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

scatter_ and scatter

Browse Source
This commit is contained in:
VictorSanh
2018-11-03 10:27:38 -04:00
parent a6efe1235f
commit 2c55568c40
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
modeling_pytorch.py
View File

@@ -502,7 +502,7 @@ class BertForQuestionAnswering(nn.Module):
def compute_loss(logits, positions): def compute_loss(logits, positions):
max_position = positions.max().item() max_position = positions.max().item()
one_hot = torch.FloatTensor(batch_size, max(max_position, seq_length) +1).zero_() one_hot = torch.FloatTensor(batch_size, max(max_position, seq_length) +1).zero_()
one_hot = one_hot.scatter(1, positions.cpu(), 1) # Second argument need to be LongTensor and not cuda.LongTensor one_hot = one_hot.scatter_(1, positions.cpu(), 1) # Second argument need to be LongTensor and not cuda.LongTensor
one_hot = one_hot[:, :seq_length].to(input_ids.device) one_hot = one_hot[:, :seq_length].to(input_ids.device)
log_probs = nn.functional.log_softmax(logits, dim = -1).view(batch_size, seq_length) log_probs = nn.functional.log_softmax(logits, dim = -1).view(batch_size, seq_length)
loss = -torch.mean(torch.sum(one_hot*log_probs), dim = -1) loss = -torch.mean(torch.sum(one_hot*log_probs), dim = -1)