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

Revert erroneous fix

Browse Source
This commit is contained in:
Lysandre
2020-02-04 16:31:07 -05:00
parent 1ebfeb7946
commit 3bf5417258
14 changed files with 37 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
examples/distillation/distiller.py
View File

@@ -109,7 +109,7 @@ class Distiller:
self.last_log = 0
self.ce_loss_fct = nn.KLDivLoss(reduction="batchmean")
self.lm_loss_fct = nn.CrossEntropyLoss()
self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
if self.alpha_mse > 0.0:
self.mse_loss_fct = nn.MSELoss(reduction="sum")
if self.alpha_cos > 0.0:
@@ -200,7 +200,7 @@ class Distiller:
-------
token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -1 where there is nothing to predict.
mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -100 where there is nothing to predict.
"""
token_ids, lengths = batch
token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
@@ -244,7 +244,7 @@ class Distiller:
)
token_ids = token_ids.masked_scatter(pred_mask, _token_ids)
mlm_labels[~pred_mask] = -1 # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
mlm_labels[~pred_mask] = -100 # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
# sanity checks
assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
@@ -265,7 +265,7 @@ class Distiller:
-------
token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -1 where there is nothing to predict.
clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -100 where there is nothing to predict.
"""
token_ids, lengths = batch
token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
@@ -273,7 +273,7 @@ class Distiller:
attn_mask = torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None]
clm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
clm_labels[~attn_mask] = -1 # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
clm_labels[~attn_mask] = -100 # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
# sanity checks
assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size