About performance in Cub-200 finetune.

[narrowsnap]

Thank you for opensource this pretrained models.
I finetuned 32x32, 32x16 and 32x8 on Cub-200 dataset with 2 1080Ti, which only achive 87.5% accuracy on testset with image size 448x448. Image size with 224x224 only get 84.43% test acc. I find biger batch size got better performance but I only have 2 GPUs.
In the paper with 32x16 pretrained model reached 89.2 accuracy.
Could you please show more details about finetune.
Thanks.

[dkm2110]

@narrowsnap what is the batch size per gpu u r using for the models? Since this is model trained on 940M images and 1.5k labels, you should expect an accuracy of about 87.9% or similar. Also for cUB it is important to do param sweep on LR, weight decay and LR schedule.

[narrowsnap]

@dkm2110
This is my experiments.

model	img size	batch size/gpu	test acc
32x8	448x448	8	0.8702(0.8752 with mixup)
32x8	224x224	32	-
32x16	448x448	4	0.8626
32x16	224x224	16	0.8443

With SGD(model.parameters(), lr=0.001, momentum=0.9, weight_decay=1e-5)
and MultiStepLR(optimizer, milestones=[30, 60], gamma=0.1)
epoches= 100.