Open vinusankars opened 5 years ago
The architecture in the paper suggests unidirectional RNNs. Is there any reason why you used bidirectional RNNs in neural_transducer.py ?
This was an arbitrary decision, without a specific reason other than bi-directional rnns are more expressive
The architecture in the paper suggests unidirectional RNNs. Is there any reason why you used bidirectional RNNs in neural_transducer.py ?