< Previous post | Next post >

[2026-01-15][Daily]

First experiments that we run on multiple datasets with growing images

First results!

Pytorch intraop vs DAHL on datasets with growing images.

Pretty strange that pytorch has a worse performance than dahl on small images, that not what we noticed on previous experiments. However here we performed the benchmarks on only 1 epoch, maybe pytorch "snowballs" and improve its performance over multiple epochs? We will definitly try with 5 epochs like in previous experiments.

Also plotting the total number of pixels on top of metrics with different scales does not make sense, maybe we should normalize?

< Previous post | Next post >