SHAP Explainer throws an error while using transformers model for abstractive summarization

[SVC04]

Issue Description

Hello,
I am trying to generate an explanation of abstractive text summarization output for a long piece of text input. I have been using various transformers models e.g. Bigbird-Pegasus, LongT5, Lonformer Encoder Decoder etc. But none of it worked and throws different errors. Following is the piece of code I used.

I get the TypeError. 'int object is not callable'.

Minimal Reproducible Example

LONG_ARTICLE = """"anxiety affects quality of life in those living
with parkinson 's disease ( pd ) more so than
overall cognitive status , motor deficits , apathy
, and depression [ 13 ] . although anxiety and
depression are often related and coexist in pd
patients , recent research suggests that anxiety
rather than depression is the most prominent and
prevalent mood disorder in pd [ 5 , 6 ] . yet ,
our current understanding of anxiety and its
impact on cognition in pd , as well as its neural
basis and best treatment practices , remains
meager and lags far behind that of depression ."""

import torch
from transformers import AutoTokenizer, LongT5ForConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained("Stancld/longt5-tglobal-large-16384-pubmed-3k_steps")
model = LongT5ForConditionalGeneration.from_pretrained("Stancld/longt5-tglobal-large-16384-pubmed-3k_steps", return_dict_in_generate=True).to("cuda")
import shap

explainer = shap.Explainer(model, tokenizer)
shap_values = explainer(s)

### Traceback

```shell
TypeError                                 Traceback (most recent call last)
Cell In [23], line 1
----> 1 shap_values = explainer(s)

File /usr/local/lib/python3.9/dist-packages/shap/explainers/_partition.py:125, in PartitionExplainer.__call__(self, max_evals, fixed_context, main_effects, error_bounds, batch_size, outputs, silent, *args)
    121 def __call__(self, *args, max_evals=500, fixed_context=None, main_effects=False, error_bounds=False, batch_size="auto",
    122              outputs=None, silent=False):
    123     """ Explain the output of the model on the given arguments.
    124     """
--> 125     return super().__call__(
    126         *args, max_evals=max_evals, fixed_context=fixed_context, main_effects=main_effects, error_bounds=error_bounds, batch_size=batch_size,
    127         outputs=outputs, silent=silent
    128     )

File /usr/local/lib/python3.9/dist-packages/shap/explainers/_explainer.py:267, in Explainer.__call__(self, max_evals, main_effects, error_bounds, batch_size, outputs, silent, *args, **kwargs)
    265     feature_names = [[] for _ in range(len(args))]
    266 for row_args in show_progress(zip(*args), num_rows, self.__class__.__name__+" explainer", silent):
--> 267     row_result = self.explain_row(
    268         *row_args, max_evals=max_evals, main_effects=main_effects, error_bounds=error_bounds,
    269         batch_size=batch_size, outputs=outputs, silent=silent, **kwargs
    270     )
    271     values.append(row_result.get("values", None))
    272     output_indices.append(row_result.get("output_indices", None))

File /usr/local/lib/python3.9/dist-packages/shap/explainers/_partition.py:150, in PartitionExplainer.explain_row(self, max_evals, main_effects, error_bounds, batch_size, outputs, silent, fixed_context, *row_args)
    148 # if not fixed background or no base value assigned then compute base value for a row
    149 if self._curr_base_value is None or not getattr(self.masker, "fixed_background", False):
--> 150     self._curr_base_value = fm(m00.reshape(1, -1), zero_index=0)[0] # the zero index param tells the masked model what the baseline is
    151 f11 = fm(~m00.reshape(1, -1))[0]
    153 if callable(self.masker.clustering):

File /usr/local/lib/python3.9/dist-packages/shap/utils/_masked_model.py:69, in MaskedModel.__call__(self, masks, zero_index, batch_size)
     66         return self._full_masking_call(full_masks, zero_index=zero_index, batch_size=batch_size)
     68 else:
---> 69     return self._full_masking_call(masks, batch_size=batch_size)

File /usr/local/lib/python3.9/dist-packages/shap/utils/_masked_model.py:146, in MaskedModel._full_masking_call(self, masks, zero_index, batch_size)
    143         all_masked_inputs[i].append(v)
    145 joined_masked_inputs = tuple([np.concatenate(v) for v in all_masked_inputs])
--> 146 outputs = self.model(*joined_masked_inputs)
    147 _assert_output_input_match(joined_masked_inputs, outputs)
    148 all_outputs.append(outputs)

File /usr/local/lib/python3.9/dist-packages/shap/models/_model.py:28, in Model.__call__(self, *args)
     27 def __call__(self, *args):
---> 28     out = self.inner_model(*args)
     29     is_tensor = safe_isinstance(out, "torch.Tensor")
     30     out = out.cpu().detach().numpy() if is_tensor else np.array(out)

File /usr/local/lib/python3.9/dist-packages/torch/nn/modules/module.py:1130, in Module._call_impl(self, *input, **kwargs)
   1126 # If we don't have any hooks, we want to skip the rest of the logic in
   1127 # this function, and just call forward.
   1128 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1129         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1130     return forward_call(*input, **kwargs)
   1131 # Do not call functions when jit is used
   1132 full_backward_hooks, non_full_backward_hooks = [], []

File /usr/local/lib/python3.9/dist-packages/transformers/models/longt5/modeling_longt5.py:1991, in LongT5ForConditionalGeneration.forward(self, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, head_mask, decoder_head_mask, cross_attn_head_mask, encoder_outputs, past_key_values, inputs_embeds, decoder_inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, return_dict)
   1988 # Encode if needed (training, first prediction pass)
   1989 if encoder_outputs is None:
   1990     # Convert encoder inputs in embeddings if needed
-> 1991     encoder_outputs = self.encoder(
   1992         input_ids=input_ids,
   1993         attention_mask=attention_mask,
   1994         inputs_embeds=inputs_embeds,
   1995         head_mask=head_mask,
   1996         output_attentions=output_attentions,
   1997         output_hidden_states=output_hidden_states,
   1998         return_dict=return_dict,
   1999     )
   2000 elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
   2001     encoder_outputs = BaseModelOutput(
   2002         last_hidden_state=encoder_outputs[0],
   2003         hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
   2004         attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
   2005     )

File /usr/local/lib/python3.9/dist-packages/torch/nn/modules/module.py:1130, in Module._call_impl(self, *input, **kwargs)
   1126 # If we don't have any hooks, we want to skip the rest of the logic in
   1127 # this function, and just call forward.
   1128 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1129         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1130     return forward_call(*input, **kwargs)
   1131 # Do not call functions when jit is used
   1132 full_backward_hooks, non_full_backward_hooks = [], []

File /usr/local/lib/python3.9/dist-packages/transformers/models/longt5/modeling_longt5.py:1396, in LongT5Stack.forward(self, input_ids, attention_mask, encoder_hidden_states, encoder_attention_mask, inputs_embeds, head_mask, cross_attn_head_mask, past_key_values, use_cache, output_attentions, output_hidden_states, return_dict)
   1392     raise ValueError(
   1393         f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time"
   1394     )
   1395 elif input_ids is not None:
-> 1396     input_shape = input_ids.size()
   1397     input_ids = input_ids.view(-1, input_shape[-1])
   1398 elif inputs_embeds is not None:

TypeError: 'int' object is not callable

```​

Expected Behavior

No response

Bug report checklist

• I have checked that this issue has not already been reported.
• I have confirmed this bug exists on the latest release of shap.
• I have confirmed this bug exists on the master branch of shap.
• I'd be interested in making a PR to fix this bug

Installed Versions

Latest version used from github repo.

[CloseChoice]

I can reproduce your example on master but do not think that this is really a shap issue here. The following lines already result in similar errors:

>>> model([LONG_ARTICLE])
AttributeError: 'list' object has no attribute 'size'
>>> model(LONG_ARTICLE)
AttributeError: 'str' object has no attribute 'size'

Can you somehow get this to run so that the model actually returns something?

[SVC04]

I have tried but it didn't work.

[CloseChoice]

Okay, this indicates that we do not really have a shap error here since the model is not running at all.

[SVC04]

The hugging face model works well and generates the output summary.
I have also tried bigbird pegasus model in place of this model it did not work either.

LONG_ARTICLE = """"anxiety affects quality of life in those living
with parkinson 's disease ( pd ) more so than
overall cognitive status , motor deficits , apathy
, and depression [ 13 ] . although anxiety and
depression are often related and coexist in pd
patients , recent research suggests that anxiety
rather than depression is the most prominent and
prevalent mood disorder in pd [ 5 , 6 ] . yet ,
our current understanding of anxiety and its
impact on cognition in pd , as well as its neural
basis and best treatment practices , remains
meager and lags far behind that of depression .
overall , neuropsychiatric symptoms in pd have
been shown to be negatively associated with
cognitive performance .


s = [LONG_ARTICLE]

from transformers import BigBirdPegasusForConditionalGeneration, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("google/bigbird-pegasus-large-pubmed")
model = BigBirdPegasusForConditionalGeneration.from_pretrained("google/bigbird-pegasus-large-pubmed")

import shap

explainer = shap.Explainer(model, tokenizer)
shap_values = explainer(s)

It also throws an error below.

AttributeError: 'numpy.ndarray' object has no attribute 'new_zeros'

Below is the full error trace.

1516             return self._compiled_call_impl(*args, **kwargs)  # type: ignore[misc]
   1517         else:
-> 1518             return self._call_impl(*args, **kwargs)
   1519 
   1520     def _call_impl(self, *args, **kwargs):

[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py](https://localhost:8080/#) in _call_impl(self, *args, **kwargs)
   1525                 or _global_backward_pre_hooks or _global_backward_hooks
   1526                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1527             return forward_call(*args, **kwargs)
   1528 
   1529         try:

[/usr/local/lib/python3.10/dist-packages/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py](https://localhost:8080/#) in forward(self, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, head_mask, decoder_head_mask, cross_attn_head_mask, encoder_outputs, past_key_values, inputs_embeds, decoder_inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, return_dict)
   2510                 )
   2511 
-> 2512         outputs = self.model(
   2513             input_ids,
   2514             attention_mask=attention_mask,

[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py](https://localhost:8080/#) in _wrapped_call_impl(self, *args, **kwargs)
   1516             return self._compiled_call_impl(*args, **kwargs)  # type: ignore[misc]
   1517         else:
-> 1518             return self._call_impl(*args, **kwargs)
   1519 
   1520     def _call_impl(self, *args, **kwargs):

[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py](https://localhost:8080/#) in _call_impl(self, *args, **kwargs)
   1525                 or _global_backward_pre_hooks or _global_backward_hooks
   1526                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1527             return forward_call(*args, **kwargs)
   1528 
   1529         try:

[/usr/local/lib/python3.10/dist-packages/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py](https://localhost:8080/#) in forward(self, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, head_mask, decoder_head_mask, cross_attn_head_mask, encoder_outputs, past_key_values, inputs_embeds, decoder_inputs_embeds, use_cache, output_attentions, output_hidden_states, return_dict)
   2364                 )
   2365 
-> 2366             decoder_input_ids = shift_tokens_right(
   2367                 input_ids, self.config.pad_token_id, self.config.decoder_start_token_id
   2368             )

[/usr/local/lib/python3.10/dist-packages/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py](https://localhost:8080/#) in shift_tokens_right(input_ids, pad_token_id, decoder_start_token_id)
     67     Shift input ids one token to the right.
     68     """
---> 69     shifted_input_ids = input_ids.new_zeros(input_ids.shape)
     70     shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
     71     shifted_input_ids[:, 0] = decoder_start_token_id

AttributeError: 'numpy.ndarray' object has no attribute 'new_zeros'

I suppose there can be a version mismatch issue.
Can you please suggest what should be done.? Would be very helpful.

[CloseChoice]

Can you show me how to generate the output with the model you are using? That's what I tried above but that failed.

[SVC04]

Following is a piece of code to generate the output using longt5.

!pip install transformers[sentencepiece] datasets sacrebleu rouge_score py7zr -q
LONG_ARTICLE = """"anxiety affects quality of life in those living
with parkinson 's disease ( pd ) more so than
overall cognitive status , motor deficits , apathy
, and depression [ 13 ] . although anxiety and
depression are often related and coexist in pd
patients , recent research suggests that anxiety
rather than depression is the most prominent and
prevalent mood disorder in pd [ 5 , 6 ] . yet ,
our current understanding of anxiety and its
impact on cognition in pd , as well as its neural
basis and best treatment practices , remains
meager and lags far behind that of depression ."""

import torch
from transformers import AutoTokenizer, LongT5ForConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained("Stancld/longt5-tglobal-large-16384-pubmed-3k_steps")
input_ids = tokenizer(LONG_ARTICLE, return_tensors="pt").input_ids

model = LongT5ForConditionalGeneration.from_pretrained("Stancld/longt5-tglobal-large-16384-pubmed-3k_steps", return_dict_in_generate=True)

sequences = model.generate(input_ids).sequences

summary = tokenizer.batch_decode(sequences)

For the second model bigbird pegasus the output can be generated using the code below.

from transformers import BigBirdPegasusForConditionalGeneration, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("google/bigbird-pegasus-large-pubmed")
model = BigBirdPegasusForConditionalGeneration.from_pretrained("google/bigbird-pegasus-large-pubmed")

text = "text article"
inputs = tokenizer(text, return_tensors='pt')
prediction = model.generate(**inputs)
prediction = tokenizer.batch_decode(prediction)

[Yangliule]

Hey I encountered a similar issue. I think the issue is SHAP passed in a numpy array instead of tensor object when calling transformers API.

[SVC04]

@Yangliule Did you find any solution.?

[Yangliule]

Hey @SVC04, I solved the issue by reverting transformers version and pytorch version. Also, make sure you are using AutoModelForSeq2SeqLM.

[CloseChoice]

@Yangliule could you please show the versions that are working?

[Yangliule]

Hey @CloseChoice , I'm using transformers 4.12.0 and torch 1.10. Basically I reverted back to when the text SHAP feature was added.