ml_mdm is a python package for efficiently training high quality text-to-image diffusion models — brought to the public by Luke Carlson, Jiatao Gu, Shuangfei Zhai, and Navdeep Jaitly.

Diffusion models are the de facto approach for generating high-quality images and videos, but learning high-dimensional models remains a formidable task due to computational and optimization challenges.

ml_mdm is an end-to-end framework for high-resolution image and video synthesis — it is named after our technique: Matryoshka Diffusion Models.

Remarkably, we can train a single pixel-space model at resolutions of up to 1024x1024 pixels, demonstrating strong zero-shot generalization using the CC12M dataset, which contains only 12 million images.

The default installation dependencies, as defined in the pyproject.toml, are selected so that you can install this library even on a CPU only machine.

Users have run this codebase with Python 3.9,3.10 and cuda_12, cuda-11.8

> pip install -e .

Developers should set up pre-commit as well with pre-commit install.

> pytest   # run test cases that can work with just cpu
> pytest  -m ''  # will run all test cases - including ones that require a gpu
> pytest -m gpu # run only gpu test cases

We've uploaded model checkpoints to:

• https://docs-assets.developer.apple.com/ml-research/models/mdm/flickr64/vis_model.pth
• https://docs-assets.developer.apple.com/ml-research/models/mdm/flickr256/vis_model.pth
• https://docs-assets.developer.apple.com/ml-research/models/mdm/flickr1024/vis_model.pth

Note: We are releasing models that were trained on 50M text-image pairs collected from Flickr. In this repo, we provide scripts for downloading CC12M and configs for training equivalent models on CC12M data.

Feel free to download the models or skip further down to train your own. Once a pretrained model is downloaded locally, you can use it in our web demo, pass it as an argument to training, sampling, and more.

export ASSET_PATH=https://docs-assets.developer.apple.com/ml-research/models/mdm

curl $ASSET_PATH/flickr64/vis_model.pth --output vis_model_64x64.pth
curl $ASSET_PATH/flickr256/vis_model.pth --output vis_model_256x256.pth
curl $ASSET_PATH/flickr1024/vis_model.pth --output vis_model_1024x1024.pth

You can run your own instance of the web demo (after downloading the checkpoints) with this command:

torchrun --standalone --nproc_per_node=1  ml_mdm/clis/generate_sample.py --port $YOUR_PORT

In the ml_mdm.models submodule, we've open sourced our implementations of:

• U-Nets
• Nested U-Nets

ml_mdm.config contains the core configuration and cli logic. Many models, clis, and functions in this codebase are configured by passing in a dataclass object. We use SimpleParsing to dynamically create clis and allow passing in yaml config representations with --config_path.

In essence, simple_parsing will convert all passed cli arguments and yaml files into clean configuration classes like ml_mdm.reader.ReaderConfig, ml_mdm.diffusion.DiffusionConfig.

ml_mdm.config stores a global mapping of names to classes in MODEL_REGISTRY, MODEL_CONFIG_REGISTRY, PIPELINE_REGISTRY, and PIPELINE_CONFIG_REGISTRY.

MODEL_REGISTRY and PIPELINE_REGISTRY store information as shown in the following example:

*_CONFIG_REGISTRY[architecture name]["model"] = model name

*_CONFIG_REGISTRY[architecture name]["config"] = configuration class

MODEL_CONFIG_REGISTRY and PIPELINE_CONFIG_REGISTRY store information as shown in the following example:

*_CONFIG_REGISTRY[architecture name]["model"] = model name

*_CONFIG_REGISTRY[architecture name]["config"] = configuration class

architecture name and model name are passed into ml_mdm.config through the function parameter *names. where *names points to "architecture name", "model name"

Once you've installed ml_mdm, download these checkpoints into the repo's directory.

curl https://docs-assets.developer.apple.com/ml-research/models/mdm/flickr64/vis_model.pth --output vis_model_64x64.pth
curl https://docs-assets.developer.apple.com/ml-research/models/mdm/flickr256/vis_model.pth --output vis_model_256x256.pth

The web demo will load each model with a corresponding configuration:

• vis_model_64x64.pth will be loaded with the settings from configs/models/cc12m_64x64.yaml
• vis_model_256x256.pth will be loaded with the settings from configs/models/cc12m_256x256.yaml
• vis_model_1024x1024.pth will be loaded with the settings from configs/models/cc12m_1024x1024.yaml

In the demo, you can change a variety of settings and peek into the internals of the model. Set the port you'd like to use by swapping in $YOUR_PORT and then run:

torchrun --standalone --nproc_per_node=1  ml_mdm/clis/generate_sample.py --port $YOUR_PORT

If you just want to step through the process of training a model and running a pipeline without downloading a large dataset, we've put together a minimal example for you. It uses the dummy data from tests/test_files/

Feel free to try changing a variety of --args either directly in the cli or by editing the config yaml file

torchrun --standalone --nproc_per_node=1 ml_mdm/clis/train_parallel.py \
 --file-list=tests/test_files/sample_training_0.tsv \
 --multinode=0 \
  --output-dir=outputs    --config_path configs/models/cc12m_64x64.yaml \
  -num_diffusion_steps=10 \
	--num-training-steps=10

You should see a outputs/vis_model_000100.pth file. Now lets do something a bit more meaningful:

(OPTIONAL) Download the first 1K files of CC12m with this sample argument

The script is based on img2dataset's CC12M script.

curl https://storage.googleapis.com/conceptual_12m/cc12m.tsv | head -n 1000 > cc12m_index.tsv

# Add headers to the file
sed -i '1s/^/url\tcaption\n/'  cc12m_index.tsv

Note: if you want all of cc12m, remove | head -n 1000 from the call

Then prepare and split into train/validation

This script requires img2dataset, either run pip install '.[data_prep]' or just pip install img2dataset

python3 -m ml_mdm.clis.scrape_cc12m \
  --cc12m_index cc12m_index.tsv \
  --cc12m_local_dir cc12m_download

After running this command you will see the following files:

training.0.tsv # train index file
validation.tsv # validation index file
cc12m_download/
   00000.parquet  00000.tar  00000.tsv  00000_stats.json  validation.tsv
   00001.parquet ....

Now that we have our training file, we can select a model config and pass any additional training arguments:

# Modify torchrun arguments to fit your GPU setup
torchrun --standalone --nproc_per_node=8 ml_mdm/clis/train_parallel.py \
  --file-list=training_0.tsv \
  --multinode=0 --output-dir=/mnt/data/outputs \
  --config_path configs/models/cc12m_64x64.yaml \
  --num-training-steps=100   --warmup-steps 10

Note: configs/models/cc12m_64x64.yaml contains many more arguments, check it out for more details.

If you've downloaded a pretrained model, you can set the --pretrained-vision-file argument to point to its location on disk

Once training completes, you'll find the model in the folder defined by the --output-dir argument:

2024-07-22:17:58:46,649 INFO     [model_ema.py:33] Saving EMA model file: /mnt/data/outputs/vis_model_000100.pth
2024-07-22:17:58:47,448 INFO     [unet.py:794] Saving model file: /mnt/data/outputs/vis_model_noema_000100.pth

Now that we have a trained model, we can generate samples from the diffusion model:

torchrun --standalone --nproc_per_node=1 ml_mdm/clis/generate_batch.py \
  --config_path configs/models/cc12m_64x64.yaml \
  --min-examples 3 --test-file-list validation.tsv \
  --sample-image-size 64 --model-file /mnt/data/outputs/vis_model_000100.pth

If you want to skip the training step, you can update --model-file to point to one of our pretrained models

For long term storage, you can optionally upload your data to s3://{your_bucket}/datasets/{datasetname}/*.[tar,tsv].

Then update configs/datasets/cc12m.yaml to point to your s3 paths.

# configs/datasets/cc12m.yaml
train:
  files:
    - s3://mlx/datasets/cc12m-64x64/images_00.*.tsv
eval:
  files:
    - s3://mlx/datasets/cc12m-64x64/validation.tsv

# configs/datasets/reader_config.yaml
reader_config:
  append_eos: true
  bucket: ${your_bucket} # add your s3 bucket
  endpoint_url: None # boto will automatically infer the endpoint

Then you can use our dataset download helper:

python -m ml_mdm.clis.download_tar_from_index \
  --dataset_config_file configs/datasets/cc12m.yaml \
  --subset train --download_tar

python -m ml_mdm.clis.download_tar_from_index \
  --dataset_config_file configs/datasets/cc12m.yaml \
  --subset eval --download_tar

Take a look at configs/datasets/cc12m.yaml.

The code allows for multiple regular expressions to be provided. Keep in mind that the regular expressions are not globs -- they are regular expressions from the python re library. So if you wanted to use only 100 of the 1000 tar files in WebImage for training you can do the following:

train:
  files:
    - s3://mlx/datasets/example-dataset-100M_64px/example-dataset-100M-00[0-1]..-[0-9]*-of-01000.tsv
eval:
  files:
    - s3://mlx/datasets/example-dataset-100M_64px/validation.tsv

You can also mix and match the files. So if you wanted to merge CC12m and imagenet you could create a new yaml file with the following contents:

train:
  files:
    - s3://mlx/datasets/imagenet-64px/imagenet-train-000??-of-00100.tsv
    - s3://mlx/datasets/cc12m-64x64/images_00.*.tsv
eval:
  files:
    - s3://mlx/datasets/cc12m-64x64/validation.tsv

The S3 Bucket contains a series of files in this format, take a look at ml_mdm/clis/scrape_cc12m.py to generate your own.

2023-04-01 01:31:30   36147200 images_00000.tar
2023-05-10 11:34:49    1108424 images_00000.tsv
2023-04-01 01:31:26   36454400 images_00001.tar
2023-05-10 11:34:49    1109588 images_00001.tsv
2023-04-01 01:31:53   36116480 images_00002.tar
...

Minimal representations of these files can be found at tests/test_files/.

If you find our work useful, please consider citing us as:

@misc{gu2023matryoshkadiffusionmodels,
      title={Matryoshka Diffusion Models},
      author={Jiatao Gu and Shuangfei Zhai and Yizhe Zhang and Josh Susskind and Navdeep Jaitly},
      year={2023},
      eprint={2310.15111},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2310.15111},
}