README.md

Setup

We use params-proto to specify the hyperparameters, and generate sweep.jsonl files. We use ml-logger to centralize metrics logging and checkpointing. We use jaynes to launch the experiments on the cloud.

conda create -n sculpting python=3.8
conda install pycurl
pip install params-proto jaynes ml-logger cloudpickle==1.3.0

Colored MNIST Experiments

There are two main experiments: Binary composition between two base diffusion models, and chained composition between three base diffusion models. The steps are:

1. train base diffusion models on various colored MNIST digit distributions
2. train binary classifier over pairs
3. sample from the binary compositions
4. train binary classifier over pairs used in the chaining experiments
5. sample from the chained compositions

The experiments folder contains the scripts used to launch the training and sampling. The models folder contains the model definitions.

diffusion_chaining
├── README.md
├── __init__.py
├── experiments
│   ├── chain.jsonl
│   ├── chain.py
│   ├── ddpm.jsonl
│   ├── ddpm.py
│   ├── sculpting.jsonl
│   └── sculpting.py
├── bcomp.py
├── bcomp_sampler.py
├── chain.py
├── chain_sampler.py
├── ddpm.py
├── ddpm_sampler.py
└── models
    ├── classifier_model.py
    ├── score_model.py
    └── util.py
3 directories, 17 files

Preparation: Training Base Diffusion Models

ddpm.py

and to sample from these models:

ddpm_sampler.py

Experiment I: Binary Composition: Training Binary Classifier

bcomp.py

now, to sample from the binary compositions

bcomp_sampler.py

Experiment II: Training Chained Classifiers

chain.py

now to sample from the chained compositions

run:

python chain_sampler.py