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

Merge pull request #2303 from patrickvonplaten/fix_error_with_repetition_penalty

Browse Source 
fix repetition penalty error in modeling_utils.py
This commit is contained in:
Thomas Wolf
2019-12-25 22:39:20 +01:00
committed by GitHub
parent 0412f3d929 18e5bdbec5
commit aeef4823ab
1 changed files with 10 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/transformers/modeling_utils.py
View File

@@ -728,7 +728,11 @@ class PreTrainedModel(nn.Module):
if repetition_penalty != 1.0: if repetition_penalty != 1.0:
for i in range(batch_size): for i in range(batch_size):
for previous_tokens in set(input_ids[i].tolist()): for previous_tokens in set(input_ids[i].tolist()):
next_token_logits[i, previous_tokens] /= repetition_penalty # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
if next_token_logits[i, previous_tokens] < 0:
next_token_logits[i, previous_tokens] *= repetition_penalty
else:
next_token_logits[i, previous_tokens] /= repetition_penalty
if do_sample: if do_sample:
# Temperature (higher temperature => more likely to sample low probability tokens) # Temperature (higher temperature => more likely to sample low probability tokens)
@@ -807,7 +811,11 @@ class PreTrainedModel(nn.Module):
if repetition_penalty != 1.0: if repetition_penalty != 1.0:
for i in range(batch_size * num_beams): for i in range(batch_size * num_beams):
for previous_tokens in set(input_ids[i].tolist()): for previous_tokens in set(input_ids[i].tolist()):
scores[i, previous_tokens] /= repetition_penalty # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
if scores[i, previous_tokens] < 0:
scores[i, previous_tokens] *= repetition_penalty
else:
scores[i, previous_tokens] /= repetition_penalty
if do_sample: if do_sample:
# Temperature (higher temperature => more likely to sample low probability tokens) # Temperature (higher temperature => more likely to sample low probability tokens)