Open machanic opened 7 years ago
Actually, this is not the way in the original paper. According to arXiv:1504.00941, initializing the weight of RNN with identity matrix may improve performance. But I think this does not affect in this case. I will delete this code later.
in train.py file: