Limitations on input image size

[Alf-Z-SymphoMe]

Hi, I tried loading a few scores (png and pdf), but I got the error message below.
I managed to make the model work with a much smaller score, as I read in an other issue that exists a maximum amount of allowed pixels per each dimension. If there is no solution yet, I'll leave it as an issue.

predictions, _ = model.predict(convert_img_to_tensor(image).unsqueeze(0).to(device),
File "site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "smt_model/modeling_smt.py", line 406, in predict
output = self.forward_decoder(encoder_output=encoder_output, last_predictions=predicted_sequence,
File "smt_model/modeling_smt.py", line 372, in forward_decoder
encoder_output_2D = self.pos2D(encoder_output)
File "site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
File "smt_model/modeling_smt.py", line 41, in forward
return x + self.get_pe_by_size(x.size(-2), x.size(-1))
RuntimeError: The size of tensor a (263) must match the size of tensor b (16) at non-singleton dimension 2

[antoniorv6]

Hi! Which SMT model you tried to use? Also, what is the resolution of your scores?

[Alf-Z-SymphoMe]

Hi, I initially tried antoniorv6/smt-grandstaff, but I ended up using PRAIG/smt-grandstaff, as suggested in the issues, due to problems in loading some of the wrights. The png file used for the error above has dimensions 2977 × 4208, not sure if you are looking for the dpi.

[antoniorv6]

Hi,
Are you introducing a system-level image or a full-page one? I guess that you are doing the first, so the PRAIG/smt-grandstaff weights should be fine. Of course, the image is too large and currently there is no control on that input size when performing tests. Although the model is able to stand some variance on height, I suggest that you reduce your image up to 256 px height maintaining the aspect ratio, which is the average size of the GrandStaff dataset.