search

lucidrains / magvit2-pytorch

Implementation of MagViT2 Tokenizer in Pytorch
MIT License
565 stars 34 forks source link

About training steps and correctness. #36

Open becauseofAI opened 9 months ago

becauseofAI commented 9 months ago

According to the same settings in the readme, I trained 40,000 coco images.

Currently, 11,700/1_000_000 steps have been trained, but reconstruction has not been possible, as shown in the figure below. image

step 20,000 image

step 39,000 image

step 54,000 image

The reconstruction results of the first few steps are shown in the figure below. step 100 image step 200 image

The training indicator curve is shown in the figure below. image

So, is the current training normal? If it's not normal, can you help locate the problem? If it is normal, how many steps does it take to train to reconstruct the image? @lucidrains

NingMa-AI commented 8 months ago

I met the same problem, I am using 100K test images of ImageNet. The saved figures and losses behave like what you show.

ziyannchen commented 7 months ago

Same here. Anyone got solutions?

kxgong commented 6 months ago

I got better results than you (not good results), and I can see the vague shapes. I used a smaller dataset.