formula-recognition-medium-scan-0001 (composite)

Use Case and High-Level Description

This is an im2latex composite model that recognizes latex formulas. The model uses vocabulary file to predict sequence of latex tokens. The model is built on the resnext50 backbone with additional attention-based text recognition head.

Vocabulary file is distributed with the model. Model can predict big and small letters, numbers, some greek letters, trigonometric functions (e.g. cos, sin, coth), logarithmic function, sqrt and superscript.

Example of the input data

Example of the output

4 7 4 W ^ { 1 } + 7 . 1 9 o ^ { 4 } - 6 - 0 . 9 6 L ^ { 1 } y

Composite model specification

Metric Value
im2latex_medium_photographed dataset, im2latex-match-images metric 81.5%
im2latex_medium_rendered dataset, im2latex-match-images metric 95.7%
Source framework PyTorch*

Im2latex-match-images metric is calculated by this algorithm

Encoder model specification

The formula-recognition-medium-scan-0001-encoder model is a ResNeXt-50 like backbone with some initialization layers for decoder

Metric Value
GFlops 16.56
MParams 1.69

Performance

Inputs

  1. Name: imgs , shape: [1x3x160x1400]. An input image in the [1xCxHxW] format. The expected channel order is BGR.

Outputs

  1. Name: hidden, shape: [1x512]. Initial context state of the LSTM cell.
  2. Name: context, shape: [1x512]. Initial hidden state of the LSTM cell.
  3. Name: init_0, shape: [1x256]. Initial state of the decoder.
  4. Name: row_enc_out, shape: [1x20x75x512]. Features from encoder that are fed to a decoder.

Decoder model specification

The formula-recognition-medium-scan-0001-decoder model is an LSTM based decoder with attention module.

Metric Value
GFlops 1.86
MParams 2.56

Performance

Inputs

  1. Name: dec_st_c , shape: [1x512]. Current context state of the LSTM cell.
  2. Name: dec_st_h , shape: [1x512]. Current hidden state of the LSTM cell.
  3. Name: output_prev, shape: [1x256]. Current state of the decoder.
  4. Name: row_enc_out, shape: [1x20x175x512]. Encoded features.
  5. Name: tgt, shape: [1x1]. Index of the previous symbol.

Outputs

  1. Name: dec_st_c , shape: [1x512]. Current context state of the LSTM cell.
  2. Name: dec_st_h , shape: [1x512]. Current hidden state of the LSTM cell.
  3. Name: output, shape: [1x256]. Current state of the decoder.
  1. Name: logit, shape: [1xVocab_Size]. Classification confidence scores in the [0, 1] range for every token.

Legal Information

[*] Other names and brands may be claimed as the property of others.