suolyer commited on
Commit
dcb4f85
1 Parent(s): f9cf421

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +2 -2
app.py CHANGED
@@ -446,7 +446,7 @@ class UniMCPredict:
446
  batch = [self.data_model.train_data.encode(
447
  sample) for sample in batch_data]
448
  batch = self.data_model.collate_fn(batch)
449
- batch = {k: v.to(self.model.model.device) for k, v in batch.items()}
450
  _, _, logits = self.model.model(**batch)
451
  soft_logits = torch.nn.functional.softmax(logits, dim=-1)
452
  logits = torch.argmax(soft_logits, dim=-1).detach().cpu().numpy()
@@ -702,7 +702,7 @@ def main():
702
  The core idea of UniMC is to convert the natural language understanding task into a multiple choice task, which allows the model to directly reuse the parameters of the MaskLM head by controlling the position encoding and attention mask. This enables UniMC to surpass 100 billion parameter models in zero-shot scenarios just by training with multiple choice datasets. In the Chinese dataset, UniMC also surpassed other models and won the first place in both FewCLUE and ZeroCLUE.
703
  """)
704
 
705
- st.info("Please input the following information「请输入以下信息...」")
706
  model_type = st.selectbox('Select task type「选择任务类型」',['Text classification「文本分类」','Sentiment「情感分析」','Similarity「语义匹配」','NLI 「自然语言推理」','Multiple Choice「多项式阅读理解」'])
707
  form = st.form("参数设置")
708
  if '中文' in language:
 
446
  batch = [self.data_model.train_data.encode(
447
  sample) for sample in batch_data]
448
  batch = self.data_model.collate_fn(batch)
449
+ batch = {k: v.to(self.model.device) for k, v in batch.items()}
450
  _, _, logits = self.model.model(**batch)
451
  soft_logits = torch.nn.functional.softmax(logits, dim=-1)
452
  logits = torch.argmax(soft_logits, dim=-1).detach().cpu().numpy()
 
702
  The core idea of UniMC is to convert the natural language understanding task into a multiple choice task, which allows the model to directly reuse the parameters of the MaskLM head by controlling the position encoding and attention mask. This enables UniMC to surpass 100 billion parameter models in zero-shot scenarios just by training with multiple choice datasets. In the Chinese dataset, UniMC also surpassed other models and won the first place in both FewCLUE and ZeroCLUE.
703
  """)
704
 
705
+ st.info("Please input the following information to experiencing UniMC「请输入以下信息开始体验 UniMC...」")
706
  model_type = st.selectbox('Select task type「选择任务类型」',['Text classification「文本分类」','Sentiment「情感分析」','Similarity「语义匹配」','NLI 「自然语言推理」','Multiple Choice「多项式阅读理解」'])
707
  form = st.form("参数设置")
708
  if '中文' in language: