Thanks for the kind words.

We did all training at 20x, but we rescale to 40x just for post-processing. This enables us to use the same post processing script as the original HoVer-Net, rather than playing around with the parameters.

In the baseline.ipynb file, in the process_segmentation() function, you will see we rescale it by a factor of 2:

# HoVerNet post-proc is coded at 0.25mpp so we resize
np_map = cv2.resize(np_map, (0, 0), fx=2.0, fy=2.0)
hv_map = cv2.resize(hv_map, (0, 0), fx=2.0, fy=2.0)
tp_map = cv2.resize(
                tp_map, (0, 0), fx=2.0, fy=2.0,
                interpolation=cv2.INTER_NEAREST)