Comparisons of timm Optimizers w/ Caution

This repo contains summaries of several sets of experiments comparing a number of optimizers with and without caution (https://huggingface.co/papers/2411.16085) enabled.

The runs were all performed training a smaller ViT (vit_wee_patch16_reg1_gap_256) for 200 epochs (10M samples seen) from scratch on the timm 'mini-imagenet' dataset, a 100 class subset of imagenet with same image sizes as originals.

So far I have results for adamw, laprop, and mars (https://huggingface.co/papers/2411.10438). You can find full results in sub-folders by optimizer names. In all of these runs, the experiments with 'c' prefix in the name have caution enabled.

This is what the 'caution' addition looks like in an optimizer:

    mask = (exp_avg * grad > 0).to(grad.dtype)
    mask.div_(mask.mean().clamp_(min=1e-3))
    exp_avg = exp_avg * mask

Train args:

./distributed_train.sh 2 --dataset hfds/timm/mini-imagenet --num-classes 100 --model vit_wee_patch16_reg1_gap_256 -j 8 --epochs 200 --warmup-prefix --sched-on-updates --warmup-lr 0 --mixup .2 --model-ema --model-ema-decay 0.999 --model-ema-warmup --aa rand-m9-mstd0.5-inc1 --remode pixel --reprob 0.25 --amp --weight-decay .05 --drop 0.1 --drop-path .1 -b 288 --opt cadamw --lr 1e-3

LaProp

LaProp Top-1 Evaluation Accuracy on Mini-ImageNet

LaProp Train Loss

AdamW

AdamW Top-1 Evaluation Accuracy on Mini-ImageNet

AdamW Train Loss

MARS

MARS Top-1 Evaluation Accuracy on Mini-ImageNet

MARS Train Loss