search

breizhn / DTLN

Tensorflow 2.x implementation of the DTLN real time speech denoising model. With TF-lite, ONNX and real-time audio processing support.
MIT License
567 stars 160 forks source link

TF_DETERMINISTIC_OPS #54

Closed StuartIanNaylor closed 2 years ago

StuartIanNaylor commented 2 years ago

run_training.py

# use the GPU with idx 0
os.environ["CUDA_VISIBLE_DEVICES"]='0'
# activate this for some reproducibility
os.environ['TF_DETERMINISTIC_OPS'] = '1'

TF_DETERMINISTIC_OPS means the model should be reproducible with the same loss, accuracy... It does slow down the training though and just wondering is there any reason why I can not just remark that out if I want as my question is that it has no other effects?

breizhn commented 2 years ago

Yes exactly, it just makes it more reproducible. Just comment it out and it will run faster, especially in recent TF versions.

StuartIanNaylor commented 2 years ago

Thanks nils I thought so but wasn't sure it has any other effect