DropNet

Instructions

<b>Dependencies required:</b>

• tensorflow2.x
• keras (comes from tensorflow.keras)
• numpy
• copy
• sklearn
• matplotlib

<b>Setup requirements:</b>

• Set up a folder named 'Images' (same directory as the Jupyter Notebook) to store the output images from the model
• Set up a folder named 'Caches' (same directory as the Jupyter Notebook) to store the caches from the model

<b> Jupyter Notebooks: </b>

• There are three notebooks for various experiments
• DropNet.ipynb is the main notebook that is used for the experiments in the main paper
• DropNet-Tiny ImageNet.ipynb is the supplementary notebook used for additional experiments on Tiny ImageNet
• DropNet-APoZ is the supplementary notebook used for additional experiments to compare DropNet's performance to that of a similar data-driven approach, Average Percentage of Zeros(APoZ)

<b>Jupyter Notebook Section descriptions (DropNet.ipynb):</b>

• <b>Section 1: Generic Initialization</b>
  • Run all the code under Section 1. These are the code to define the main functions required for the experiments
• <b>Section 2: Load Cache to Print Graph (if required)</b>
  • Run the codes here only if you need to print graphs from a cahce in the Caches folder
• <b>Section 3: MNIST</b>
  • Run the codes here for experiments on the MNIST dataset.
    • Section 3.1: Customize Model
      • Allows you to customize the model used for the experiment.
    • Section 3.2: Evaluate Model
      • Allows you to run code to evaluate performance of various metrics.
    • Section 3.3: Oracle Comparison
      • Allows you to run code to evaluate performance of min, max and min_layer metrics against the oracle
    • Section 3.4: Random Initialization
      • Allows you to run code to compare performance of random initialization on the final pruned model vs original initialization
    • Section 3.5: Compare Percent
      • Allows you to run code to compare effect on accuracy based on proportion of nodes/filters dropped per training cycle
• <b>Section 4: CIFAR10</b>
  • Run the codes here for experiments on the CIFAR10 dataset
    • Section 4.1: Customize Model
      • Allows you to customize the model used for the experiment.
    • Section 4.2: Evaluate Model
      • Allows you to run code to evaluate performance of various metrics.
    • Section 4.3: Oracle Comparison
      • Allows you to run code to evaluate performance of min, max and min_layer metrics against the oracle
    • Section 4.4: Random Initialization
      • Allows you to run code to compare performance of random initialization on the final pruned model vs original initialization
    • Section 4.5: Compare Percent
      • Allows you to run code to compare effect on accuracy based on proportion of nodes/filters dropped per training cycle

<b>Jupyter Notebook Section descriptions (DropNet-Tiny ImageNet.ipynb):</b>

• <b>Section 1: Generic Initialization</b>
  • Run all the code under Section 1. These are the code to define the main functions required for the experiments
• <b>Section 2: Tiny ImageNet</b>
  • Run the codes here for experiments on the Tiny ImageNet dataset
    • Section 2.1: Customize Model
      • Allows you to customize the model used for the experiment.
    • Section 2.2: Evaluate Model
      • Allows you to run code to evaluate performance of various metrics.

<b>Jupyter Notebook Section descriptions (DropNet-APoZ.ipynb):</b>

• <b>Section 1: Generic Initialization</b>
  • Run all the code under Section 1. These are the code to define the main functions required for the experiments. The flattenmask() function has been updated here to compare with Average Percentage of Zeros (APoZ).
• <b>Section 2: CIFAR10</b>
  • Run the codes here for experiments on the CIFAR10 dataset
    • Section 2.1: Customize Model
      • Allows you to customize the model used for the experiment.
    • Section 2.2: Evaluate Model
      • Allows you to run code to evaluate performance of various metrics.