Hi
First of all, thank you for this course. I have a remark about lab 4 of fine tuning pretrained models for text classification. In the last step you are partially fine tuning the model and compare it with the fully fine tuned model, which results in “the same performance”. However, the partially fine tuned model is initialized with the weights of the fully fine tuned model, instead of starting from the pretrained weights, so that is not really a fair comparison.
Instead of further training the fully fine tuned model I think you should redefine the bert_model in the last section before training the partial one:
bert_model, bert_tokenizer = helper_utils.load_bert(model_path, num_classes=2)
Using these settings I get these results: (epochs and loss is still the same, just the initial weights are different)
Which also alters the conclusion of this Lab “This perfectly illustrates the core benefit of partial fine-tuning: achieving comparable, if not better, performance while saving valuable time and computational resources.” You sacrifice a bit of performance for efficiency.