DAI-Lab/RivaGAN

Poor performance

Closed this issue · 3 comments

llbbcc commented
  • Python version: 3.9
  • Operating System: Linux

Description

from rivagan import RivaGAN

model = RivaGAN()
model.fit('./data/hollywood2', epochs=300, batch_size=12, use_critic=True, use_adversary=True)
model.save('./checkpoints/model.pt')

When I use this script for training, I get a few warnings as follows,

[mp3float @ 0xe49f1240] Header missing
[mp3float @ 0xe49f1240] Header missing
[mp3float @ 0xe4c01e00] Header missing

I don't know if this affects the model.
My model accuracy has been stuck at about 54%.

What I use is Hollywood2 dataset and it has been processed for successful loading.

Has anyone else experienced the same problem, or has anyone reproduced the results of the paper.

langyo commented

It's audio data and wouldn't affect the result.

llbbcc commented

It's audio data and wouldn't affect the result.

Thanks for your reply! The reason I'm having this problem is that Hollywood2 is too small for training a 64bit length model. Moments In Time is great.

langyo commented

It's audio data and wouldn't affect the result.

Thanks for your reply! The reason I'm having this problem is that Hollywood2 is too small for training a 64bit length model. Moments In Time is great.

If you need, here are some pre-trained onnx models inside the demo: https://github.com/ShieldMnt/invisible-watermark/tree/main/imwatermark