/ERA_S7_Assignment

Primary LanguageJupyter Notebook

Session 7 Assignment

CODE 1: APPLY TECHNIQUE BATCH NORMALIZATION, REGULARIZATION, and GLOBAL AVERAGE POOLING to a Basic Skeleton

Target:

  • Use nn.Sequential
  • Add BatchNorm
  • Apply dropOut on each layer
  • Use GAP in last layer

Results:

  • Parameters: 5.1k
  • Best Training Accuracy: 93.31
  • Best Test Accuracy: 97.78

Analysis:

  • Model is way to much lighter
  • It seems model is under-fitting cause of very less number of parameters

With BatchNorm

  • Parameters: 10 k
  • Best training accuracy = 99.88
  • Best Test accuracy = 99.26%

With BatchNorm and DropOut

  • Parameters: 10 k
  • Best training accuracy = 99.04
  • Best Test accuracy = 99.13%

With BatchNorm, DropOut and GAP

  • Parameters: 5.1 k
  • Best training accuracy = 93.91
  • Best Test accuracy = 97.78%

CODE 2: INCREASE MODEL CAPACITY AND FINE TUNE MAXPOOLING POSITION

Target:

  • Increase model capacity at the end (add layer after GAP)
  • Perform MaxPooling at RF=5 and using only one maxpooling layer

Results:

  • Parameters: 7.9k
  • Best Training Accuracy: 99.20%
  • Best Test Accuracy: 99.39%

Analysis:

  • Model is very good.
  • No overfitting
  • Still model is not able to get 99.4%

CODE 3: APPLY IMAGE AUGMENTATION AND FINE TUNE LEARNING RATE, ADD StepLR SCHEDULER

Target:

  • Add rotation, of (-7 to 7) degrees.
  • Add StepLR scheduler

Results:

  • Parameters: 7.9k
  • Best training accuracy = 99.30
  • Best Test accuracy = 99.47%

Analysis:

  • Model is awesome!!!
  • No overfittng
  • Target achieved