[Moreh] Running DDIM on Moreh AI Framework

Data will be downloaded automatically by the program. No preparation needed.

This repo supports three datasets: cifar10, LSUN (bedroom & church categories) and CelebA (see in configs/). Currently, LSUN and CelebA cannot be downloaded due to their server side being down. The instructions below only run with cifar10.

git clone https://github.com/loctxmoreh/ddim
cd ddim

Using a100env.yml to create a conda environment:

conda create -f a100env.yml
conda activate ddim

Using hacenv.yml to create a conda environment:

conda create -f hacenv.yml
conda activate ddim

After create env by hacenv.yml, re-run this command to install torch with Moreh:

conda install -y torchvision torchaudio numpy protobuf==3.13.0 pytorch==1.7.1 cpuonly -c pytorch

With cifar10, edit configs/cifar10.yml and change training.n_epochs to a small value for testing

During training, the program writes running-related data to a separate directory. From now on, this directory is refered to as the work_dir.

First, edit train.sh and change work_dir to the desired directory. work_dir should have write permission. Then run the training script:

./train.sh

Edit sampling.sh and change work_dir to the same one used in training. Then, run the sampling script:

./sampling.sh

You can either use the local pretrained checkpoints (obtained in the previous training step), or the pretrained models provided by the authors, to sample new images. Just edit sampling.sh and comment out the irrelevant command.

The generated images are stored in $work_dir/image_samples.

These are instructions to reproduce legacy loading error.

conda env create -f hacenv2.yml
conda activate ddim
update-moreh --force --driver-only --target 0.2.0

./sample.py

The script will failed when tried to load the downloaded checkpoint stored at $HOME/.cache/diffusion_models_converted.