hassanmohsin / dlscore-cnn

Hi there 👋

dlscore-cnn's Issues

would you like to provide a toy dataset?

Thank you for your repo. I want to try your code, but I met difficulties on install htmd. Would you like to provide a toy dataset that I can try your code? Many thanks.

May I know how many data points did you used?

I noticed that you didn't do data augmentation like in Kdeep original post.

DLSCORE-CNN/pytorch/test.py

Line 53 in c3e3a15

pred[i] = _pred.mean().cpu().numpy()

Hello!

I'm refactoring your code as an important part of my master's project.

After I completed the entire process of pytorch-based programming (referring to your code), I immediately conducted a performance evaluation test of the model. The result is this:
{'RMSE': 1.5596018, 'Spearman Correlation': 0.6496260994698196, 'Pearson Correlation': 0.6565482435856405}

This is extremely inconsistent with the performance evaluation results in the original text of Kdeep, and then I carefully read the original text and your code. I found the following issues:

original text: We augment our data by rotating each subgrid 90 deg, providing 24 times our initial training size. This augmenting methodology is also used at prediction time and then averaged out, reducing variance.
but you: I didn't find any code for data augmentation, which also appeared in the validation of the model (It may also be because my programming ability is too weak or this part of the code has not been seriously discovered).

So I would like to ask you, do you think data augmentation should be done in this model? What is the performance of the model trained in your environment?

Bast wish to u, Im waiting for your reply. :)