search

jik876 / hifi-gan

HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis
MIT License
1.92k stars 506 forks source link

MPD and MSD Two discriminators take up a lot of memory #158

Open a897456 opened 10 months ago

a897456 commented 10 months ago

image image What's going on here? How can I reduce the memory?

egaznep commented 1 month ago

@a897456 I think the parameter counts given in the paper are only for the generator, since this is what matters for inference. While the total number of parameters (and hence the model checkpoint sizes) are relatively big, the training memory footprint is primarily due to the generator.