Radius for topaz train

[annwhoorma]

train command accepts radius as an optional argument, and in most places it's set to either 1, 2, or 3 (default), regardless of what actual radius in pix is. For example, this is the radius description in train script:

training.add_argument('-r', '--radius', default=3, type=int, help='pixel radius around particle centers to consider positive (default: 3)')

However, the actual positive area particle center is usually larger (8 and more, even with downscaling). Even extract command reflects radius boundaries (for tuning) between 5 and 100.

When I tried to increase the radius for training, I got worse results than with a smaller radius. Why the radius for training don't seem to correspond to an actual radius of particles?

[tbepler]

The radius for train and the radius for extract do not relate to the same underlying concepts. The train radius specifies the set of pixels around the center coordinate of each labeled particle that are included as positive labels for training the classifier. Setting this to a small number (much smaller than the radius of the particle) is important for getting well centered predictions. The extract radius specifies how close predicted particles can be to each other when coordinates are identified from the pixel-level probabilities via peak finding. This radius should be close to the actual size of your particle, because it prevents multiple high scoring pixels within a single particle being called as centers of different particles by the peak finding algorithm.

[annwhoorma]

Understood, thanks a lot for giving a thorough explanation!

it'd be nice if the code or documentation had this explanation to avoid confusion :)

[tbepler]

There's some description of these parameters in this walkthrough, but I agree we should improve this documentation. Thanks for the feedback!