AzzedineNed
commented
2 months ago Hello. To clarify, I have not used "FID" but rather "pr50k3_full" for evaluation. Regarding your concern about the high score, there are only a couple of factors that I can think of:
1- Dataset Imbalance: The MEL class has significantly fewer images compared to the nevus class. This imbalance can lead to challenges in training neural networks like StyleGAN3 effectively. With fewer examples, the model may struggle to learn diverse and representative features, potentially impacting the quality of generated images.
2-Initialization of Training: Another factor to consider is the initialization of training weights. Starting with random weights can affect how the model learns and adapts to the dataset. Proper initialization and possibly pre-training on related tasks could help mitigate this issue.
Hello~I have a question that's been bothering me for half a month now; I'd like to seek your advice on it. Why is it that when I use MEL images with hair removal and a size of 256x256 for StyleGAN3 training, the resulting FID score is very high, exceeding 500 right from the start? According to your training, the FID should begin in the single digits.I'd appreciate it if you could take the time to address this lingering doubt of mine when it's convenient for you. Thank you.