About fine-tune on small-scale datasets

[sha310139]

Hi,

Thanks for sharing this amazing work.
I'm interested in how to fine-tune on small-scale datasets.
Can you share fine-tuning commands and config files? It will be very helpful for me.

Thanks.

[VideoNetworks]

Training command and config is similar to train k400.

cmd = "python -u main_ddp_shift_egaze.py
--multiprocessing-distributed --world-size 1 --rank 0
--dist-ur tcp://127.0.0.1:23677
--name k400
-resume2 path-to-k400-weights
--cfg ./config/custom/egaze/egaze_vit_config_D_TK_8x8_Div4.yml"
os.system(cmd)

For main_ddp_shift_egaze.py, an extra resum2 is insert before resume args.

[sha310139]

Thanks for your reply !
I will try to modify and train it.