#we have concatenated the negative features and did a forward pass and can slice out the corresponding parts of of cap_logits and targets
  #PAD=0
  #SOS=1
  #EOS=2
  cap_crit = nn.CrossEntropyLoss()
  pos_cap_logits = cap_logits[:batch_size,:]
  target_cap = target_cap.long().view(-1)
  pos_target = target_cap[:batch_size]
  orig_pos_target = pos_target
  mask =pos_target.ge(1.1)# mask out padding
  pos_target=torch.masked_select(pos_target, mask) #(batch,)
  pos_cap_logits=pos_cap_logits[mask,:] #(batch, vocab)
  pos_caption_loss = cap_crit(pos_cap_logits, pos_target.long().view(-1))

  visual_unpaired_cap_logits = cap_logits[batch_size: batch_size*2, :]
  vp_target = target_cap[batch_size: batch_size*2]
  mask =vp_target.ge(1.1)# mask out padding
  vp_target=torch.masked_select(vp_target, mask) #(batch,)
  visual_unpaired_cap_logits=visual_unpaired_cap_logits[mask,:] #(batch, vocab)
  vp_caption_loss = cap_crit(visual_unpaired_cap_logits, vp_target.long().view(-1))

  lang_unpaired_cap_logits = cap_logits[batch_size*2:, :]
  lup_target = target_cap[batch_size*2:]
  mask =lup_target.ge(1.1)# mask out padding
  lup_target=torch.masked_select(lup_target, mask) #(batch,)
  lang_unpaired_cap_logits=lang_unpaired_cap_logits[mask,:] #(batch, vocab)
  lup_caption_loss = cap_crit(lang_unpaired_cap_logits, lup_target.long().view(-1))
  margin=1.0
  visual_rank_weight = 1.0
  lang_rank_weight = 0.1
  visual_rank_loss = visual_rank_weight * torch.clamp(margin + visual_unpaired - pos_caption_loss, 0)
  lang_rank_loss = lang_rank_weight * torch.clamp(margin + lang_unpaired - pos_caption_loss, 0)
  mmi_loss = (visual_rank_loss + lang_rank_loss)

  generation_loss = pos_caption_loss + mmi_loss