OCR-09 googoo

Requirements

• Python 3
• PyTorch

All dependencies can be installed with PIP.

pip install tensorboardX tqdm pyyaml psutil

pip install -r requirements.txt

현재 검증된 GPU 개발환경으로는

• Pytorch 1.0.0 (CUDA 10.1)
• Pytorch 1.4.0 (CUDA 10.0)
• Pytorch 1.7.1 (CUDA 11.0)

Supported Models

• CRNN
• SATRN

Supported Data

• Aida (synthetic handwritten)
• CROHME (online handwritten)
• IM2LATEX (pdf, synthetic handwritten)
• Upstage (print, handwritten)

Usage

Training

python train.py

Evaluation

python evaluate.py

Config

configs/SATRN.yaml에서 config 수정 가능.

network: model to use.
input_size:
  height: height of input image
  width: width of input image
SATRN:
  encoder:
    hidden_dim: size of hidden dimension of encoder
    filter_dim: size of intermediate dimension of feedforward network.
    layer_num: number of encoder layer.
    head_num: number of heads in multi-head attn.
  decoder:
    src_dim: size of input vector.
    hidden_dim: size of hidden dimension of decoder.
    filter_dim: size of intermediate dimension of feedforward network.
    layer_num: number of decoder layer.
    head_num: number of heads in multi-head attn.
Attention:
  src_dim: size of input vector.
  hidden_dim: size of hidden dimension of lstm.
  embedding_dim: size of embedding dimension of input.
  layer_num: number of latm layer
  cell_type: LSTM or GRU
checkpoint: file path of checkpoint
prefix: "./log/satrn"
version: ''

data:
  train:
    - "/opt/ml/input/data/train_dataset/gt.txt"
  test:
    - ""
  token_paths:
    - "/opt/ml/input/data/train_dataset/tokens.txt"  # 241 tokens
  dataset_proportions:  # proportion of data to take from train (not test)
    - 1.0
  random_split: if True, random split from train files
  test_proportions: 0.2 # only if random_split is True
  crop: True
  rgb: 1    # 3 for color, 1 for greyscale
  
batch_size: 34
num_workers: 8
num_epochs: 60
print_epochs: 1
dropout_rate: 0.1
teacher_forcing_ratio: 0.5
max_grad_norm: 2.0
seed: 1234
optimizer:
  optimizer: 'Adam' # Adam, Adadelta
  lr: 5e-4 # 1e-4
  weight_decay: 1e-4
  is_cycle: True

experiment report

T1219_최현진 : https://github.com/bcaitech1/p4-fr-9-googoo/wiki/%5BT1219_%EC%B5%9C%ED%98%84%EC%A7%84%5DSATRN-%EC%8B%A4%ED%97%98

T1243_김선욱 : https://github.com/23ksw10/Math-Expression-Recognition/blob/main/README.md