BookCorpus training not generating meaningful sentences.

[shikunyu8]

Hi, thank you for sharing this implementation. I have several questions about your code.

1. Did you train your model on the BookCorpus dataset? Can the model generate meaningful previous and next sentence given an input sentence?

2. I went through your implementation and found that you are using LSTM instead of GRU model, and the skip thoughts vector is 1200 dimensional instead 2400 dimensional. Are there other differences to the original paper?

3. I've trained this model for roughly a day, but the loss does not decrease much and the model is generating garbage. Could you share your training strategy that can let the model generate meaningful sentences?

Thanks!