search

domerin0 / rnn-speech

Character level speech recognizer using ctc loss with deep rnns in TensorFlow.
MIT License
77 stars 31 forks source link

cluster size #44

Open anurag10 opened 6 years ago

anurag10 commented 6 years ago

can I know the cluster size you have used?

AMairesse commented 6 years ago

Hi, I'm not sure to understand what you mean by "cluster size". Are you talking about the number of samples used together for one step of training ? If this is the case the english model was trained with a batch size of 10 files and a mini-batch of 3. That mean that the train step (updating the network) was based on the error found on 30 files each time. I found that it was a good compromise between the stability of the training and the performance. Is that what you were asking ?

anurag10 commented 6 years ago

Thanks for your answer. Actually I wanted to know if you run it on single machine or on multiple machines?

AMairesse commented 6 years ago

Yes it's running on a single machine : training and/or inference (a good GPU is highly recommended) You could probably use a machine with multiple GPU because tensorflow would manage it but I'm not even sure it would be efficient. I mean the code would probably not make a optimal use of it. And about multiple machine I haven't look in supporting it. It could be done but would require quite a lot of changes.