Unexpected key(s) in state_dict and size mismatch when running python3 pix2tex.py about latex-ocr

[lukas-blecher]

Hi,
this seems to be an issue with the timm package. Just now I've closed PR #31, which allows to use the newest version.

Please update the package

pip install -U timm

and make sure you are using the new version of this repository.

If the error persists please let me know.

[ajq2020]

thanks, the program is running now, but the output result is wrong.
I tested with the image

and the output is

c^{2}\nabla\times B={\frac{\frac{\underline{3}}}{2}{5}}{5}{5}{5}{5}{6}{\delta+{\delta u}}}\sin^{\left(-16}c^{\right.\frac{\lambda}}c^{\left(\right.\right.\right.\left(\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.\right.

https://katex.org/?data=%7B%22displayMode%22%3Atrue%2C%22leqno%22%3Afalse%2C%22fleqn%22%3Afalse%2C%22throwOnError%22%3Atrue%2C%22errorColor%22%3A%22%23cc0000%22%2C%22strict%22%3A%22warn%22%2C%22output%22%3A%22htmlAndMathml%22%2C%22trust%22%3Afalse%2C%22code%22%3A%22c%5E%7B2%7D%5C%5Cnabla%5C%5Ctimes%20B%3D%7B%5C%5Cfrac%7B%5C%5Cfrac%7B%5C%5Cunderline%7B3%7D%7D%7D%7B2%7D%7B5%7D%7D%7B5%7D%7B5%7D%7B5%7D%7B5%7D%7B6%7D%7B%5C%5Cdelta%2B%7B%5C%5Cdelta%20u%7D%7D%7D%5C%5Csin%5E%7B%5C%5Cleft%28-16%7Dc%5E%7B%5C%5Cright.%5C%5Cfrac%7B%5C%5Clambda%7D%7Dc%5E%7B%5C%5Cleft%28%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cleft%28%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%5C%5Cright.%22%7D

katex cannot parse, so I think the result is not right.

[lukas-blecher]

Ok, so the main issue is resolved.

Regarding the prediction: When I feed the exact same picture to the model I get

c^{2}\nabla\times B={\frac{j}{\varepsilon_{0}}}+{\frac{\partial E}{\partial t}}c^{2}

link to katex. The result is not perfect, the vectors are not bold but other than that it is correct.

Note, the output is not deterministic. You can try to predict the image again or to change a couple of parameters, e.g. the temperature.

In case that you did not download the image_resizer.pth, I'd suggest you try it again. It can improve performance greatly, especially for quite large images.

[lukas-blecher]

Might be related to #33
Please downgrade x-tranformers to 0.15.0

pip install -U x-transformers==0.15.0

Thanks