Alwin4Zhang

model generate with different batch size but get different results

2

### System Info I'm using MT5ForConditionalGeneration by transformers to generate summaries,but when I use the arguments below,I will get different results when using beam search + do_sample + top_k +...

postion_ids的定义

3

``` top_len = max(context_len, triple_len, goal_len) pos = list(range(top_len - context_len + 1, top_len + 1)) + list(range(top_len - triple_len + 1, top_len + 1)) if goal_len > 0: pos.extend(list(range(top_len...

使用DUEE 1.0数据 Epoch=200 前面的几个epoch argu的F1都非常低这是正常的嘛？ Train argu F1: 0.000011 100% 1498/1498 [00:13

```python # model_unilm.py def compute_attention_bias(self, segment_ids): """实现unilm的attention: bert encoder self_attention + look_ahead_mask attention""" idxs = torch.cumsum(segment_ids, dim=1) mask = idxs[:, None, :]