- Vietnamese Handwriting Dataset: https://www.dropbox.com/s/zk82qkidkqma8bj/vn_handwriting_data.zip?dl=0
export PYTHONPATH=path-to-project
cp src/example_config.py src/config.py
-
Model trained with 25 epochs: https://github.com/huyhoang17/Colab_Temporary/blob/master/Training_CTC_Vietnamese_Recognition_25epochs.ipynb
-
[Overfitting model] Model trained with 10 epochs (initial weights from 25 epochs), reduced LR from 0.02 to 0.005: https://github.com/huyhoang17/Colab_Temporary/blob/master/%5BOverfitting%5DTraining_CTC_Vietnamese_Recognition_40epochs.ipynb
- Add Dockerfile
- Deploy simple demo with Tensorflow Serving
- Simple API
- Refactor code
- Add Visual Callback ✔️
- Train code with Colab's TPU?
- Convert to Pytorch code
- Data Augmentation ✔️
- Generate more data
- Simplifly the model
- Early Stopping
- Cross Validation
- Dropout (NN)
- Use Transfer Learing!
- ...
Colab Notebook:
Papers
- https://arxiv.org/pdf/1804.01527.pdf
- https://arxiv.org/abs/1801.08624
- https://arxiv.org/abs/1308.0850
CTC loss:
- https://www.dlology.com/blog/how-to-train-a-keras-model-to-recognize-variable-length-text/
- https://hackernoon.com/latest-deep-learning-ocr-with-keras-and-supervisely-in-15-minutes-34aecd630ed8
- https://github.com/SeanNaren/deepspeech.pytorch/blob/master/model.py
Kaggle ctc loss:
Explained ctc loss:
- https://gab41.lab41.org/speech-recognition-you-down-with-ctc-8d3b558943f0
- https://distill.pub/2017/ctc/
- https://stats.stackexchange.com/questions/320868/what-is-connectionist-temporal-classification-ctc
CTC loss param:
Overfitting