/dgm-medfusion

Implementation of Medfusion for the DGM Image Statistics Challenge

Primary LanguagePythonMIT LicenseMIT

dgm-medfusion

Implementation of Medfusion for the DGM Image Statistics Challenge

Requirements

  • Docker
  • Nvidia Drivers
  • Nvidia Docker for GPU-Accelerated Containers

Running Sampling

  1. Prepare the script to run

    chmod a+x sample.sh

  2. (Optional) Select sampling parameters

  3. Run the sample.sh script

    sudo ./sample.sh

Sampling Parameters

Default:

AUTOENCODER_CHECKPOINT="project/model_checkpoints/vae.ckpt"
PIPELINE_CHECKPOINT="project/model_checkpoints/pipeline.ckpt"
OUTPUT_PATH="project/generated_datasets/"
STEPS="300"
SEED="0"
BATCH_SIZE="20"
DATASET_SIZE="10000"
NUM_WORKERS="4"

Note: You can change this parameters in the sample.sh script

Parameters Description

  • AUTOENCODER_CHECKPOINT: Path to the weights of the Variational Autoencoder used in the Diffusion Pipeline
  • PIPELINE_CHECKPOINT: Path to the weights of the Diffusion Pipeline that generates the dataset
  • OUTPUT_PATH: Path to the generated dataset directory
  • STEPS: Steps in the sampling pipeline. 300 by default. If you want a faster sampling you can decrease to 100 or 200 in detriment of image quality. If you want higher quality in your images you can increase to 400, 500+ (This increases the sampling time significantly)
  • SEED: Random seed used to generate the dataset
  • BATCH_SIZE: Number of images generated each batch. 20 images by default. This batch size peaks ~15 GB VRAM usage. If you are training in a larger GPU, increase this value for faster sampling.
  • DATASET_SIZE: Number of images generated in total. 10000 by default. In the default values, this usually takes about 8 hours.
  • NUM_WORKERS: Number of CPUs used to write the images.