Quickstart: Using deepCR

Quick download of a HST ACS/WFC image

from deepCR import deepCR
from astropy.io import fits
image = fits.getdata("jdba2sooq_flc.fits")
mdl = deepCR(mask="ACS-WFC",

binary_mask = mdl.clean(image, threshold = 0.5, segment=True, patch=1024)
binary_mask, cleaned_image = mdl.clean(image, threshold = 0.5, inpaint=True, segment=True, patch=1024)

The binary CR mask is generated by applying a threshold [0, 1] to the raw deepCR output, which is a probability map. The keywords, segment=True, patch=1024, directs deepCR to process the image in patches of 1024 by 1024, which limits memory footprint to ~2gb. Using a smaller patch size slows things down. Make sure your patch size don’t result in edge patches that has dimensions less than 256. For example, using a patch size of 256 for a 300*300 image will result in a tiny 44*44 patch at the opposite corner.

For CPU, you should specify n_jobs=-1 to enable multiprocessing unless your pytorch code is MKL optimized. Monitor CPU usage – if CPU usage for single process is > 100% it means intel MKL is probablly in place. You should never use this option for GPU which is bottlenecked by memory, not FLOPS.