Open TeddyPorfiris opened 1 month ago
There are numerous reasons why your model could perform well on one few-shot task and not on another. The good thing with few-shot learning is that since the volumes are small, it is easier to investigate image-wise what went wrong with the prediction, and why it was associated to a wrong class.
I suggest you use classic investigation tools (confusion matrix, etc...) and complete them with a visualization of poorly classified images and to what particular images from the support set they are the closest.
Hello! Thanks so much for easyfsl, it's fantastic. I am testing my Prototypical Network (trained on mini imagenet) with SupportSetFolder. When I test it on the folder I attached called dataset1 (containing photos from internet), I get very accurate results. But when I test it on the folder I attached called dataset2 (containing photos I took), I get very inaccurate results. If you could help me figure out why this is, I'd appreciate it so much. Thanks again.
Link to download MIN_model.pth: https://drive.google.com/file/d/1q6sfNYcYSTUJzEiHq1T-nJ5R31EZ8dio/view?usp=sharing dataset1.zip dataset1.zip dataset2.zip dataset2.zip