, 2)}),
            (MaxPool2DLayer, {'pool_size': (1, 2)}),
            (DenseLayer, {'num_units': 64}),
            (DropoutLayer, {}),
            (DenseLayer, {'num_units': 64}),
            (DenseLayer, {'num_units': 2, 'nonlinearity': softmax})]

    lr = theano.shared(np.float32(1e-4))

    net = NeuralNet(
            layers=layers,
            max_epochs=100,
            update=adam,
            update_learning_rate=lr,
            train_split=TrainSplit(eval_size=0.1),
            on_epoch_finished=[
                AdjustVariable(lr, target=1e-8, half_life=20)],
            verbose=4)

    if resume != None:
        net.load_params_from(resume)

    net.fit(X_train, y_train)

    net.save_params_to('./models/net_params.pkl')