Neural Network Session Project

Directory Structure

To use train.py and visualisation.py, you need to be in the ./src folder.

.
├── .gitignore
├── AUTHORS.md                          <- The names of the project's authors
├── README.md                           <- This file
├── requirements.txt                    <- The required packages
├── results                             <- The generated results
└── src                                 <- The source code
    ├── models                          <- Base class, base blocks and implementation of the models
    │   ├── DNNBaseModel.py             <- The base class of which our model's inherit
    │   ├── DNNBlocks.py                <- The base blocks which make up our models
    │   ├── LeNet.py                    <- The LeNet neural network
    │   ├── ResNeXt.py                  <- WIP : The ResNeXt neural network
    │   └── VGGNet.py                   <- The VGGNet neural network
    ├── scenarii                        <- Main scenario and strategies
    │   ├── BaseScenario.py             <- The base class of which our scenarios inherit
    │   └── PoolBaseSamplingScenario.py <- The implementation of the pool sampling scenario and its strategies
    ├── scripts                         <- The scripts detailed below
    │   ├── launch_calcul_canada.sh     <- Used to launch runner.py on calcul canada
    │   ├── mean.py                     <- Used to compute the mean values of all results in a folder
    │   ├── mean-visualisation.py       <- Plots the computed mean values
    │   ├── runner.py                   <- Launches sequentially train.py instances
    │   └── visualisation.py            <- Plots the various metrics of a single result file
    ├── DataManager.py                  <- The manager of the DataLoader objects
    ├── train.py                        <- The main entry point of our project
    └── TrainTestManager.py             <- The class which hosts the active learning's logic

Requirements

This project is meant to be used with python 3.7.

Usage

Training :

The main entry point script is train.py. It launches one full active learning loop. To use it, type :

cd src (important !)

python3 scripts/train.py model dataset [parameters]

The parameters are :
--model : The model to use (values : vggnet / lenet) (default = vggnet).
--dataset : The dataset to use (values : CIFAR-10 / mnist) (default = cifar10).
--strategy : The data selection criterion (values : least_confidence / entropy_sampling / margin_sampling / random_sampling) (default = least_confidence).
--stopping_criterion_value : The quantity of data to add to the seed dataset (default = 100).
--seed_size : The size of the initial training dataset (default = 100).
--batch_szie : The batch size used when training the model (default = 20). Reduce this value if you have memory errors.
--query_size : The quantity of data added to the seed dataset after one iteration of the active learning's loop (default = 20).
--optimizer : The optimizer used to train the model (values : SGD or Adam) (default = Adam).
--num_epochs : The number of epochs to train each model (default = 10).
--validation : The share of data used for the validation dataset (default = 0.1).
--lr : The learning rate (default = 0.001).

To launch multiple instances of train.py one after the other, edit runner.py to fit your needs and simply launch it with :

cd src

python scripts/runner.py

Visualisation :

There are 3 scripts useful for data visualisation.

1. visualisation.py

   The first one is visualisation.py. It displays graphs with the following metrics : training and validation losses, training and validation accuracies.
   To launch it, use :

   cd src

   python scripts/visualisation.py [any number of result files]

2. mean.py

   To be able to plot the mean validation accuracy curves, you first need to compute the mean values using the mean.py script.
   To use it, type :

   cd src

   python scripts/mean.py [folder path]

   It will compute the mean of every result file of the given folder path and write the result inside the folder ./results/mean

3. mean-visualisation.py Lastly, this script will allow you to plot the mean learning curves with the variance around each curve. To use it, type :

   cd src

   python scripts/mean-visualisation.py [any number of mean result files]

License

This code is released under the MIT license. See the LICENSE file for more information.