search

tiongleslie / 3D-C2FT

MIT License
11 stars 1 forks source link

Training failure problem #2

Open WFLiu0327 opened 1 year ago

WFLiu0327 commented 1 year ago

Hello, I am recently been working on 3D reconstruction. I tried to run your 3D-C2RT project. I just replaced the CNN encoder with ConvNeXt-Large which has better representation ability, as well as modified the size of batch-size, but my loss function does not drop when I train up. What do you think this may be due to? We will cite your article in our near future work, hope to get your reply.

tianbao-li commented 1 year ago

I have encountered the same problem as you. I reproduced the original code without any modifications, but after training for 300 epochs, the loss has not decreased.

tianbao-li commented 1 year ago

I tried the learning rate settings in the paper, but it didn't work.

WFLiu0327 commented 1 year ago

Do you think it's possible that the code just doesn't work well?

WFLiu0327 commented 1 year ago

Let's communicate directly in Chinese.

tianbao-li commented 1 year ago

I want to know if there is a problem with my reproduction code, or the training time is not long enough, I wonder if the author can provide the training loss curve

george-kalitsios commented 1 year ago

I'm facing the same issue; even after 400 epochs of training, the loss has not decreased. Did you come up with a solution?