CMU-02699-Image-Segmentation-via-GANs2

Spring 2022 Bioimage Informatics (Self-Study) project

This project explores using a deep generative adversarial network (GAN) to perform semi-supervised image segmentation on the 2015 MICCAI Gland Challenge dataset. This is an improvement over version 1, adding triplet loss and hard negative mining to the training process.

See version 1 here.

See the final report here.

How to run on an EC2 instance

1. Copy octopus library to Deep Learning instance.
2. Copy data to instance. Data was is found at here. I organized the files into the following format.

warwick
    ├── a
    ├── a_anno
    ├── b
    ├── b_anno
    ├── training
    └── training_anno

I also padded all the numbering in the filenames (i.e. testA_1_anno.bmp -> testA_01_anno.bmp)

4. Copy this codebase to instance.
5. Log into instance.
6. (Optional) Mount drive on EC2.

bash octopus/bin/mount_drive

6. Activate conda environment.

conda activate pytorch_p38

7. Install wandb.

bash octopus/bin/setup_wandb

8. Run code for a single run or a sweep.

Single run

python /home/ubuntu/CMU-02699-Image-Segmentation-via-GANs2/run_octopus.py --filename=/home/ubuntu/CMU-02699-Image-Segmentation-via-GANs2/configs/remote/config_remote_GAN_only_002.txt

Sweep

wandb sweep /home/ubuntu/CMU-02699-Image-Segmentation-via-GANs2/sweeps/remote/sweep_remote_grid_023.yaml