EVA4 Assignment S7, TSAI
- Setup the code and define the baseline architecture.
- Set the device to GPU instead of CPU.
- Exploratory data to understand the data normalization and data augmentation.
- Total parameters = 2,435,744
- For 2 epochs, Train Accuracy 63%
- For 2 epochs, Test Accuracy 62%
- Not able to achieve target accuracy of 80% and parameters less than 1M.
- Set the image normalization and image augmentation
- Total parameters = 2,435,744
- For 2 epochs, Train Accuracy 60%
- For 2 epochs, Test Accuracy 62%
- Accuracy didn't improve as we tried only for two epochs.
- Refining the architecture to meet parameter target < 1M
- Total parameters = 1,092,768
- For 20 epochs, Train Accuracy 84%
- For 20 epochs, Test Accuracy 81%
- Accuracy meets the target >80% but slightly over the parameter targets 1M.
Implemented depth wise convolution and dilated layer. Even optimized parameter to meet target of 1M
- Total parameters = 943,552
- For 20 epochs, Train Accuracy 82%
- For 20 epochs, Test Accuracy 81%
- Results of accuracy is more than 80% and parameters target < 1M.
- Models is performing well also as gap between train and test accuracy is small.