catdx

This repository contains a Vivit fine-tuning project on medical data.

Need to Know

• A .secrets file must be placed in the repo's root directory to store your wandb API key.
• The src directory contains the main codebase, while the scripts directory contains scripts for running the project.
• The configs directory contains configuration files for the project.
• Set up the dataset directory and update the configuration at src/scripts/configs/config.yaml accordingly.
• The wandb project integration is pending (#TODO).

Python (Single Run, Single GPU) Container

! WARNING: This container is discontiued, use the Accelerate container instead since it can run both single and multi GPU.

Accelerate (Single Run, Multi GPU) Container

To run the script using the Accelerate container on HSSH, use the following command:

submit \
  --name vivit-contrastive \
  --gpus N \
  --host hssh2 \
  --mount "$(pwd)":/scratch/catdx \
  --mount /home/vitturini/shared/RVENetCropRszMp4:/scratch/catdx/RVENetCropRszMp4 \
  eidos-service.di.unito.it/vitturini/vivit:accelerate

The script will use by default every GPU allocated to the container, everything else is the same as the Python container.

Devcontainer

Run the development container on HSSH using:

submit \
  --name vivit-dev \
  --gpus N \
  --mount /mnt/fast-scratch/vitturini/catdx:/scratch/catdx \
  --mount /home/vitturini/shared/fix_apical4:/scratch/catdx/fix_apical4 \
  eidos-service.di.unito.it/vitturini/vivit:dev

Check which machine your devcontainer is running on:

docker service ps my_container

When finished, scale the container down:

docker service scale my_container=0

To reconnect (scale up) with the devcontainer:

docker service scale my_container=1

To run the multi-gpu training script from inside the container:

accelerate launch -m src.scripts.main

You can add arguments to the accelerate launch command, for example:

accelerate launch --num_processes 1 --mixed_precision fp16 -m src.scripts.main

Sweep Container

! WARNING: This container is not yet functional. It is a work in progress and should not be used for now.

Building and Pushing the Container

To build the container, run the following command:

docker build -t eidos-service.di.unito.it/vitturini/vivit:base -f docker/Dockerfile_base .

To push the container to the repository, use:

docker push eidos-service.di.unito.it/vitturini/vivit:base

Dataset format

To make the code work as it is, the dataset must be in the following format: a dir with every video and a csv file with the labels.

dataset/
|-- video1.mp4
|-- video2.mp4
...
|-- videoN.mp4
|-- all_files_with_partition.csv

The csv file must contain the following columns:

• file_name: the name of the video file with the extension (e.g., video1.mp4)
• CO: the label for the CO
• Every other label
• partition: the partition of the video (train, val, test)

Nevertheless, the code uses the Huggingface Datasets library, so it is possible to use any dataset format supported by the library with a limited amount of work.

WARNINGS

• A refactoring of the code is scheduled soon.
• The sweep container mounts the source code, so code should not be modified from anywhere when running the sweep.