Tiny-Imagenet Challenge

EIP phase-1 Project

Goal: Achieve atleast 60% validation accuracy on Tiny-Imagenet Dataset in less than 200 epochs.

Challenges:

  • Due to lack of infrastructure, google colab was the only option available.

Approach:

  • Trained a custom made DenseNet Model (21 layers) after taking receptive field and object size into consideration.
  • The model has 5 DenseBlocks each with 3 convolution layers and 4 transition blocks in between. There is no 5th transition block.
  • Substitued Flattern + Dense with Global Average Pooling layer to enable progeressive resizing.
  • Trained the model on scaled down images (16x16 and 32x32) for initial few epochs to extract the essential features and later on original size to improve the further accuracy.
  • Train additionally on poorly performing classes identified using F1 score.
  • Triangular Cyclic LR for faster convergence.
  • 14 different type of image augmentation techniques.

Details:

  • Batch size: 320
  • Total Parameters: 8.9M
  • Learning Rate: MaxLR = 0.05, MinLR=0.001 and StepSize=2000

Results:

  • Validation Accuracy 62.95%
  • Epochs 95

You can find the notebook here