Esempio n. 1
0
def main():

    args = parse_args()

    with open(args.config_file) as fp:
        config = yaml.safe_load(fp)

    input_timesteps = config["input_timesteps"]
    output_timesteps = config["output_timesteps"]
    batch_size = config["batch_size"]
    max_epochs = config["max_epochs"]
    patience = config["patience"]

    train_dataset = [(load_wave(_[0]).reshape(-1, 1),
                      load_wave(_[1]).reshape(-1, 1))
                     for _ in config["train_data"]]
    train_dataflow = flow(train_dataset, input_timesteps, batch_size)

    val_dataset = [(load_wave(_[0]).reshape(-1,
                                            1), load_wave(_[1]).reshape(-1, 1))
                   for _ in config["val_data"]]
    val_dataflow = flow(val_dataset, input_timesteps, batch_size)

    model = build_model(input_timesteps)
    model.compile(loss=LossFunc(output_timesteps), optimizer="adam")

    train(model, train_dataflow, val_dataflow, max_epochs, patience)
Esempio n. 2
0
def main():

    args = parse_args()

    with open(args.config_file) as fp:
        config = yaml.safe_load(fp)

    input_timesteps = config["input_timesteps"]
    output_timesteps = config["output_timesteps"]
    batch_size = config["batch_size"]

    data = load_wave(args.input_file)

    # padding and rounded up to the batch multiple
    block_size = output_timesteps * batch_size
    prepad = input_timesteps - output_timesteps
    postpad = len(data) % block_size
    padded = np.concatenate(
        (np.zeros(prepad, np.float32), data, np.zeros(postpad, np.float32)))
    x = sliding_window(padded, input_timesteps, output_timesteps)
    x = x[:, :, np.newaxis]

    model = load_model(args.model_file,
                       custom_objects={"LossFunc": LossFunc(output_timesteps)})

    y = model.predict(x, batch_size=batch_size)
    y = y[:, -output_timesteps:, :].reshape(-1)[:len(data)]
    save_wave(y, args.output_file)
Esempio n. 3
0
def main():

    args = parse_args()

    with open(args.config_file) as fp:
        config = yaml.safe_load(fp)

    input_timesteps = config["input_timesteps"]
    output_timesteps = config["output_timesteps"]
    batch_size = config["batch_size"]

    data = load_wave(args.input_file)
    print("data.shape is:", data.shape)
    print("data.len is:", len(data))
    """
    from nnabla.ext_utils import get_extension_context
    cuda_device_id = 0
    ctx = get_extension_context('cudnn', device_id=cuda_device_id)
    print("Context: {}".format(ctx))
    nn.set_default_context(ctx)  # Set CUDA as a default context.
    """

    # padding and rounded up to the batch multiple
    block_size = output_timesteps * batch_size
    prepad = input_timesteps - output_timesteps
    postpad = len(data) % block_size
    print("postpad", block_size - postpad)
    padded = np.concatenate((np.zeros(prepad, np.float32), data,
                             np.zeros(block_size - postpad, np.float32)))
    x = sliding_window(padded, input_timesteps, output_timesteps)
    x = x[:, :, np.newaxis]
    y = np.zeros_like(x)

    batchlen = x.shape[0]
    print("x.length is:", batchlen)

    xx = nn.Variable((batch_size, input_timesteps, 1))

    nn.load_parameters("best_result.h5")

    print("xx.shape is:", xx.shape)

    yy = build_model(xx)

    print("yy.shape is:", yy.shape)

    print("x.shape in the loop is:", x[32:32 + batch_size, :, :].shape)

    start1 = time()

    for step in range(0, batchlen, batch_size):

        xx.d = x[step:step + batch_size, :, :]

        yy.forward()

        y[step:step + batch_size, :, :] = yy.d

        proc_time = time() - start1
        print(proc_time)
        print(step)

    y = y[:, -output_timesteps:, :].reshape(-1)[:len(data)]
    save_wave(y, args.output_file)

    print("finished\n")
    proc_time = time() - start1
    print(proc_time)
Esempio n. 4
0
def main():

    args = parse_args()

    with open(args.config_file) as fp:
        config = yaml.safe_load(fp)

    input_timesteps = config["input_timesteps"]
    output_timesteps = config["output_timesteps"]
    batch_size = config["batch_size"]
    max_epochs = config["max_epochs"]
    patience = config["patience"]

    train_dataset = [(load_wave(_[0]).reshape(-1, 1),
                      load_wave(_[1]).reshape(-1, 1))
                     for _ in config["train_data"]]
    train_dataflow = flow(train_dataset, input_timesteps, batch_size)

    val_dataset = [(load_wave(_[0]).reshape(-1,
                                            1), load_wave(_[1]).reshape(-1, 1))
                   for _ in config["val_data"]]
    val_dataflow = flow(val_dataset, input_timesteps, batch_size)
    """
    from nnabla.ext_utils import get_extension_context
    cuda_device_id = 0
    ctx = get_extension_context('cudnn', device_id=cuda_device_id)
    print("Context: {}".format(ctx))
    nn.set_default_context(ctx)  # Set CUDA as a default context.
    """

    timestamp = datetime.datetime.now()

    # For tensorboard
    tb_log_dir = "tensorboard/{:%Y%m%d_%H%M%S}".format(timestamp)
    if not os.path.exists(tb_log_dir):
        os.makedirs(tb_log_dir)
    tb_writer = SummaryWriter(tb_log_dir)

    x = nn.Variable((batch_size, input_timesteps, 1), need_grad=True)
    y = nn.Variable((batch_size, input_timesteps, 1), need_grad=True)

    vx = nn.Variable((batch_size, input_timesteps, 1), need_grad=True)
    vy = nn.Variable((batch_size, input_timesteps, 1), need_grad=True)

    t = build_model(x)
    vt = build_model(vx)

    loss = F.mean(
        F.squared_error(t[:, -output_timesteps:, :], y[:,
                                                       -output_timesteps:, :]))
    solver = S.AdaBound()
    solver.set_parameters(nn.get_parameters(grad_only=False))

    tloop = tqdm.trange(1, max_epochs + 1)
    es = EarlyStopping(patience=patience)

    steps = 100
    validation_time = 10

    cp_dir = "checkpoint/{:%Y%m%d_%H%M%S}".format(timestamp)
    if not os.path.exists(cp_dir):
        os.makedirs(cp_dir)

    idx = 0

    for i in tloop:

        # TRAINING

        tloop.set_description("Training Epoch")
        st = tqdm.trange(0, steps)
        losses = AverageMeter()
        for j in st:

            x.d, y.d = train_dataflow.__next__()
            st.set_description("Training Steps")
            loss.forward()
            solver.zero_grad()
            loss.backward()
            solver.update()
            losses.update(loss.d.copy().mean())
            st.set_postfix(train_loss=losses.avg)

            # write train graph
            tb_writer.add_scalar('train/loss', losses.avg, global_step=idx)
            idx += 1

        # VALIDATION
        vlosses = AverageMeter()
        for j in range(validation_time):
            vx.d, vy.d = val_dataflow.__next__()

            vloss = F.mean(
                F.squared_error(vt[:, -output_timesteps:, :],
                                vy[:, -output_timesteps:, :]))
            vloss.forward(clear_buffer=True)
            vlosses.update(vloss.d.copy().mean())

        tloop.set_postfix(train_loss=losses.avg, val_loss=vlosses.avg)

        tb_writer.add_scalar('test/loss', vlosses.avg, global_step=i)

        stop = es.step(vlosses.avg)
        is_best = vlosses.avg == es.best

        # save current model
        nn.save_parameters(os.path.join(cp_dir, 'checkpoint_{}.h5'.format(i)))

        if is_best:
            nn.save_parameters(os.path.join(cp_dir, 'best_result.h5'))
            """
            contents = {
                'networks': [
                    {'name': 'MyChain',
                    'batch_size': batch_size,
                    'outputs': {'t': t },
                    'names': {'x': x}}],
                'executors': [
                    {'name': 'runtime',
                    'network': 'MyChain',
                    'data': ['x'],
                    'output': ['t']}]}
            nnabla.utils.save.save(os.path.join(cp_dir ,'MyChain.nnp'), contents)
            """
        if stop:
            print("Apply Early Stopping")
            break