Learnt how to write the dataloader, the model and the training code . The following analysis was done:
- With Batch Norm
- Adding new layers
- With Dropout
- Different activation functions at the end
- Different pooling strategies
- Different optimizers
- Basic Augmentation like Rotation, Translation, Color Change
Also experimented with: Experiment with
- Residual Blocks
- Different learning strategies
It is available here.There are approx 9233 images in train and 484 images in test set.
The submission is for this challenge on ai-crowd portal. This was done using fastai and resnet-50 pre-trained model . Using this code I achieved an accuracy of 0.616