Hey, I’m trying to use TorchVision’s get_graph_node_names() to fetch the names of nodes available for feature extraction in the Huggingface’s gpt2-small model using the following code:
from torchvision.models.feature_extraction import get_graph_node_names
nodes, _ = get_graph_node_names(gpt_small)
print(nodes)
However, I’m getting the following error:
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Input In [14], in <cell line: 2>()
1 from torchvision.models.feature_extraction import get_graph_node_names
----> 2 nodes, _ = get_graph_node_names(gpt_small)
3 print(nodes)
File ~/miniconda3/envs/biofeedback/lib/python3.10/site-packages/torchvision/models/feature_extraction.py:253, in get_graph_node_names(model, tracer_kwargs, suppress_diff_warning)
251 is_training = model.training
252 train_tracer = NodePathTracer(**tracer_kwargs)
--> 253 train_tracer.trace(model.train())
254 eval_tracer = NodePathTracer(**tracer_kwargs)
255 eval_tracer.trace(model.eval())
File ~/miniconda3/envs/biofeedback/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:587, in Tracer.trace(self, root, concrete_args)
585 for module in self._autowrap_search:
586 _autowrap_check(patcher, module.__dict__, self._autowrap_function_ids)
--> 587 self.create_node('output', 'output', (self.create_arg(fn(*args)),), {},
588 type_expr=fn.__annotations__.get('return', None))
590 self.submodule_paths = None
592 return self.graph
File ~/miniconda3/envs/biofeedback/lib/python3.10/site-packages/transformers/models/gpt2/modeling_gpt2.py:1047, in GPT2LMHeadModel.forward(self, input_ids, past_key_values, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, encoder_hidden_states, encoder_attention_mask, labels, use_cache, output_attentions, output_hidden_states, return_dict)
1039 r"""
1040 labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1041 Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
1042 `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
1043 are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
1044 """
1045 return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-> 1047 transformer_outputs = self.transformer(
1048 input_ids,
1049 past_key_values=past_key_values,
1050 attention_mask=attention_mask,
1051 token_type_ids=token_type_ids,
1052 position_ids=position_ids,
1053 head_mask=head_mask,
1054 inputs_embeds=inputs_embeds,
1055 encoder_hidden_states=encoder_hidden_states,
1056 encoder_attention_mask=encoder_attention_mask,
1057 use_cache=use_cache,
1058 output_attentions=output_attentions,
1059 output_hidden_states=output_hidden_states,
1060 return_dict=return_dict,
1061 )
1062 hidden_states = transformer_outputs[0]
1064 # Set device for model parallelism
File ~/miniconda3/envs/biofeedback/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:577, in Tracer.trace.<locals>.module_call_wrapper(mod, *args, **kwargs)
573 return _orig_module_call(mod, *args, **kwargs)
575 _autowrap_check(patcher, getattr(getattr(mod, "forward", mod), "__globals__", {}),
576 self._autowrap_function_ids)
--> 577 return self.call_module(mod, forward, args, kwargs)
File ~/miniconda3/envs/biofeedback/lib/python3.10/site-packages/torchvision/models/feature_extraction.py:84, in NodePathTracer.call_module(self, m, forward, args, kwargs)
82 self.current_module_qualname = module_qualname
83 if not self.is_leaf_module(m, module_qualname):
---> 84 out = forward(*args, **kwargs)
85 return out
86 return self.create_proxy("call_module", module_qualname, args, kwargs)
File ~/miniconda3/envs/biofeedback/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:573, in Tracer.trace.<locals>.module_call_wrapper.<locals>.forward(*args, **kwargs)
572 def forward(*args, **kwargs):
--> 573 return _orig_module_call(mod, *args, **kwargs)
File ~/miniconda3/envs/biofeedback/lib/python3.10/site-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 ~/miniconda3/envs/biofeedback/lib/python3.10/site-packages/transformers/models/gpt2/modeling_gpt2.py:768, in GPT2Model.forward(self, input_ids, past_key_values, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, encoder_hidden_states, encoder_attention_mask, use_cache, output_attentions, output_hidden_states, return_dict)
765 return_dict = return_dict if return_dict is not None else self.config.use_return_dict
767 if input_ids is not None and inputs_embeds is not None:
--> 768 raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
769 elif input_ids is not None:
770 input_shape = input_ids.size()
ValueError: You cannot specify both input_ids and inputs_embeds at the same time
I don’t understand why both input_ids and inputs_embeds are getting passed to the gpt2-small during the symbolic tracing. Any help would be greatly appreciated.
Pytorch version: 1.12.0
Thanks!