search

shiyegao / DDA

Official repository of "Back to Source: Diffusion-Driven Test-Time Adaptation"
68 stars 8 forks source link

Generated images sharing #12

Closed dineshdaultani closed 8 months ago

dineshdaultani commented 8 months ago

Hi DDA authors, Thanks for your great work. I am currently generating the adapted images using image_adapt.sh script. However, as you might know the generation process is very slow. It would take around 40+ days on my current GPU environment just to generate imagenet-c severity-5 images.

Hence, I was thinking if it would be possible for you to share the DDA generated images using any platform like google drive/one drive? It would be really helpful and save a lot of computing resources.

Please let me know about it. Will look forward to hearing from you. Thanks!

shiyegao commented 8 months ago

Thanks for your appreciation. Similar replies can be found in #1 and #6.

We used multiple GPUs to generate samples. If you hope to do some experiments using only one GPU, it may be a good idea to use a part of the dataset, such as the top 500 or 5k images, instead of 50k.

Besides, improving the generation speed can also be considered.