Get the answers+confidence from a BLIP-2 model for a VQA task.

blip2_confidence_snippet.py

outputs = blip2.generate(
    pixel_values=pixel_values,
    input_ids=input_ids,
    attention_mask=attention_mask,
    do_sample=False,
    num_beams=5,
    max_new_tokens=10,
    min_length=1,
    length_penalty=-1,
    return_dict_in_generate=True,
    output_scores=True,
)

output_text = blip2_tokenizer.batch_decode(
    outputs.sequences, skip_special_tokens=True
)

output_scores = torch.exp(outputs.sequences_scores)