rain.sh | Let It Rain! -

My belief is that you have some bugs in your code. One thing you didn't explain to me is why you are calling stats.mode(labels) on your labels for example. Also, I can't tell whether or not you used SMOTE correctly and so on.

However, since you cannot provide any data here on stackoverflow I took the Rain in Australia dataset from Kaggle.

It looks very similar to yours. The big difference is that we have only two classes for "rain tomorrow": yes and no. But it's also quite imbalanced:

This is often due to environmental variables being passed on, or not passed on, to the container without noticing. To ensure this is not an issue, you can use -e or --cleanenv. This will prevent any variables not prefixed with SINGULARITYENV_ from being loaded into the container.

That said, the warning WARNING: skipping mount of sysfs: no such file or directory is also concerning: singularity was unable to mount /sys into the image because it doesn't exist on the host server. That particular python error also seems to be specific to windows 10. Singularity doesn't currently support windows, even with the magic of WSL2.

I think the problem is that you are clearing the session before training the model, doing this would make no sense, because clearing the session cleans the model structures in memory, so there would be no model representation in the TensorFlow side, making training fail.

So do not juse K.clear_session() in this case. It does not seem to be needed.

I assume NumpyLogReg works very well on binary classification. Make use of the same class for multi-class classification using the One-Vs-Rest (ovr) technique.

Let's assume the dataset is having 3 classes A, B, C

1. Call the binary classification model with the class label A as +ve class & B, C as -ve class and note down the prob scores
2. Repeat the same by considering B as +ve & A, C as -ve and C as +ve & A, B as -ve. Note down the respective prob scores.
3. Basically, if there are n classes, there will be n binary classifier models i.e, fitting one classifier per class
4. By careful inspection of the classifier of each of the classes (i.e., by analyzing the prob values), you can achieve multi-class classification & the model will be highly interpretable.

Please refer this guideline for more detailed explanation