SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

fix #1223

Browse Source
This commit is contained in:
thomwolf
2019-09-17 10:25:06 +02:00
parent c88f05163d
commit 45de034bf8
1 changed files with 10 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
pytorch_transformers/modeling_xlnet.py
View File

@@ -743,8 +743,9 @@ class XLNetModel(XLNetPreTrainedModel):
if data_mask is not None: if data_mask is not None:
# all mems can be attended to # all mems can be attended to
mems_mask = torch.zeros([data_mask.shape[0], mlen, bsz]).to(data_mask) if mlen > 0:
data_mask = torch.cat([mems_mask, data_mask], dim=1) mems_mask = torch.zeros([data_mask.shape[0], mlen, bsz]).to(data_mask)
data_mask = torch.cat([mems_mask, data_mask], dim=1)
if attn_mask is None: if attn_mask is None:
attn_mask = data_mask[:, :, :, None] attn_mask = data_mask[:, :, :, None]
else: else:
@@ -755,7 +756,8 @@ class XLNetModel(XLNetPreTrainedModel):
if attn_mask is not None: if attn_mask is not None:
non_tgt_mask = -torch.eye(qlen).to(attn_mask) non_tgt_mask = -torch.eye(qlen).to(attn_mask)
non_tgt_mask = torch.cat([torch.zeros([qlen, mlen]).to(attn_mask), non_tgt_mask], dim=-1) if mlen > 0:
non_tgt_mask = torch.cat([torch.zeros([qlen, mlen]).to(attn_mask), non_tgt_mask], dim=-1)
non_tgt_mask = ((attn_mask + non_tgt_mask[:, :, None, None]) > 0).to(attn_mask) non_tgt_mask = ((attn_mask + non_tgt_mask[:, :, None, None]) > 0).to(attn_mask)
else: else:
non_tgt_mask = None non_tgt_mask = None
@@ -775,8 +777,11 @@ class XLNetModel(XLNetPreTrainedModel):
##### Segment embedding ##### Segment embedding
if token_type_ids is not None: if token_type_ids is not None:
# Convert `token_type_ids` to one-hot `seg_mat` # Convert `token_type_ids` to one-hot `seg_mat`
mem_pad = torch.zeros([mlen, bsz], dtype=torch.long, device=device) if mlen > 0:
cat_ids = torch.cat([mem_pad, token_type_ids], dim=0) mem_pad = torch.zeros([mlen, bsz], dtype=torch.long, device=device)
cat_ids = torch.cat([mem_pad, token_type_ids], dim=0)
else:
cat_ids = token_type_ids
# `1` indicates not in the same segment [qlen x klen x bsz] # `1` indicates not in the same segment [qlen x klen x bsz]
seg_mat = (token_type_ids[:, None] != cat_ids[None, :]).long() seg_mat = (token_type_ids[:, None] != cat_ids[None, :]).long()