[PyTorch newbie] I have a dataset of images, where for each image we have 20+ attributes. I need to train a classifier which takes images as input, and returns the predicted attributes as output. As a first step, I would like to fine-tune ResNet. I need to complete the task using PyTorch. How do you recommend to proceed? Thank you!
If you have an architecture that returns logits (often the case), you wouldn’t need to change the model architecture instead. I would recommend finetuning via a multilabel loss function e.g., MultiLabelSoftMarginLoss — PyTorch 1.11.0 documentation and iterating on the model/data from there.
Thanks for the hint! I’ll check what’s the loss function of ResNet and I’ll try to change is as you suggested.