Execute models

• Extract the downloaded .tar file of the cifar-100 dataset

• Give permissions to script

chmod u+x script_name.sh

• Execute script

./script_name.sh

Models

HD_CNN Baseline

The initial baseline of the HD-CNN Paper

ResNet Baseline

A simple vanilla ResNet-52

Baseline Architecture

A modified ResNet-52 that in its middle a coarse prediction head is added and at the end a fine prediction head.

ResNet Attention

Same modified ResNet as "Baseline Architecture" but with an attention layer added in between as described below.

Algorithm

1. Define the Coarse Classifier (CC) and the Fine Classifier (FC) for ResNet-50

CC consists of:

• ResNet Block: First two ConvBlocks of ResNet-50 (Input layer - "conv2_block3_out")
  • Input: Image (32x32x3)
  • Output: RO1 feature matrix
• Attention Block: Create heatmap of the images
  • Input: Output RO1 (8x8x256)
  • Output: Input RI1 (8x8x256)
• Prediction Layer:
  • Input: Output RO1 (8x8x256)
  • Output: 20 softmax units (Coarse labels)

FC consists of:

• ResNet Block: Third to Fifth ConvBlocks of ResNet-50 ("conv3_block1_1_conv" - "conv5_block3_out")
  • Input: RI1 + 20 coarse prediction labels
  • Output: RO2 feature matrix
• Prediction Bloack:
  • Input: RO2
  • Output: 100 softmax units (Fine labels)

2. Train CC

3. Freeze CC, Train FC

4. Fine-tune CC + FC together