Hey there, fellow deep learning enthusiasts! I recently came across a Reddit post from someone struggling with f-AnoGAN, a network for anomaly detection. They were getting poor results, with ROC, AUC, and PR scores all hovering around 0.3. If you’re facing similar issues, this post is for you.
First, let’s break down the problem. f-AnoGAN is a powerful tool, but it can be finicky. The original poster had a dataset with 2242 normal images for training and 2242 normal images for testing, plus 3367 abnormal images. They followed the standard steps for training and testing, but still got poor results.
So, what’s going wrong? One possibility is that the model is overfitting or underfitting. Have you tried tweaking the hyperparameters or experimenting with different architectures? Another issue could be with the data itself – are the normal and abnormal images properly labeled and balanced?
If you’re stuck, I recommend checking out the original f-AnoGAN paper or the GitHub repository for inspiration. You might also want to try different evaluation metrics or techniques, like ensembling or transfer learning.
Lastly, don’t be afraid to ask for help! The deep learning community is all about collaboration and learning from each other. Share your experiences and insights, and let’s troubleshoot f-AnoGAN together.
What are your thoughts on f-AnoGAN? Have you had any success with it, or are you struggling like the original poster?