This project utilizes equivariant imaging as described by this paper applied to a deep cascade of convolutional neural networks, the model for which has been adapted from here
All used packages are listed in the Anaconda environment.yml file. You can create an environment and run
conda env create -f environment.yml
To test a model, simply run the following command with the appropriate filename
python test.py --ckp "./ckp/<filename>"
By default, it will load the provided ckp_final.pth.tar file.
Additional flags
--model-name: Text displayed on top of the output image--sample-to-show: Index of test images to display to the screen-h: Shows help for the above flags
To train EI, run
python train.py
Additional flags
--schedule: List of epochs when to drop the learning rate. Default--cos: Use cosine decay for learning rate and overrides schedule if set--epochs: Number of training epochs to perform--lr: Initial learning rate--wd: Initial weight decay-b: Batch size--ckp-interval: How often to save. Regardless, the model will be saved once training is finished--dataset: Path to the dataset used. Has to be a MATLAB file--ei-trans: Number of transformations--ei-alpha: Equivariance strength--views: Number of subsample views for radon transform-h: Shows help for the above flags
All of these have default values and the code will still work if you run the command above.