Hello everyone, in this thread I hope to detail my approach to working on the Cassava Leaf Disease Classification Competition on Kaggle.
About The Competition
This is a Image classification task wherein the objective is to classify an image and associate it with a label. The task is to enable farmers to quick identify if the the image is affected by one of the diseases for which there are examples there are in the dataset, and leaves which are healthy.
About the Dataset
This dataset is provided in the form of
TFRecords and Images. There are a total of 21,367 Training Images and 5 Classes. These images were collected on mobile cameras, and would be interesting to see the differences in image statistics due to this.
The classes are as follows:
0: Cassava Bacterial Blight (CBB)
1: Cassava Brown Streak Disease (CBS**D)
2: Cassava Green Mottle (CGM)
3: Cassava Mosaic Disease (CMD)
I intend to work with Remo to visualize, create train-valid splits and also perform Exploratory Data Analysis easily, and outline how it could be useful to perform this.
The initial plan for the model exploration (Subject to change, through iteration)
- Creating a supervised transfer learning baseline
- Creating a self-supervised transfer learning baseline
- Self-supervised learning on the training set + subsequent finetuning.
I will be updating this thread as I progress, and adding insights and work towards a good position in the competition.
|Model||Type of Training||Training Accuracy||Validation Accuracy|