Created by: JiayiFeng

In the test_understand_sentiment_lstm.py, the training code is like this:

def main():
    word_dict = paddle.dataset.imdb.word_dict()
    cost, acc = lstm_net(dict_dim=len(word_dict), class_dim=2)

    batch_size = 100
    train_data = paddle.batch(
        paddle.reader.buffered(
            paddle.dataset.imdb.train(word_dict), size=batch_size * 10),
        batch_size=batch_size)

    data = chop_data(next(train_data()))

    place = core.CPUPlace()
    tensor_words, tensor_label = prepare_feed_data(data, place)
    exe = Executor(place)
    exe.run(framework.default_startup_program())

    while True:
        outs = exe.run(framework.default_main_program(),
                       feed={"words": tensor_words,
                             "label": tensor_label},
                       fetch_list=[cost, acc])
        cost_val = np.array(outs[0])
        acc_val = np.array(outs[1])

        print("cost=" + str(cost_val) + " acc=" + str(acc_val))
        if acc_val > 0.9:
            break

Obviously, every batch the same data are fed into the model. It might be a good way to test basic functions of lstm, but it's not enough to check its performance on the real data.