Advanced Training

Assignment:

1. Pick your last code
2. Make sure to Add CutOut to your code. It should come from your transformations (albumentations)
3. Use this repo: https://github.com/davidtvs/pytorch-lr-finder (Links to an external site.)
   1. Move LR Finder code to your modules
   2. Implement LR Finder (for SGD, not for ADAM)
   3. Implement ReduceLROnPlatea: https://pytorch.org/docs/stable/optim.html#torch.optim.lr_scheduler.ReduceLROnPlateau (Links to an external site.)
4. Find best LR to train your model
5. Use SDG with Momentum
6. Train for 50 Epochs.
7. Show Training and Test Accuracy curves
8. Target 88% Accuracy.
9. Run GradCAM on the any 25 misclassified images. Make sure you mention what is the prediction and what was the ground truth label.
10. Submit

Solution

Github-Notebook : https://github.com/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/10_AdvancedTraining/AdvancedTraining.ipynb

Google-Colab: https://colab.research.google.com/github/satyajitghana/TSAI-DeepVision-EVA4.0/blob/master/10_AdvancedTraining/AdvancedTraining.ipynb

Model Stats

Epochs: 50
Max Train Accuracy: 98.27
Max Test Accuracy: 92.03

Model Metrics

Grad-CAM

Misclassifications Grad-CAM