I finetuned a VisualBertForPreTraining model in order to perform a binary classification task using the Hateful Memes Dataset. However, my issue is that among the model’s outputs, obtained in the following way:
outputs = model( input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, visual_embeds=visual_embeds, visual_attention_mask=visual_attention_mask, visual_token_type_ids=visual_token_type_ids, labels=token_labels )
there are no probabilities for the classification. Therefore, I believe it’s necessary to construct a classification head for the prediction_logits present in the model’s output.
Specifically, what I obtain as
output.prediction_logits during the evaluation phase, for a batch of 16 inputs, is a tensor of size
torch.Size([16, 150, 30522]).
How can I transform this tensor into probabilities?
Thank you very much in advance for your help.