Setup

Aside from the requirements in requirements.txt you'll need to install reproject from source, specifically from https://github.com/adobles96/reproject. Clone the repo locally and then run pip install -e path/to/local/reproject.

Files

• proj_tools.py: Utility file that contains most of the heavy lifting around data manipulation:
  • Projections to and from healpix to flat
  • Smoothing logic to make patches uniresolution and to create the training inputs
  • Tiling logic (similar to ForSE+)
  • etc.
• gen_dataset.py: A script used to generate two sets of patches:
  1. training_patches.pkl: A dict containing both input and output image patches at different resolutions for trianing.
  2. inference_patches.pkl: A dict containing input image patches for inference.
• dataset.py: A module containing torch.utils.data.Dataset objects to be used during training and inference.
• train.py: The training script.
• inference.py: A script that runs a trained model on inference patches, to produce a super-resolution version of the full sky. It generates three files:
  1. sr_patches.pkl: A dict mapping resolution to an array of (N, C, H, W) patches.
  2. sr_map.npy: A numpy array of (C, npix_full_sky) holding a full sky healpix map for Q and U.
  3. overlap_map.npy: A numpy array of (C, npix_full_sky) holding a map of the full sky counting the number of overlaping patches per pixel.
  4. effective_overlap_map.npy: A numpy array of (C, npix_full_sky) holding a map of the full sky counting the number of overlaping patches per pixel, weighted by their apodizations.
• plotting.py: Plotting utilities used to generate plots for Minkowski Functionals and Power Spectra at the patch level.

Workflow

1. Run gen_dataset.py with desired arguments. Eg: python gen_dataset.py --smooth_before_resize --npix_out 80
2. Decide on a run config (config.yml)
3. Run train.py. Eg: python train.py
4. Run inference.py with the desired checkpoint produced during training. This will generate super-resolution patches and the full sky map. Eg: python inference.py --model_checkpoint 'path/to/checkpoint.ckpt' --patchsize_multiple 20
5. Use outputs to run analyses. See sample_plots.ipynb for examples.

Things to try?

• MSELoss instead of L1Loss
• Different patchsize_multiple
• Increase patch overlap to generate more training patches (decrease overlap for inference patches)
• Shift centers to generate more training data
• Rotate patches in full sky to generate more training data.
• Architectural changes
  • Transfer Learning
  • Pure convolution architecture
  • Pure vision transformer?