finetune strategy on 1k data about imagenet21k HOT 4 CLOSED

why do you want to finetune on 1k ? i dont really see the point
anyway, similar to standard 1k training, only with lower learning rate, and you can try adam instead SGD.

from imagenet21k.

Thanks. I want to reproduce the results in the table.
The finetune result on Imagenet-1K is only with lower learning rate and adam?

from imagenet21k.

basically yes (compared to standard 21K training)

from imagenet21k.

@pawopawo have you reproduced the acc?
@mrT23 In the paper, Table 3, the imagenet21k pretrained model were obtained using single label or semantic softmax training strategy? Did the official site mention if they actually use which strategy? Or by default people are all using semantic softmax way?

from imagenet21k.