EVA4 Session 5 assginment
- Set up and defining skeleton with Convolution block, GAP and Convolution block.
- Parameters: 13,584
- Best Train Accuracy 98.53%
- Best Test Accuracy 98.34%
- Basic skeleton model doesn't meet target accuracy of 99.4% and model is slightly over-fitting.
- Parameters count is >10k
- Adding batch-norm and drop-out to reduce over-fitting and improve the model efficiency.
- Iterate to find the best drop-out value (0.05 to 0.2)
- Parameters: 13,584
- Best results with 0.05 as Dropout value where difference b/w test and train accuracy is small.
- Best Train Accuracy 99.12%
- Best Test Accuracy 99.45%
- Model is under-fitting and total parameters also > 10k.
- Eventhough the model is under-fitting and we need to add more parameters. We need to optimizing of output channels to decrease the parameters < 10 K to meet the requirement.
- Parameters: 9,752
- Best Train Accuracy 99.24%
- Best Test Accuracy 99.39%
- Total parameters < 10k
- Model performance is good but not achieving 99.4% accuracy target.
- Add image augmentation w random rotation and random affine to improve the model performance.
- Parameters: 9,752
- Best Train Accuracy 98.03%
- Best Test Accuracy 99.26%
- Total parameters < 10k
- Data augmentation didn't improve the accuracy.
- Adding LR Scheduler
- Iterated on learning rate (0.01, 0.02)
- Iterated on removing batch-norm and drop-out in initial layer of convolution.
- Parameters: 9,712
- Best Train Accuracy 99.22%
- Best Test Accuracy 99.48%
- Total parameters < 10k
- Accuracy > 99.40 consistently for 4 instances in 15 epochs.
- Model performance is good as difference between train and test is small.