search

brandontrabucco / da-fusion

Effective Data Augmentation With Diffusion Models
MIT License
221 stars 18 forks source link

Where in the code is randomized intensities? #19

Open bansh123 opened 9 months ago

bansh123 commented 9 months ago

I couldn't find it in the code. It seems like just setting t_0=0.5.

brandontrabucco commented 9 months ago

Hello,

The code for randomizing the intensity strength is in two parts: https://github.com/brandontrabucco/da-fusion/blob/dbcdc9c6aeb0a25aa1cb2bdddc7104be2ee63bf4/train_classifier.py#L113-L117

This first part defines multiple augmentations with different intensities to sample from. The following second part samples one of the intensities: https://github.com/brandontrabucco/da-fusion/blob/dbcdc9c6aeb0a25aa1cb2bdddc7104be2ee63bf4/semantic_aug/augmentations/compose.py#L49-L56

By setting multiple values for --strength when calling train_classifier.py, such as as done here: https://github.com/brandontrabucco/da-fusion/blob/dbcdc9c6aeb0a25aa1cb2bdddc7104be2ee63bf4/scripts/stacking/launch_textual_inversion%3D1.0-0.75-0.5-0.25_coco.sh#L22-L26

the intensity will be sampled uniformly from a list given via the command line

bansh123 commented 9 months ago

Thanks! I have one more question. Can you provide instruction to obtain tokens from textual inversion for reproducing the results in figure 5? ( Or download link for learned tokens )