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

Merge pull request #2130 from huggingface/ignored-index-coherence

Browse Source 
[BREAKING CHANGE] Setting all ignored index to the PyTorch standard
This commit is contained in:
Thomas Wolf
2019-12-21 14:55:40 +01:00
committed by GitHub
parent 18601c3b6e dc667ce1a7
commit 6f68d559ab
17 changed files with 42 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/contrib/run_openai_gpt.py
View File

@@ -75,7 +75,7 @@ def pre_process_datasets(encoded_datasets, input_len, cap_length, start_token, d
n_batch = len(dataset)
input_ids = np.zeros((n_batch, 2, input_len), dtype=np.int64)
mc_token_ids = np.zeros((n_batch, 2), dtype=np.int64)
lm_labels = np.full((n_batch, 2, input_len), fill_value=-1, dtype=np.int64)
lm_labels = np.full((n_batch, 2, input_len), fill_value=-100, dtype=np.int64)
mc_labels = np.zeros((n_batch,), dtype=np.int64)
for i, (story, cont1, cont2, mc_label), in enumerate(dataset):
with_cont1 = [start_token] + story[:cap_length] + [delimiter_token] + cont1[:cap_length] + [clf_token]

10
examples/distillation/distiller.py
View File

@@ -112,7 +112,7 @@ class Distiller:
self.last_log = 0
self.ce_loss_fct = nn.KLDivLoss(reduction='batchmean')
self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
if self.alpha_mse > 0.:
self.mse_loss_fct = nn.MSELoss(reduction='sum')
if self.alpha_cos > 0.:
@@ -186,7 +186,7 @@ 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)
@@ -224,7 +224,7 @@ class Distiller:
_token_ids = _token_ids_mask * (probs == 0).long() + _token_ids_real * (probs == 1).long() + _token_ids_rand * (probs == 2).long()
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
@@ -246,7 +246,7 @@ 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)
@@ -254,7 +254,7 @@ 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

2
examples/run_lm_finetuning.py
View File

@@ -150,7 +150,7 @@ def mask_tokens(inputs, tokenizer, args):
special_tokens_mask = [tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()]
probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
masked_indices = torch.bernoulli(probability_matrix).bool()
labels[~masked_indices] = -1 # We only compute loss on masked tokens
labels[~masked_indices] = -100 # We only compute loss on masked tokens
# 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices

2
examples/utils_ner.py
View File

@@ -94,7 +94,7 @@ def convert_examples_to_features(examples,
pad_on_left=False,
pad_token=0,
pad_token_segment_id=0,
pad_token_label_id=-1,
pad_token_label_id=-100,
sequence_a_segment_id=0,
mask_padding_with_zero=True):
""" Loads a data file into a list of `InputBatch`s