MODE '22 workshop data challenge - Giles Solution

My solution to: https://github.com/GilesStrong/mode_diffprog_22_challenge

Final score: private IOU = 0.824

See included PDF for details.

Notebooks

The notebooks contain all the specific code for the problem, and all the data processing starting from the original challenge data.

0_unet-pixelshuffle_depth-encode_data-aug_se-net_f8r2_full.ipynb
- A pure UNet model with a score of ~0.812
1_unet-pixelshuffle_depth-encode_data-aug_gravnet_osa_f32_full.ipynb
- Extends the model to replace the centre of the UNet with a GravNet based network, IOU ~= 0.818
2_unet-pixelshuffle_depth-encode_data-aug_gravnet_osa_f32_full-long_ensemble5.ipynb
- Trains an ensemble of 5 models over 100 epochs each, IOU ~= 0.821
3_unet-pixelshuffle_depth-encode_data-aug_gravnet_osa_f64_full-long_studentval_ensemble5-GPU1.ipynb
- Uses pseudo-labels for the test data, generated by the predictions of the previous ensemble, and trains an ensemble of 10 models over 200 epochs.
- IOU ~= 0.822
- The notebook by default trains 5 models, and so either needs to be either copied to train models over 2 GPUs in parallel, or set to train 10 models sequentially

Setup

For GPU usage, I'd recommend installing PyTorch yourself, based on driver versions, etc. See https://pytorch.org/get-started/locally/ for pip and conda commands.

The solutions rely on the LUMIN front-end to PyTorch. Due to various bug-fixes and changes, these solutions require the current bleeding-edge version. For now (close to 2022/09/05), install from source with:

git clone git@github.com:GilesStrong/lumin.git
cd lumin
pip install .

But hopefully later, this will be enough:

pip install lumin==0.8.1

if that doesn't work, try:

pip install lumin==0.9.0

Additionally, you'll need:

pip install sparse

Any problems, open an issue.

GilesStrong/mode_diffprog_22_challenge_giles_solution

MODE '22 workshop data challenge - Giles Solution

Notebooks

Setup