ShenXiaolei comments

Results 5 comments of


                                            ShenXiaolei

Proposal - implement MaskDiT technique for fast training

training with mae strategy, there will be one more decoder during reasoning. it is a serious shortcoming

When running CIFAR10 demo, NAN appears in loss.

> I got the same problem when enable modulate kernel. me too

Doesn't converge when I train with my own data

How did you build your dataset?

pipeline at dev branch

![IMG_0426](https://github.com/user-attachments/assets/eca0d3be-740b-469d-a03f-f3272e679915) input and output images @zengyh1900

> ![IMG_0426](https://private-user-images.githubusercontent.com/31148634/371379647-eca0d3be-740b-469d-a03f-f3272e679915.jpeg?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3Mjc0MzUwNzksIm5iZiI6MTcyNzQzNDc3OSwicGF0aCI6Ii8zMTE0ODYzNC8zNzEzNzk2NDctZWNhMGQzYmUtNzQwYi00NjlkLWEwM2YtZjMyNzJlNjc5OTE1LmpwZWc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwOTI3JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDkyN1QxMDU5MzlaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT0wYTdiNDQxMGI1MDBhYjc1MmNjZTVlYzVmMDg3MDUyNTNmYmIzMDM1MjI0OTZjMGI4NmUzODJiNGEzZWI1Nzg4JlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCJ9.t3Tvenf1qAGHCUm2-Yc5sKIOB4eWFKQmEq8iepbMFns) input and output images @zengyh1900 differences between dev and main branch 1. dev: conditioning_latents = torch.concat([mask, conditioning_latents], 1) main: conditioning_latents = torch.concat([conditioning_latents, mask], 1) 2. dev: original_mask =...