Training vit_base_patch16_224_in21k on Cifar10 #1392
Unanswered
animesh-007
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
Hi, I was trying to train vit_b_16 on cifar10 using timm library. I am currently using the following command with no augmentation.
python train.py ./datasets/cifar10 -d "torch/cifar10" --dataset-download -b=64 --img-size=224 --epochs=50 --color-jitter=0 --amp --sched='cosine' --model-ema --model-ema-decay=0.995 --min-lr=1e-8 --warmup-epochs=3 --train-interpolation=bilinear --aa=v0 --pretrained --lr=2e-4 --model=vit_base_patch16_224_in21k --opt=adam --weight-decay=1e-4 --num-classes 10
I am getting the best accuracy on the 0th epoch only.
Test (EMA): [ 0/156] Time: 0.126 (0.126) Loss: 0.2100 (0.2100) Acc@1: 93.7500 (93.7500) Acc@5: 100.0000 (100.0000) Test (EMA): [ 50/156] Time: 0.072 (0.076) Loss: 0.1656 (0.1602) Acc@1: 98.4375 (97.5797) Acc@5: 100.0000 (99.9081) Test (EMA): [ 100/156] Time: 0.074 (0.075) Loss: 0.1504 (0.1661) Acc@1: 98.4375 (97.4010) Acc@5: 100.0000 (99.8917) Test (EMA): [ 150/156] Time: 0.076 (0.075) Loss: 0.1578 (0.1669) Acc@1: 98.4375 (97.4234) Acc@5: 100.0000 (99.8965) Test (EMA): [ 156/156] Time: 0.020 (0.075) Loss: 0.0986 (0.1665) Acc@1: 100.0000 (97.4500) Acc@5: 100.0000 (99.8900) *** Best metric: 97.77 (epoch 0)
Can anybody guide me on how I can improve the performance?
Beta Was this translation helpful? Give feedback.
All reactions