This repository hosts the official PyTorch implementation of the paper: Ambient Diffusion: Learning Clean Distributions from Corrupted Data.
Authored by: Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alexandros G. Dimakis, Adam Klivans
Abstract : We present the first diffusion-based framework that can learn an unknown distribution using only highly-corrupted samples. This problem arises in scientific applications where access to uncorrupted samples is impossible or expensive to acquire. Another benefit of our approach is the ability to train generative models that do not memorize any individual training sample, since they never observe clean training data. Our main idea is to introduce additional measurement distortion during the diffusion process and require the model to predict the original corrupted image from the further corrupted image. We prove that our method leads to models that learn the conditional expectation of the full uncorrupted image given this additional measurement corruption. This holds for any corruption process that satisfies some technical conditions (and in particular includes inpainting and compressed sensing). We train models on standard benchmarks (CelebA, CIFAR-10 and AFHQ) and show that we can learn the distribution even when all the training samples have $90%$ of their pixels missing. We also show that when we finetune foundation models using small corrupted datasets (e.g. MRI scans with block corruptions), we learn the clean distribution without memorizing the training set.
The recommended way to run the code is with an Anaconda/Miniconda environment. First, clone the repository:
git clone https://github.com/giannisdaras/ambient-diffusion.git
.
Then, create a new Anaconda environment and install the dependencies:
conda env create -f environment.yml -n ambient
You will also need to have diffusers
installed from source. To do so, run:
pip install git https://github.com/huggingface/diffusers.git
We provide pre-trained models for AFHQ, CelebA-HQ and CIFAR-10 trained under different corruption levels. The checkpoints are available through the link: link. You will neeed ~16GB of disk space for all the models. To download from the terminal, simply run:
wget https://zenodo.org/record/7964925/files/checkpoints.zip?download=1
You might also need to download dataset and dataset statistics for training and for FID calculation. To do so, follow the instructions provided here.
To evaluate FID on the pre-trained models, run the following command:
EXPERIMENT_NAME=<experiment_name>
GPUS_PER_NODE=<number_of_gpus>
torchrun --standalone --nproc_per_node=$GPUS_PER_NODE \
eval.py --network=<path_to_folder_that_contains_the_pkl_file OR > \
--ref=<path_to_dataset_statistics> \
--outdir=<out_dir> \
--experiment_name=$EXPERIMENT_NAME \
--seeds=0-49999 --batch=16 \
--corruption_probability=<corruption_probability> --delta_probability=<delta_probability> --mask_full_rgb=True \
--num_masks=1 --guidance_scale=0.0 \
--training_options_loc=<path_to_training_options.json> \
--num=50000 --with_wandb=False
To train a new model on CelebA, run the following command:
torchrun --standalone --nproc_per_node=<GPUS_PER_NODE> \
train.py --outdir=<output_dir> --experiment_name=<experiment_name> --dump=200 \
--cond=0 --arch=ddpmpp --precond=ambient --cres=1,2,2,2 --lr=2e-4 --dropout=0.1 --augment=0.15 \
--data=<path_to_celeba_dataset> \
--norm=2 --max_grad_norm=1.0 --mask_full_rgb=True --corruption_probability=<corruption_probability> --delta_probability=<corruption_probability> --batch=256 --max_size=30000
To finetune the IF model, run the following command:
accelerate launch --mixed_precision="no" --multi_gpu train_text_to_image.py --resolution=64 --center_crop --random_flip --train_batch_size=32 --gradient_accumulation_steps=1 --max_train_steps=15000 --max_grad_norm=1.0 --learning_rate=3e-6 --lr_scheduler="constant" --lr_warmup_steps=0 --output_dir=<output_dir> --validation_prompts "person," --report_to="wandb" --corruption_probability=<corruption_probability> --delta_probability=<delta_probability> --use_8bit_adam --allow_tf32 --corruption_pattern="dust" --train_data_dir=<path_to_dataset> --empty_text --max_train_samples=3000 --seed=0
To generate images from the finetuned model, run:
python if_inference.py --checkpoint_path=<path_to_state_dict.pt> --output_dir=<output_dir> --batch_size=4 --corruption_probability=<corruption_probability>
This will generate 50000 images in the folder <output_dir>
.
You can run the following command to find for each generated image the nearest neighbors in the dataset using DINO.
python find_dataset_neighbors.py --input_dir=<model_outputs> --output_dir=<output_dir>
--data=<dataset_path> --features_path=celeba_features.npy --max_size=10000