tiny-imagenet

Tiny Imagenet Experiments

This code is based on openai/guided-diffusion.

Download pre-trained model

We have released checkpoint for the class conditional Tiny imagenet model.

• 64x64 class conditional DM: ema_0.9999_120000.pt

Training

To train a conditional diffusion model, first download the Tiny Imagenet dataset from here and unconditional Imagenet checkpoint from the openai/improved-diffusion repo here . Then store the dataset in a folder called tiny-imagenet-200 in the datasets folder. Then run the following command:

python scripts/image_train.py --image_size 64 --num_channels 128 --num_res_blocks 3 --learn_sigma True --diffusion_steps 1000 --noise_schedule cosine --lr 1e-4 --batch_size 128 --class_cond True --resume_checkpoint checkpoints_openai/imagenet64_uncond_100M_1500K.pt --data_dir datasets/tiny-imagenet-200/train --save_interval 20_000 --num_samples 32

To train the rejection classifier used in our experiments, run the following command after updating paths in the script:

python scripts/train_rejection_classifier.py

Sampling

For vanilla conditional sampling, run the following command:

python scripts/image_sample.py <training_args> --use_ddim True --timestep_respacing ddim100

For rejection sampling with classifier-free guidance, run:

python scripts/sample_from_rejection_classifier.py --target_dir <dir_to_save_images> --threshold <rejection_threshold>  --target_class <target_class> --guidance_str <cf_guidance_str>

The above script uses ddim sampling with 100 steps.

For vanilla classifier-free guidance without rejection sampling, run the same command with --threshold 0.0.