Skip to content
This repository has been archived by the owner. It is now read-only.


Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


Image Segmentation Techniques on the ISBI 2012 dataset:

All results are from src/models/ implementation, which has an Argparse CLI that should be a good starting point.

Running the code

  • All of the data is kept in the /data directory, so no need to download anything.
  • Make sure dependencies in requirements.txt are installed (you'll know when you run it).
  • Run the unet model to train: python src/models/ train without saved weights, or with saved weights python src/models/ train --weights /path/to/weights_file.hdf5.
  • To make a submission (i.e. predictions on testing data): python src/models/ submit --weights /path/to/weights_file.hdf5 --tiff /path/to/saved_submission.tiff.


  • 3/13/17

    • Rand score ~0.9620, information score 0.9828.
    • Commit
    • Submission
    • Used a weighted cost function to put 5x error on the boundary areas, similar to approach described in UNet paper.
  • 3/11/17:

    • Rand score ~0.9571, information score ~0.9821.
    • Commit
    • Submission
    • Re-seeding the RNG before every new batch for both training and validation gave the best information score so far. Didn't train this network for very long, so hopefully more training will improve the rand score.
  • 3/11/17:

    • Rand score ~0.9641, information score ~0.9805.
    • Commit
    • Submission
    • Using all 30 images for both training and validation gave the best rand score so far, but just barely.
  • 3/9/17:

    • Rand score ~0.9511, information score ~0.9805, would have made it on the leaderboard but it was worse than my prior submission.
    • Commit
    • Submission
    • Most significant difference is that I trained on all the images and validated with 9 images. I trained longer and got the loss lower than on the previous submission, so this leads me to believe there must have been a little overfitting.
  • 3/8/17:

    • Rand score ~0.9637, information score ~0.9814, ~30th place.
    • UNet based on JocicMarko example, sampling 256x256 tiles from a montage of images.
    • Commit
    • Submission
    • The biggest difference seems to have been using the he_normal initialization for all of the convolutional layers, inspired by this blog post.


No releases published


No packages published