How to reproduce A3 procedure (ResNet strikes back)

[EthanChen1234]

Hi @rwightman

the paper, ResNet strikes back: An improved training procedure in timm.
I'm trying to reproduce A3 procedure, refer to #924

version = '0.5.2'，runs on 4xV100, batch 512 per card

my config yaml, has a little difference with _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

• train_interpolation：random -> bicubic
• opt：fusedlamb -> lamb
• bce_target_thresh: 0.2

the best top1 acc is 76.278%, could you help find the gap?

the config yaml and train log, as follows:
args.yaml.txt
bicubic.log

[rwightman]

@EthanChen1234 did you try to eval the best checkpoint at 224x224 crop 0.95?

Also, since paper was published, train at 176 was found to be a better match for eval at 224 (but increases the compute/time a bit)

[EthanChen1234]

Hi, rwightman, thanks very much for your quick reply.
The top1 acc 76.278% was the evaluation result when training the model, which img-size was 160x160.

After eval the best checkpoint at 224x224 crop 0.95, the top1 acc is 78.226%.

[JeanKaddour]

[EthanChen1234]

1. add --bce-target-thresh 0.2, you can refer the detailed information from Training configuration for ResNets Strike Back, A3 configuration #924
2. the val dir, should have the same directory structure as train dir

[JeanKaddour]

1. oh, I actually did add --bce-target-thresh 0.2. Sorry, did not mention this. I then used the config you linked.
2. thanks, will compare with that!

[JeanKaddour]

It was indeed the wrong dataset... sorry y'all and thank you!