Added ~60k more images

Lots more image classifiers trained, and existing classifiers improved

Training Notes:

~300k images from Danbooru, r-34, e621 etc...

LR 2e-4

TE_LR 5e-5

batch 4

GA 32

dim 64 (resized with kohya, with sv_ratio=10, and max dim 64)

conv_dim 32

alpha 32

conv_alpha 32 (I meant to set this to conv_dim/2, but ohh well)

scheduler: cosine_with_restarts, 33 restarts

base model NAI

No --vae

flip aug

clip skip 2

225 token length

bucketing at 1024 max 1024 (I should have stuck with max_bucket_size=768 for reduced training time)

tag drop chance 0.1

network_dropout 0.33

scale_weight_norms 1 (either this or dropout resulted in less blown out results, which allowed for longer training at higher LR)

tag shuffling

50,000 unique tags, but about 500 frequently occurring ones (see attached training data for tag list)

about 14 days training time (slower than last because of bucket size choice)