-
Hi, is there a training script / configuration (as #917 for A2) that exactly corresponds to A3 from the ResNets Strike Back paper? While we can guess most of the configuration from the hyperparams + timm config settings, we want the training script to be as exact as possible for reproducibility reasons. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 2 replies
-
Two configs attached, one for seed 21, one for seed 0 with train interpolation fixed at bicubic (otherwise randomly selects per sample). They both exceeded the paper 78.1 as you can see. This was run on 4xV100, batch 512 per card, so you'll need to scale LR appropriately if that differs. _78_25-fusedlamb-cosine-lr0.00800-wd0.020000-n0-rand-m6-mstd0.5-inc1-m0.1-sd0.0-d0.0-ls0.0-100-100-resnet50-args.yaml.txt |
Beta Was this translation helpful? Give feedback.
Two configs attached, one for seed 21, one for seed 0 with train interpolation fixed at bicubic (otherwise randomly selects per sample). They both exceeded the paper 78.1 as you can see. This was run on 4xV100, batch 512 per card, so you'll need to scale LR appropriately if that differs.
_78_25-fusedlamb-cosine-lr0.00800-wd0.020000-n0-rand-m6-mstd0.5-inc1-m0.1-sd0.0-d0.0-ls0.0-100-100-resnet50-args.yaml.txt
_78_23-fusedlamb-cosine-lr0.00800-wd0.020000-n0-rand-m6-mstd0.5-inc1-m0.1-sd0.0-d0.0-ls0.0-101-100-resnet50-args.yaml.txt