Cassava Leaf Disease Classification - Pytorch project

Leaf disease image classification model deployed with Streamlit. Project contains also implemented from scratch MLP-Mixer [1] architecture.

Content

• main.py - Streamlit server code
• eval.py - supplementary functions for model inference in Streamlit
• models.py - model definitions and Pytorch Lightning model wrapper
• datasets.py - contains CassavaDataset class
• train.py - training script with early stopping and tensorboard logging.

Data

Dataset comes from Cassava Leaf Disease Kaggle competition[2] as .png files labeled with .csv file.

Models

MlpMixer - implemented neural network as described in paper[1].

TransferredInception - class for transfering Inception_v3 architecture.

LightningModelWrapper - Pytorch Lightning wrapper for models above (or any other for this task). Comes with logging per epoch (for more readable learnig process plots in Tensorboard)

Deployment

Application is served on Microsoft Azure VM with Streamlit on address: http://51.13.72.28

Used tools

• Pytorch
• Pytorch Lightning
• Albumentations
• Streamlit

References

[1]I. Tolstikhin et al., ‘MLP-Mixer: An all-MLP Architecture for Vision’, arXiv:2105.01601 [cs], Jun. 2021, Accessed: Mar. 14, 2022. [Online]. Available: http://arxiv.org/abs/2105.01601

[2]E. Mwebaze, T. Gebru, A. Frome, S. Nsumba, and J. Tusubira, ‘iCassava 2019 Fine-Grained Visual Categorization Challenge’, arXiv:1908.02900 [cs], Dec. 2019, Accessed: Mar. 14, 2022. [Online]. Available: http://arxiv.org/abs/1908.02900