-
Notifications
You must be signed in to change notification settings - Fork 17
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Added batch image generation and extended logging. #35
Conversation
Hi, thank you for your efforts in enhancing this repository! TL;DR - We will get back to you after reviewing the functionality, generation score, GPU memory usage, and inference time. Appreciate your patience.
It looks great and seems working properly. We'll verify its functionality and reach out to you.
We’d like to clarify that our |
@bokyeong1015 Thanks for your response. Surely, reproducibility is crucial. I hope you'll find these modifications beneficial. |
Hi, thanks again for your PR and sorry for the late response. Before proceeding, would you kindly create separate PRs with two distinct branches? If this is OK for you, please let us know. We'll then set up the corresponding issues and branches and get back to you.
Reviewing improved-wandb-logger (
|
Eval Batch Size | 1 | 4 | 8 |
---|---|---|---|
GPU Memory | 4978MiB (4.9GB) | 11590MiB (11.3GB) | 20034MiB (19.6GB) |
Generation Time | 34130s (9.5h) | 27564s (7.7h) | 26313s (7.3h) |
FID | 16.98 | 17.16 | 16.97 |
IS | 31.68 | 31.62 | 31.22 |
CLIP Score | 0.2677 | 0.2677 | 0.2675 |
@bokyeong1015 glad to hear you found these modifications helpful. Yeah I agree, that two pull requests for two separate features are better for maintaining the structure of the project, and I'll take into account your comments and points in the new pull requests. |
Thanks for your kind reply! Could you please create two separate PRs for the branches Either of the two options below is fine, so please decide at your convenience. Thank you for your time and help.
|
Hi, thanks for your work. The results are quite impressive. Below I added a few modifications that can be helpful for experimentation.
The original
src/generate.py
generates images one-by-one which leads to under utilization of GPU and as consequence, generation of 30k images takes a while. I've added batched generation of images to speed-up it.In addition, the base training script
src/kd_train_text_to_image.py
logs only the total loss to W&B and one may be interested in each particular contribution. I added image logging to W&B as well.