My bad, I missed the top blog link with more complete code examples, maybe that would help me https://github.com/ariG23498/timm-wrapper-examples/blob/main/%2304_sft.ipynb
Trainer directly in custom script, I made a Colab notebook that illustrates what I'm currently trying, in the prediction test at the end the labels are not food:
https://colab.research.google.com/drive/14jTpetYR61B6EVoJ6o8_B8gi6-SiizCA?usp=sharing
Very excited for this, thank you! We were just about to switch to timm and this makes it even easier!
Maybe a simple naive question, I'm trying to put together a demo training script that loads the base model from:TimmWrapperForImageClassification.from_pretrained("timm/mobilenetv4_conv_medium.e500_r256_in1k").to("cuda")
but then trains on the food101 dataset, just to illustrate training on a new custom dataset. Training works fine, but inference return animal names as labels.
Should I be setting: label2id, num_labels, id2label, params when I load the finetuned model? It seems to store the data somewhere during training, but trainingarguments don't allow me to set TypeError: TrainingArguments.__init__() got an unexpected keyword argument 'label2id'
