question about neg_out from blip pretrain files

[jumbokun]

Hello Lavis team,

I kinda don't get why this snippet would deliever a negative text for each image

        # select a negative text for each image
        text_ids_neg = []
        text_atts_neg = []
        for b in range(bs):
            neg_idx = torch.multinomial(weights_i2t[b], 1).item()
            text_ids_neg.append(encoder_input_ids[neg_idx])
            text_atts_neg.append(text.attention_mask[neg_idx])

so encoder_input_ids is simply input ids of text..? You picked some token ids from that instead of picking up from queue?.. Could you please explain it a little bit?