Revert erroneous fix

examples/contrib/run_openai_gpt.py

@@ -81,7 +81,7 @@ 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]

examples/distillation/distiller.py

@@ -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

examples/run_lm_finetuning.py

@@ -207,7 +207,7 @@ def mask_tokens(inputs: torch.Tensor, tokenizer: PreTrainedTokenizer, args) -> T
         padding_mask = labels.eq(tokenizer.pad_token_id)
         probability_matrix.masked_fill_(padding_mask, 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

src/transformers/modeling_albert.py

@@ -632,8 +632,8 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
         r"""
         masked_lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the masked language modeling loss.
-            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with
             labels in ``[0, ..., config.vocab_size]``
 
     Returns:

src/transformers/modeling_bert.py

@@ -846,8 +846,8 @@ class BertForPreTraining(BertPreTrainedModel):
         r"""
         masked_lm_labels (``torch.LongTensor`` of shape ``(batch_size, sequence_length)``, `optional`, defaults to :obj:`None`):
             Labels for computing the masked language modeling loss.
-            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
         next_sentence_label (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`, defaults to :obj:`None`):
             Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair (see :obj:`input_ids` docstring)
@@ -948,13 +948,13 @@ class BertForMaskedLM(BertPreTrainedModel):
         r"""
         masked_lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the masked language modeling loss.
-            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
         lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the left-to-right language modeling loss (next word prediction).
-            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Returns:
@@ -1015,7 +1015,7 @@ class BertForMaskedLM(BertPreTrainedModel):
         # 2. If `lm_labels` is provided we are in a causal scenario where we
         #    try to predict the next token for each input in the decoder.
         if masked_lm_labels is not None:
-            loss_fct = CrossEntropyLoss()  # -1 index = padding token
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
             outputs = (masked_lm_loss,) + outputs
 

src/transformers/modeling_ctrl.py

@@ -479,8 +479,8 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
-            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            Indices are selected in ``[-100, 0, ..., config.vocab_size]``
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Return:

src/transformers/modeling_distilbert.py

@@ -517,8 +517,8 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
         r"""
         masked_lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the masked language modeling loss.
-            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Returns:

src/transformers/modeling_gpt2.py

@@ -547,8 +547,8 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
-            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            Indices are selected in ``[-100, 0, ..., config.vocab_size]``
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Return:
@@ -655,7 +655,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
         mc_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size)`, `optional`, defaults to :obj:`None`)
             Labels for computing the multiple choice classification loss.

src/transformers/modeling_openai.py

@@ -516,8 +516,8 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
-            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            Indices are selected in ``[-100, 0, ..., config.vocab_size]``
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Return:
@@ -621,7 +621,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
         mc_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size)`, `optional`, defaults to :obj:`None`)
             Labels for computing the multiple choice classification loss.

src/transformers/modeling_roberta.py

@@ -200,8 +200,8 @@ class RobertaForMaskedLM(BertPreTrainedModel):
         r"""
         masked_lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the masked language modeling loss.
-            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
 
     Returns:

src/transformers/modeling_t5.py

@@ -802,8 +802,8 @@ class T5WithLMHeadModel(T5PreTrainedModel):
     r"""
         **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the masked language modeling loss.
-            Indices should either be in ``[0, ..., config.vocab_size]`` or -1 (see ``input_ids`` docstring).
-            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            Indices should either be in ``[0, ..., config.vocab_size]`` or -100 (see ``input_ids`` docstring).
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``.
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
@@ -906,7 +906,7 @@ class T5WithLMHeadModel(T5PreTrainedModel):
         if lm_labels is not None:
             shift_logits = lm_logits[..., :-1, :].contiguous()
             shift_labels = lm_labels[..., 1:].contiguous()
-            loss_fct = CrossEntropyLoss()
+            loss_fct = CrossEntropyLoss(ignore_index=-100)
             loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
             decoder_outputs = (
                 loss,

src/transformers/modeling_transfo_xl.py

@@ -858,8 +858,8 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
-            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            Indices are selected in ``[-100, 0, ..., config.vocab_size]``
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Return:

src/transformers/modeling_xlm.py

@@ -667,8 +667,8 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
-            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            Indices are selected in ``[-100, 0, ..., config.vocab_size]``
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Return:

src/transformers/modeling_xlnet.py

@@ -993,8 +993,8 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
             Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
-            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-1`` are ignored (masked), the loss is only
+            Indices are selected in ``[-100, 0, ..., config.vocab_size]``
+            All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
 
     Return: