The Dataset provided has an excel sheet with columns as image_id, split_type(train or test), Target class(class 1 to class 29) which is the binarized conversion of the target Very similar to One-hot encoding
The Training dataset has 64674 images and test dataset has 7186 images. All the preprocessing is done on the training dataset and validation is done on the test dataset. The Training dataset is highly imbalanced and images have different sizes (1024x1024x3, 475x475x3) etc. Resizing is performed on the images to a standard size of 400 x 400 x 3
Primary metrics used for the Project is F1 micro and F1 macro. The Best-fit model has a F1 macro score of 0.459 and F1 micro of 0.39 on eval set
I trained 2 models - Pretrained resNet model and a custom CNN model. The Best fit is a Convolutional Neural Network with 7 convolutional layers. Every Conv2d layer is developed with a BatchNorm and a pooling layer. Some hyperparameters like activation function, batch sizes and learning rates were experimented with and optimized during the course of the project.