I have a dataset with the following info

Image1 x1 x2 x3 y

Image2 x1 x2 x3 y ...

Where x1, x2 & x3 are categorical features. My goal is to extract features from the images and use those features combined with x1 x2 and x3 as an input to another model to classify y.

My question is the following. Assuming I use a pre-trained resnet model, should I retrain the last layer using my images and y's to fit the resnet better to my problem or should I directly use the features extracted from the resnet as an input to my other model without retraining resnet?


1 Answer 1


Most pre-trained ResNet are pretty capable of returning a rather informative feature representation, even without fine-tuning. This, surprisingly, seems to be the case also for image datasets that do not resemble ImageNet. For example, I've used the feature representation of ResNet18 (the most basic ResNet, if you will) of the MVTec dataset, and the representations were pretty impressive (in terms of how well they were classified later on in my pipeline).

Of course, one can usually expect better results when finetuned to a specific problem, but I would start off with the representation vectors as is, and see how they perform.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .