Reputation: 8026
I want a summary of a PyTorch
model downloaded from huggingface.
Am I doing something wrong here?
from torchinfo import summary
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
summary(model, input_size=(16, 512))
Gives the error:
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
/usr/local/lib/python3.7/dist-packages/torchinfo/torchinfo.py in forward_pass(model, x, batch_dim, cache_forward_pass, device, **kwargs)
257 if isinstance(x, (list, tuple)):
--> 258 _ = model.to(device)(*x, **kwargs)
259 elif isinstance(x, dict):
11 frames
/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1050 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051 return forward_call(*input, **kwargs)
1052 # Do not call functions when jit is used
/usr/local/lib/python3.7/dist-packages/transformers/models/bert/modeling_bert.py in forward(self, input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, labels, output_attentions, output_hidden_states, return_dict)
1530 output_hidden_states=output_hidden_states,
-> 1531 return_dict=return_dict,
1532 )
/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1070
-> 1071 result = forward_call(*input, **kwargs)
1072 if _global_forward_hooks or self._forward_hooks:
/usr/local/lib/python3.7/dist-packages/transformers/models/bert/modeling_bert.py in 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)
988 inputs_embeds=inputs_embeds,
--> 989 past_key_values_length=past_key_values_length,
990 )
/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1070
-> 1071 result = forward_call(*input, **kwargs)
1072 if _global_forward_hooks or self._forward_hooks:
/usr/local/lib/python3.7/dist-packages/transformers/models/bert/modeling_bert.py in forward(self, input_ids, token_type_ids, position_ids, inputs_embeds, past_key_values_length)
214 if inputs_embeds is None:
--> 215 inputs_embeds = self.word_embeddings(input_ids)
216 token_type_embeddings = self.token_type_embeddings(token_type_ids)
/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1070
-> 1071 result = forward_call(*input, **kwargs)
1072 if _global_forward_hooks or self._forward_hooks:
/usr/local/lib/python3.7/dist-packages/torch/nn/modules/sparse.py in forward(self, input)
159 input, self.weight, self.padding_idx, self.max_norm,
--> 160 self.norm_type, self.scale_grad_by_freq, self.sparse)
161
/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py in embedding(input, weight, padding_idx, max_norm, norm_type, scale_grad_by_freq, sparse)
2042 _no_grad_embedding_renorm_(weight, input, max_norm, norm_type)
-> 2043 return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
2044
RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got torch.cuda.FloatTensor instead (while checking arguments for embedding)
The above exception was the direct cause of the following exception:
RuntimeError Traceback (most recent call last)
<ipython-input-8-4f70d4e6fa82> in <module>()
5 else:
6 # Can't get this working
----> 7 summary(model, input_size=(16, 512)) #, device='cpu')
8 #print(model)
/usr/local/lib/python3.7/dist-packages/torchinfo/torchinfo.py in summary(model, input_size, input_data, batch_dim, cache_forward_pass, col_names, col_width, depth, device, dtypes, row_settings, verbose, **kwargs)
190 )
191 summary_list = forward_pass(
--> 192 model, x, batch_dim, cache_forward_pass, device, **kwargs
193 )
194 formatting = FormattingOptions(depth, verbose, col_names, col_width, row_settings)
/usr/local/lib/python3.7/dist-packages/torchinfo/torchinfo.py in forward_pass(model, x, batch_dim, cache_forward_pass, device, **kwargs)
268 "Failed to run torchinfo. See above stack traces for more details. "
269 f"Executed layers up to: {executed_layers}"
--> 270 ) from e
271 finally:
272 if hooks is not None:
RuntimeError: Failed to run torchinfo. See above stack traces for more details. Executed layers up to: []
Upvotes: 7
Views: 8770
Reputation: 1036
I encountered a similar issue, where I obtained a layer-by-layer summary of the model the following way:
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained(
pretrained_model_name_or_path='bert-base-uncased',
num_labels=2
)
for layer_name, params in model.named_parameters():
print(layer_name, params.shape)
# bert.embeddings.word_embeddings.weight torch.Size([30522, 768])
# bert.embeddings.position_embeddings.weight torch.Size([512, 768])
# bert.embeddings.token_type_embeddings.weight torch.Size([2, 768])
# bert.embeddings.LayerNorm.weight torch.Size([768])
# bert.embeddings.LayerNorm.bias torch.Size([768])
# bert.encoder.layer.0.attention.self.query.weight torch.Size([768, 768])
# ...
Upvotes: 1
Reputation: 1376
There's a bug [also reported] in torchinfo
library [torchinfo.py
] in the last line shown. When dtypes
is None
, it is by default creating torch.float
tensors whereas forward
method of bert
model uses torch.nn.embedding
which expects only int/long
tensors.
def process_input(
input_data: Optional[INPUT_DATA_TYPE],
input_size: Optional[INPUT_SIZE_TYPE],
batch_dim: Optional[int],
device: Union[torch.device, str],
dtypes: Optional[List[torch.dtype]] = None,
) -> Tuple[CORRECTED_INPUT_DATA_TYPE, Any]:
"""Reads sample input data to get the input size."""
if input_size is not None:
if dtypes is None:
dtypes = [torch.float] * len(input_size)
If you try modifying the line to the following, it works fine.
dtypes = [torch.int] * len(input_size)
EDIT (Direct solution w/o changing their internal code):
from torchinfo import summary
from transformers import AutoModelForSequenceClassification, AutoTokenizer
model = AutoModelForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
summary(model, input_size=(2, 512), dtypes=['torch.IntTensor'])
Alternate:
For a simple summary, you could use print(model)
instead of summary
function.
Upvotes: 8