Model fails with index out of range in self when used on a text-classification pipeline

[riolcrt]

Steps to reproduce

nlp = pipeline("text-classification", "pysentimiento/robertuito-sentiment-analysis")
nlp("long text"*300)

IndexError                                Traceback (most recent call last)
Input In [23], in <cell line: 1>()
----> 1 nlp("long text"*300)

File /opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py:125, in TextClassificationPipeline.__call__(self, *args, **kwargs)
     [92](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=91) def __call__(self, *args, **kwargs):
     [93](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=92)     """
     [94](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=93)     Classify the text(s) given as inputs.
     [95](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=94) 
   (...)
    [123](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=122)         If `self.return_all_scores=True`, one such dictionary is returned per label.
    [124](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=123)     """
--> [125](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=124)     result = super().__call__(*args, **kwargs)
    [126](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=125)     if isinstance(args[0], str):
    [127](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=126)         # This pipeline is odd, and return a list when single item is run
    [128](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=127)         return [result]

File /opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py:1027, in Pipeline.__call__(self, inputs, num_workers, batch_size, *args, **kwargs)
   [1025](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=1024)     return self.iterate(inputs, preprocess_params, forward_params, postprocess_params)
   [1026](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=1025) else:
-> [1027](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=1026)     return self.run_single(inputs, preprocess_params, forward_params, postprocess_params)

File /opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py:1034, in Pipeline.run_single(self, inputs, preprocess_params, forward_params, postprocess_params)
   [1032](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=1031) def run_single(self, inputs, preprocess_params, forward_params, postprocess_params):
   [1033](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=1032)     model_inputs = self.preprocess(inputs, **preprocess_params)
-> [1034](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=1033)     model_outputs = self.forward(model_inputs, **forward_params)
   [1035](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=1034)     outputs = self.postprocess(model_outputs, **postprocess_params)
   [1036](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=1035)     return outputs

File /opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py:944, in Pipeline.forward(self, model_inputs, **forward_params)
    [942](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=941)     with inference_context():
    [943](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=942)         model_inputs = self._ensure_tensor_on_device(model_inputs, device=self.device)
--> [944](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=943)         model_outputs = self._forward(model_inputs, **forward_params)
    [945](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=944)         model_outputs = self._ensure_tensor_on_device(model_outputs, device=torch.device("cpu"))
    [946](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/base.py?line=945) else:

File /opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py:137, in TextClassificationPipeline._forward(self, model_inputs)
    [136](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=135) def _forward(self, model_inputs):
--> [137](file:///opt/conda/lib/python3.8/site-packages/transformers/pipelines/text_classification.py?line=136)     return self.model(**model_inputs)

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1102, in Module._call_impl(self, *input, **kwargs)
   [1098](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1097) # If we don't have any hooks, we want to skip the rest of the logic in
   [1099](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1098) # this function, and just call forward.
   [1100](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1099) if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   [1101](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1100)         or _global_forward_hooks or _global_forward_pre_hooks):
-> [1102](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1101)     return forward_call(*input, **kwargs)
   [1103](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1102) # Do not call functions when jit is used
   [1104](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1103) full_backward_hooks, non_full_backward_hooks = [], []

File /opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py:1204, in RobertaForSequenceClassification.forward(self, input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, labels, output_attentions, output_hidden_states, return_dict)
   [1196](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1195) r"""
   [1197](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1196) labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
   [1198](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1197)     Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
   [1199](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1198)     config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
   [1200](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1199)     `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
   [1201](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1200) """
   [1202](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1201) return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-> [1204](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1203) outputs = self.roberta(
   [1205](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1204)     input_ids,
   [1206](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1205)     attention_mask=attention_mask,
   [1207](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1206)     token_type_ids=token_type_ids,
   [1208](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1207)     position_ids=position_ids,
   [1209](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1208)     head_mask=head_mask,
   [1210](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1209)     inputs_embeds=inputs_embeds,
   [1211](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1210)     output_attentions=output_attentions,
   [1212](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1211)     output_hidden_states=output_hidden_states,
   [1213](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1212)     return_dict=return_dict,
   [1214](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1213) )
   [1215](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1214) sequence_output = outputs[0]
   [1216](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=1215) logits = self.classifier(sequence_output)

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1102, in Module._call_impl(self, *input, **kwargs)
   [1098](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1097) # If we don't have any hooks, we want to skip the rest of the logic in
   [1099](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1098) # this function, and just call forward.
   [1100](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1099) if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   [1101](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1100)         or _global_forward_hooks or _global_forward_pre_hooks):
-> [1102](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1101)     return forward_call(*input, **kwargs)
   [1103](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1102) # Do not call functions when jit is used
   [1104](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1103) full_backward_hooks, non_full_backward_hooks = [], []

File /opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py:843, in RobertaModel.forward(self, input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, encoder_hidden_states, encoder_attention_mask, past_key_values, use_cache, output_attentions, output_hidden_states, return_dict)
    [836](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=835) # Prepare head mask if needed
    [837](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=836) # 1.0 in head_mask indicate we keep the head
    [838](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=837) # attention_probs has shape bsz x n_heads x N x N
    [839](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=838) # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
    [840](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=839) # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
    [841](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=840) head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
--> [843](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=842) embedding_output = self.embeddings(
    [844](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=843)     input_ids=input_ids,
    [845](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=844)     position_ids=position_ids,
    [846](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=845)     token_type_ids=token_type_ids,
    [847](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=846)     inputs_embeds=inputs_embeds,
    [848](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=847)     past_key_values_length=past_key_values_length,
    [849](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=848) )
    [850](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=849) encoder_outputs = self.encoder(
    [851](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=850)     embedding_output,
    [852](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=851)     attention_mask=extended_attention_mask,
   (...)
    [860](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=859)     return_dict=return_dict,
    [861](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=860) )
    [862](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=861) sequence_output = encoder_outputs[0]

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1102, in Module._call_impl(self, *input, **kwargs)
   [1098](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1097) # If we don't have any hooks, we want to skip the rest of the logic in
   [1099](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1098) # this function, and just call forward.
   [1100](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1099) if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   [1101](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1100)         or _global_forward_hooks or _global_forward_pre_hooks):
-> [1102](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1101)     return forward_call(*input, **kwargs)
   [1103](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1102) # Do not call functions when jit is used
   [1104](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1103) full_backward_hooks, non_full_backward_hooks = [], []

File /opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py:136, in RobertaEmbeddings.forward(self, input_ids, token_type_ids, position_ids, inputs_embeds, past_key_values_length)
    [134](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=133) embeddings = inputs_embeds + token_type_embeddings
    [135](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=134) if self.position_embedding_type == "absolute":
--> [136](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=135)     position_embeddings = self.position_embeddings(position_ids)
    [137](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=136)     embeddings += position_embeddings
    [138](file:///opt/conda/lib/python3.8/site-packages/transformers/models/roberta/modeling_roberta.py?line=137) embeddings = self.LayerNorm(embeddings)

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1102, in Module._call_impl(self, *input, **kwargs)
   [1098](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1097) # If we don't have any hooks, we want to skip the rest of the logic in
   [1099](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1098) # this function, and just call forward.
   [1100](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1099) if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   [1101](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1100)         or _global_forward_hooks or _global_forward_pre_hooks):
-> [1102](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1101)     return forward_call(*input, **kwargs)
   [1103](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1102) # Do not call functions when jit is used
   [1104](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py?line=1103) full_backward_hooks, non_full_backward_hooks = [], []

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/sparse.py:158, in Embedding.forward(self, input)
    [157](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/sparse.py?line=156) def forward(self, input: Tensor) -> Tensor:
--> [158](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/sparse.py?line=157)     return F.embedding(
    [159](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/sparse.py?line=158)         input, self.weight, self.padding_idx, self.max_norm,
    [160](file:///opt/conda/lib/python3.8/site-packages/torch/nn/modules/sparse.py?line=159)         self.norm_type, self.scale_grad_by_freq, self.sparse)

File /opt/conda/lib/python3.8/site-packages/torch/nn/functional.py:2044, in embedding(input, weight, padding_idx, max_norm, norm_type, scale_grad_by_freq, sparse)
   [2038](file:///opt/conda/lib/python3.8/site-packages/torch/nn/functional.py?line=2037)     # Note [embedding_renorm set_grad_enabled]
   [2039](file:///opt/conda/lib/python3.8/site-packages/torch/nn/functional.py?line=2038)     # XXX: equivalent to
   [2040](file:///opt/conda/lib/python3.8/site-packages/torch/nn/functional.py?line=2039)     # with torch.no_grad():
   [2041](file:///opt/conda/lib/python3.8/site-packages/torch/nn/functional.py?line=2040)     #   torch.embedding_renorm_
   [2042](file:///opt/conda/lib/python3.8/site-packages/torch/nn/functional.py?line=2041)     # remove once script supports set_grad_enabled
   [2043](file:///opt/conda/lib/python3.8/site-packages/torch/nn/functional.py?line=2042)     _no_grad_embedding_renorm_(weight, input, max_norm, norm_type)
-> [2044](file:///opt/conda/lib/python3.8/site-packages/torch/nn/functional.py?line=2043) return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)

IndexError: index out of range in self

[finiteautomata]

Hmmmm, strange. As far as I can see on tokenizer.json, there is a max_length set. Let me check it

[finiteautomata]

Do not understand why this fails, as the tokenizer has the model_max_len property set. Please report on transformers.