Thanks for open-sourcing your work, I have been trying to use CMC on my custom toy dataset which has 2 views (RGB, Depth). The number of RGB and Depth images is 4000. I have 2 dataloader, one for RGB data and another for depth. And use corresponding RGB and depth to be positive pair.
When I use NCECriterion, the reslut is like,
If use Softmax-CE, the result becomes,
Could you please give me some suggestion about train with dual-modal datasets or small datasets ?
Thanks a lot!
Thanks for open-sourcing your work, I have been trying to use CMC on my custom toy dataset which has 2 views (RGB, Depth). The number of RGB and Depth images is 4000. I have 2 dataloader, one for RGB data and another for depth. And use corresponding RGB and depth to be positive pair.
When I use NCECriterion, the reslut is like,
If use Softmax-CE, the result becomes,
Could you please give me some suggestion about train with dual-modal datasets or small datasets ?
Thanks a lot!