Model fails for a simple tex

[wklm]

Hi,

it's a great project, many thanks for it! The model needs some work though, I found it failing for relatively simple examples like this one: $\mathcal{X} \times \Theta \to \overbar{R}$ spitting out very different results in consecutive prediction attempts. Fundamentally this issue could be solved by the introduction of confidence thresholds...

Cheers,
Wojtek

[lukas-blecher]

I've never experienced this issue. I know sometimes the model is not able to translate some image. But from my experience it will predict almost always the same latex code.
Please make sure you have the correct dependency versions. Especially x_transformers==0.15.0.
If that's not it can you please upload a screen grab from the equation in question. I tried to reproduce it with the equation above (katex) but the prediction was consistently the same. Wrong but always the same.

[lukas-blecher]

I think this can only be solved by more training data.

[lukas-blecher]

As it turns out it was another problem which is fixed now in #119

returns now X\times\Theta\rightarrow\bar{R}