def play()

in src/model/self_play_qgen_vilbert.py [0:0]

• 59 lines of code
• 13 McCabe index (conditional complexity)

    def play(self, qgen_img_feats, qgen_bboxs, tgt_cat, tgt_bbox, cats, bboxs, bboxs_mask, 
             sos_token, pad_token, eoq_token, eod_token, 
             answer2id, answer2token, max_q_len, greedy=True, max_turns=8, answer_as_sos=True):
        device = qgen_img_feats.device
        batch_size = qgen_img_feats.size(0)
        num_bboxs = qgen_img_feats.size(1)
        end_of_dialog = torch.zeros(batch_size).bool().to(device)
        sos = torch.zeros(batch_size, 1).fill_(sos_token).long().to(device)
        last_wrd = torch.zeros(batch_size).fill_(sos_token).long().to(device)
        last_state = None


        pi = self.qgen.state_handler.init_state(batch_size, num_bboxs, device)
        
        dialog = [torch.LongTensor(0).to(device) for _ in range(batch_size)]
        q_log = [[] for _ in range(batch_size)]
        a_log = [[] for _ in range(batch_size)]
        a_conf_log = [[] for _ in range(batch_size)]
        for turn in range(max_turns):
            # print(turn, ':', pi[0].argmax())
            q, q_len, state, end_of_dialog_next = self.qgen.generate_sentence(
                last_wrd, qgen_img_feats, qgen_bboxs, eoq_token, eod_token, end_of_dialog, 
                max_q_len=max_q_len, pi=pi, last_state=last_state, greedy=greedy
            )

            pad_q = pad_sequence(q, batch_first=True, padding_value=pad_token)
            # HACK: length == 0 can not forward in RNN
            fake_q_len = q_len.clone()
            fake_q_len[q_len == 0] = 1
            a = self.oracle(pad_q, tgt_cat, tgt_bbox, fake_q_len)
            a_confidence = nn.functional.softmax(a, dim=-1)
            a_idx = a.argmax(dim=-1)
            a = oracle_output_to_answer_token(a_idx, answer2id, answer2token)
            for b in range(batch_size):
                if not end_of_dialog[b]:
                    _q = q[b][:q_len[b]]
                    q_log[b].append(_q)
                    dialog[b] = torch.cat([dialog[b], _q])
                if not end_of_dialog_next[b]:
                    _a = a[b].view(-1)
                    a_log[b].append(_a)
                    a_conf_log[b].append(a_confidence[b, a_idx[b]])
                    dialog[b] = torch.cat([dialog[b], _a])

            if end_of_dialog_next.sum().item() == batch_size:
                break
            end_of_dialog = end_of_dialog_next
            if answer_as_sos:
                last_wrd = a
            last_state = state

            #pi = self.qgen.refresh_pi(pi, a, last_state[0,0], obj_repr, input_token=True)
            txt_attn_mask = [[1] * (ql+1) + [0] * (q_len.max().item() - ql) for ql in q_len]
            txt_attn_mask = torch.tensor(txt_attn_mask).to(device)
            pi, _ = self.qgen.state_handler.forward_turn(
                torch.cat([sos, pad_q], dim=-1), 
                a_idx, 
                None, # cats 
                qgen_img_feats,
                qgen_bboxs, 
                curr_state=pi,
                attention_mask=txt_attn_mask,
                update_vilbert=False,
            )
            
        dial_len = torch.LongTensor([len(dial) for dial in dialog]).to(device)
        dial_pad = pad_sequence(dialog, batch_first=True, padding_value=pad_token)
        guess = self.guesser(dial_pad, dial_len, cats, bboxs, bboxs_mask)
        return guess, dialog, q_log, a_log, a_conf_log