javadan/CGIAR-Crop-Damage-Classification-Challenge

zindi.africa # CGIAR-Crop-Damage-Classification-Challenge zindi.africa First attempt at a Computer Vision competition. Came 100/347. Got my best score with code that I subsequently lost (i.e. overwrote, with newer code). My top result was .65 with efficientvit_l2.r384_in1k and some random cropping and CutMix and some other augmentations. First time using FastAI. It seems super buggy to me. But nice when it works. e.g. Save as .pth or export as .pkl?: sometimes it won't let you export as .pkl and sometimes it won't let you load from .pth. e.g. Using some complex batch_tfms like CutMix? Then some things like lr_find and some item_tfms options won't work. Image_tiling_CGIAR_2023.ipynb - image tiling, chopping up images into chunks. I thought we'd get more detail from the high res images. Validation score went up to 1.3, so apparently that was worse. CNN_and_histogram_CGIAR_2023.ipynb WIP- ran out of time 1. basic efficientvit_l2.r384_in1k, (about .7 ish on public LB) then 2. caformer_s36.sail_in22k, (about .7 ish on public LB) ,then 3. convnextv2_large.fcmae with attempt to merge histograms, which in hindsight, is, i dunno. That data is already encoded in the pixels. Right? We're gonna have to ask GPT4 about this one, and maybe even Claude. GPT4 suggests isotonic regression. There's always more to learn. got weird output. still meaningful output? logits: ID DR G ND WD other ID_SJBCSZ -1.5997245 -0.21110994 -4.5409064 -0.5731514 -4.0951037 ID_4UMPJL -1.5987856 -0.21640371 -4.540986 -0.56841195 -4.09604 ID_AB4DZF -1.5956886 -0.21697117 -4.5416307 -0.570616 -4.0943494 ID_H257JS -1.6071882 -0.20562026 -4.5432496 -0.56928974 -4.099599 ID_JKST46 -1.5893495 -0.22358689 -4.5396028 -0.5720792 -4.091543 probs: ID,DR,G,ND,WD,other ID_SJBCSZ,0.16802011,0.44741762,0.010551222,0.36050996,0.016381206 ID_4UMPJL,0.16815141,0.44610924,0.010550389,0.36160335,0.01636613 ID_AB4DZF,0.16858506,0.44596905,0.010543662,0.3610 …