Open NSC508 opened 1 week ago
Hi @NSC508,
Currently, angle_emb only outputs Spearman's correlation when a validation set is provided to the fit() function. We will consider adding evaluation loss output in the next version.
Thank you!
@SeanLee97 Thank you for the quick response! I see, understood. Follow up question in the meantime then - is there a way for me to manually calculate loss for a given dataset of format Dataset.A? My idea was to use the AngleDataCollator
to tokenize my dataset, then pass it through the model backbone to get the loss. However, this does not work
collator = AngleDataCollator(angle.tokenizer, padding=True, return_tensors='pt', max_length=angle.max_length, filter_duplicate=True, coword_random_mask_rate=0
outputs = angle.backbone(collator(eval_ds)['input_ids].to('cuda'), labels=torch.tensor(eval_ds['label']).to('cuda')
Hello Author, thank you so much for publishing this library - it's incredibly useful. I ran into an issue while finetuning the model for my downstream task, and was wondering if I am using the function wrong. I am following the documentation for custom training and the training runs as normal, but I would like to view the validation loss as well as train loss so that I can identify potential overfitting. However, the function only seems to be logging train loss. Is there any way for me to log the validation loss as well during training?