Help needed

Need someone to review https://github.com/lucidrains/pixel-level-contrastive-learning/blob/main/pixel_level_contrastive_learning/pixel_level_contrastive_learning.py#L305

The distance threshold in the paper was said to be 0.7, but that did not work for me (resulted in nearly all positive pixel pairs).

[conradry]

To me, the size of a “feature map bin” is equal the downsampling factor of the model; i.e. for a resnet each bin is 32x32. The diagonal length of the bin would be ~45 and the threshold distance (0.7 * 45) would be ~32. It at least sounds plausible, but the paper isn’t very clear.

[lucidrains]

ok, I decided to go with the assumption of having the diagonal of each ‘pixel’ in the feature map normalized to 1

https://github.com/lucidrains/pixel-level-contrastive-learning/commit/8fa3425a243eb4892cfa5a7b453c616370fae2c2