I decided to start journaling because I have a bit of traction, and the analysis isn’t totally wrong. Also Lacker said I should have a notebook, but I never write stuff with pens.
Ran 5000 training batches last night over night, noticed the data wasn’t being properly randomized. Patch size 6.
Fixed it, ran a new 50 batches. Not very accurate at that level of training, 50/50. Patch size 6.
Ran a new test of 500 batches. 52% accurate. Patch size 6.
Ran a new test of 500 batches with a bigger patch (8px patch instead of 6px). Only 48% accurate, and only predicted 8.5% of test set on.
Ran a new test of 500 batches with a bigger patch (7px patch). Only 50% accurate, 50% of labels on.
Started a new test of 5000 batches over night, with a patch size of 5. Again, only 50% accurate.