sujitojha1 / EVA4-S5

EVA4 Session 5 assginment

S5

EVA4 Session 5 assginment

Step 1

Target .

1. Set up and defining skeleton with Convolution block, GAP and Convolution block.

Results .

1. Parameters: 13,584
2. Best Train Accuracy 98.53%
3. Best Test Accuracy 98.34%

Analysis

1. Basic skeleton model doesn't meet target accuracy of 99.4% and model is slightly over-fitting.
2. Parameters count is >10k

Step 2

Target .

1. Adding batch-norm and drop-out to reduce over-fitting and improve the model efficiency.
2. Iterate to find the best drop-out value (0.05 to 0.2)

Results .

1. Parameters: 13,584
2. Best results with 0.05 as Dropout value where difference b/w test and train accuracy is small.
3. Best Train Accuracy 99.12%
4. Best Test Accuracy 99.45%

Analysis

1. Model is under-fitting and total parameters also > 10k.

Step 3

Target .

1. 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.

Results .

1. Parameters: 9,752
2. Best Train Accuracy 99.24%
3. Best Test Accuracy 99.39%

Analysis

1. Total parameters < 10k
2. Model performance is good but not achieving 99.4% accuracy target.

Step 4

Target .

1. Add image augmentation w random rotation and random affine to improve the model performance.

Results .

1. Parameters: 9,752
2. Best Train Accuracy 98.03%
3. Best Test Accuracy 99.26%

Analysis

1. Total parameters < 10k
2. Data augmentation didn't improve the accuracy.

Step 5

Target .

1. Adding LR Scheduler
2. Iterated on learning rate (0.01, 0.02)
3. Iterated on removing batch-norm and drop-out in initial layer of convolution.

Results .

1. Parameters: 9,712
2. Best Train Accuracy 99.22%
3. Best Test Accuracy 99.48%

Analysis

1. Total parameters < 10k
2. Accuracy > 99.40 consistently for 4 instances in 15 epochs.
3. Model performance is good as difference between train and test is small.