Salesforce

/

blip2-flan-t5-xl

like 59

Follow

Salesforce 843

Hi.

All of a sudden, this error occurred. It was fine when I used this model about 5 days ago.
The error was shape mismatch: value tensor of shape [393216] cannot be broadcast to indexing result of shape [0].

I was trying to figure out why this happened, although my code was not changed...

But I finally figure out this is because of the Update for transformers (#5)

Here is the error.

What should I do?

RuntimeError
shape mismatch: value tensor of shape [393216] cannot be broadcast to indexing result of shape [0]
  File "/ssd01/jisu_sohee/PAVLM/jacob/transformers/models/blip_2/modeling_blip_2.py", line 2324, in generate
    inputs_embeds[special_image_mask] = language_model_inputs.flatten()
    ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^
  File "/ssd01/jisu_sohee/PAVLM/jacob/ad_setting10.py", line 515, in forward
    captionA_outputs = self.blip2_model.generate(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/ssd01/jisu_sohee/PAVLM/jacob/trainer_setting10.py", line 171, in validate_per_epoch
    loss,acc = self.model(
               ^^^^^^^^^^^
  File "/ssd01/jisu_sohee/PAVLM/jacob/trainer_setting10.py", line 199, in train
    _ = self.validate_per_epoch(-1, initial_infernece=True)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/ssd01/jisu_sohee/PAVLM/jacob/trainer_setting10.py", line 239, in <module>
    trainer.train(ckpt_dir="checkpoints/mvtec_setting10")
RuntimeError: shape mismatch: value tensor of shape [393216] cannot be broadcast to indexing result of shape [0]