[Help]: Is there any loss that linearly correlate to performance of TTA autoencoder?

[Jiang-Stan]

Problem Overview

(Briefly and clearly describe the issue you're facing and seeking help with.)
I tried to reproduce the TTA recipe, but it seems that it is hard to assess performance through any loss.

Thanks in advance for your sicerely help!

[Jiang-Stan]

Can you share training detail of TTA autoencoderKL? I tried to reproduce the result with the origin repo with your preprocessed mel data and training configs, but I failed to get result similar to opensource ckpt. Or is there any way determining performance of ckpts of different iter?

[HeCheng0625]

Hi, the detail can be found in https://github.com/open-mmlab/Amphion/blob/main/egs/tta/autoencoderkl/exp_config_latent_4_10_78.json

[fpicetti]

So the rest of the parameters are inherited from config/autoencoderkl.json, right?