How to reproduce the results of LSTM on AN4

[Jia-zb]

I was interested in your work on PPoPP'22, thank you for making the code open source. I tried to run the LSTM AN4 code, but cannot achieve the results claimed in the paper(WER=0.309 or 0.368)，only reach 0.46. I know I'm using a different environment, Perhaps you can give me some suggestions to improve the WER?

Here is the environment I use:

• 8*A100 within one server
• Horovod 0.22.1
• Here are the parameters I use: horovodrun -np 8 python horovod_trainer.py --dnn lstman4 --dataset an4 --max-epochs 1000 --batch-size 2 --nworkers 8 --data-dir ./audio_data --lr 0.001 --nwpernode 8 --nsteps-update 1

In addition, I also adjusted learning rate decay rate(In dl_trainer.py/_Adjust_Learning_Rate_LSTMan4() ). The original 1.01 May not be suitable for my environment, so I changed it to 1.005.

Thank you for seeing this, do you have any suggestions?

[Shigangli]

Hi, could you try to use batch-size=8 instead of 2? I used global batch size = 64.

[Jia-zb]

Thank you for your reply, I tried Batchsize 4 and 8 respectively, but the results were even worse(the former was 0.73 and the latter was 0.92). I don't know why, but it seems to prefer the smaller batchsize.