'tuple' object has no attribute 'dim' in generation

[pengshancai]

Hi there, During training, the code report the following bug

File "doha.py", line 756, in main() File "doha.py", line 746, in main qa.train() File "doha.py", line 528, in train results = self.evaluate(dev_data) File "doha.py", line 613, in evaluate pred, total_eval_loss, total_words = self.predict((dev_examples, dev_features)) File "doha.py", line 575, in predict top_p=0.9, File "/root/anaconda3/envs/wow/lib/python3.7/site-packages/torch/autograd/grad_mode.py", line 28, in decorate_context return func(*args, *kwargs) File "/root/anaconda3/envs/wow/lib/python3.7/site-packages/transformers/generation_utils.py", line 396, in generate encoder_outputs: ModelOutput = encoder(input_ids, attention_mask=attention_mask, return_dict=True) File "/root/anaconda3/envs/wow/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(input, **kwargs) File "/root/anaconda3/envs/wow/lib/python3.7/site-packages/transformers/modeling_bart.py", line 331, in forward attention_mask = invert_mask(attention_mask) File "/root/anaconda3/envs/wow/lib/python3.7/site-packages/transformers/modeling_bart.py", line 135, in invert_mask assert attention_mask.dim() == 2 AttributeError: 'tuple' object has no attribute 'dim'

I am using transformer 3.1.0, it seems to me that during encoding, the attention mask must be a tensor instead of a tuple containing 2 tensors

Any suggestion on this?

[shrimai]

Have you used the patch for generation_utils provided in the code? You have to copy patch/generation_utils.py to your path /root/anaconda3/envs/wow/lib/python3.7/site-packages/transformers/

[pengshancai]

Oh, I did not notice that. Thanks!