/diffssl

Code for the paper "Do text-free diffusion models learn discriminative visual representations?"

Primary LanguagePythonMIT LicenseMIT

Diffusion SSL (diffssl)

Code for the paper "Do text-free diffusion models learn discriminative visual representations?".

(This work supersedes "Diffusion Models Beat GANs on Image Classification".)

Unconditional Diffusion Models as Self-Supervised Representation Learners.

Setting up the environment

  1. Clone this repository and navigate to it in your terminal.
  2. install python==3.9
  3. pip install -e .
    This should install the guided_diffusion python package that the scripts depend on.
  4. pip install -r requirements.txt

Running the scripts

  • Download this checkpoint from the original guided-diffusion repo and place it in checkpoints/256x256_diffusion_uncond.pt.

  • The scripts for different modes are located in the scripts directory. The scripts can be run using the following commands. Make sure to specify TRAIN_DATA and VAL_DATA in the scripts before running.

    • GD(L): bash scripts/linear.sh
    • Attention head: bash scripts/attention.sh
    • DifFormer: bash scripts/fusion.sh
    • DifFeed: bash scripts/feedback.sh

For more information on the arguments, refer to the finetune.py:create_argparser().