Image Classification Project

Table of Contents

• Introduction
• Dependencies
• How to Run
• Main Function
• Config Files
• Sweep Config Files
• Authors

Introduction

This project is designed to perform image classification using different Convolutional Neural Network (CNN) models in Keras. It supports three functionalities:

1. Inference and Class Activation Maps

2. Training

3. Knowledge Distillation

Dependencies

The following python packages are required to run this project:

• tensorflow-gpu
• keras
• wandb
• opencv-python
• numpy
• yaml

How to Run

To run this project, use the following command in your terminal:

python main.py --config <path to yaml config file> --wandb <name of wandb project> --inference <True/False> --distilation <True/False>

The config file is used to specify the hyperparameters, data paths, etc. The --wandb flag is used to specify the name of the W&B project, and if set to None, W&B will not be used. The --inference flag is used to run the project in inference mode and the --distilation flag is a boolean flag to train the specified student network from the the knowledge of a teacher one.

To run hyperparameter search using Sweep, use the following command in your terminal:

python main_sweep.py --sweep=<path_to_yaml> --n_runs=<num_runs> --wandb=<project_name> --inference=<inference_mode> --distilation=<distilation_mode>

The code can be run by calling python main_sweep.py and passing in the required parameters. The user can specify the sweep configuration file (--sweep), the number of runs (--n_runs), the name for the WandB project (--wandb), whether to run the code in inference mode (--inference), and whether to run the code in distillation mode (--distilation). The code will automatically train the model using the specified parameters and save the final model to disk.

In the wandb project, you will find a Sweep section to analyise the hyperparameter search. For instance:

Config Files

To run the main, there are 3 neural networks created and possible to use, each with a different config file in config/folder. The different CNNs are:

• Default Keras ResNet50 with modified output layer for certain number of classes and default weigths to be specified in the config file. You can also select stages you want by selecting the desired output layer.

• Custom ResNet with the number of blocks and channel to be specified:

• ShuffleNetV2 with the scale_factor, bottleneck_ratio and shuffle_units to be specified:

It is important to know that the structure of the train and validation folder must be a seperate folder for each class.

Sweep config file

Sweep config file will overwrite the parameters of the main config file. In each run, sweep will automatically generate a value of each hyperparameter and train with these hyperparameters, in seek of maximizing or minimizing the metric specified, in a random way, bayesian or grid search.

To make hyperparameter search just add to the parameters, one with the same name as in the config file. For instance, if you want to try different preprocessing functions, in the sweep file write this:

parameters:
  dataset.preprocessing_function:
    values:
      - tf.keras.applications.resnet50.preprocess_input
      - preprocess_input

And in the config file will be overwritten with the chosen one:

dataset:
  preprocessing_function: tf.keras.applications.resnet50.preprocess_input

Authors

Josep Bravo (Me) and Miquel Romero (https://github.com/0Miquel)