About average_checkpoints in neural_sp/bin/eval_utils.py.

[lfgogogo]

I have reproduced the result of aishell,but when i check the weight files,i find model-avg10 is about one fifth size of the others.
I only see computing average in the average_checkpoints，how can this process decreases the size of the model,just like quantization.
I also use the distiller to make comparison，the model-avg10 is almost the same size as quantization model.

[hirofumi0810]

@lfgogogo That is because parameters in the optimizer were removed after model averaging.

[lfgogogo]

@hirofumi0810 Why do model average computing?For speeding up?But i test both,it doesn't seem to improve much,on my test dataset which consists of 30 thousands speeches,the original model.epoch-25 cost 15+ hours,while the averaged model cost 12 huors.Do you have some suggestions for speeding up the process,hiro?Thank you very much.

[hirofumi0810]

@lfgogogo If you have a lot of training data, checkpoint averaging might be ineffective. Please try to change n_average in score.sh.
But averaging is not related to speed performance at all.