one of the variables needed for gradient computation has been modified by an inplace operation

[TTzhangheng]

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.LongTensor [4, 90]] is at version 1; expected version 0 instead. Hint: why？

[Junyiliu0]

Same question.Did you solve it?

[Junyiliu0]

I have solved it. The main problem is in the loss_icr. Just change the bill.py in line 163 to
text_encoder_input_ids = text.input_ids.clone()
text_output = self.text_encoder(text_encoder_input_ids, attention_mask=text.attention_mask,
return_dict=True, mode='text')
Maybe the problem is the input_ids of the text, because the tensor shape is [8,90](my batch_size = 8).

[Paarth002]

@mlii0117, I am still getting this error 'RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation'. Please suggest how to solve this error.

[confesion]

I have solved it. The main problem is in the loss_icr. Just change the bill.py in line 163 to text_encoder_input_ids = text.input_ids.clone() text_output = self.text_encoder(text_encoder_input_ids, attention_mask=text.attention_mask, return_dict=True, mode='text') Maybe the problem is the input_ids of the text, because the tensor shape is [8,90](my batch_size = 8).

Yeah,it is blip.py instead of bill.py.

[kongkong935]

@confesion Hello, I have also encountered this problem. According to your method, it did not solve it. Can you explain it in more detail

[kongkong935]

@Paarth002 I have also encountered this problem and would like to ask if you have solved it? Can you help me

[kongkong935]

@TTzhangheng I have also encountered this problem and would like to ask if you have solved it? Can you help me