search

Thunfischpirat / SpokenDigits

This is our submission for the final graded project for the WS22/23 course "Neural Networks: Theory and Implementation" at Saarland University.
1 stars 0 forks source link

Model validation accuracy collapsed #20

Closed Thunfischpirat closed 1 year ago

Thunfischpirat commented 1 year ago

Somehow the neural network model's performance during training doesn't surpass 10% anymore. Why is that the case?

Thunfischpirat commented 1 year ago

When we updated the models to return class embeddings instead of class probabilities so that we could create the t-SNE embeddings for task II.3, we forgot to adjust the training loop of train_model(). in line 111, we now applied the softmax to the model output. This fixes the issues. #12