quarticcat / detypify Goto Github PK

We need an extra step to download detypify assets. What's worse is that we should keep in mind and update the assets on detypify upgrade its pre-trained model.

Myriad-Dreamin/tinymist#260

Possible solution: publish and install it via npm: yarn add detypify

import {Detypify} from 'detypify';
console.log((await Detypify.create()).candidates(strokes));

https://github.com/Myriad-Dreamin/tinymist/blob/01f72c3d75b099afff7ff59c394067e57b79027a/tools/editor-tools/src/features/symbol-view.detypify.ts

The original training data come from Detexify is in the form of strokes that can be described as follows:

type Point = tuple[float, float, int]  # (x, y, timestamp)
type Stroke = list[Point]
type Strokes = list[Stroke]

The implementation of Detexify is KNN where distances are calculated by Greedy DTW. This model is too large to ship with the website. Besides, I don't know how to handle variable-length inputs. Therefore, I found another way.

Current model of Detypify (0.1.0) is naive CNN. It first normalizes strokes (center + scale), then draw them on a 32x32 canvas (pixelated, no anti-aliasing), then CNN. This approach has some drawbacks:

• It cannot distinguish prime.rev and backslash very well. Normalization makes them look the same. There might be more such examples.
• It cannot recognize uncommon symbols very well. This is more related to the problem of training data, which is extremely unbalanced: the numbers of samples vary from 10- to 3900+. But it seems like Detexify can recognize them.

Could a stroke-based model be better? If so, which model should I use?

Or, is there any potential improvement to the current model?

Currently, all training data come from Detexify. The dataset is extremely unbalanced. Some symbols have more than 3900 samples, while some have less than 10 samples. Many uncommon symbols are nearly untrained.

Apart from that, due to the difference between LaTeX and Typst, some Typst symbols don't even have corresponding samples in Detexify's training data. For example, the join series (there's only \bowtie).

That is to say, we really need contributions from the community. Here comes the question: I don't know which platform is suitable for this need.

In addition to that, we can probably explore some data augmentation methods to enhance the training set. And unfortunately, I know nothing about that as well.

How to draw a $\wp$ (\wp, ℘) ? At present, it seems that all inputs of the trajectory will be recognized as φ and other results. Is this because the dataset does not contain ℘?