The Prerequisites

This section will guide you through installing the various prerequisites of the BYNN project.

Images

First and foremost, to create an object detection neural network, you will need images of the objects you wish to detect with your model… many images. Usually, the best the best accuracy of a model is obtained by using 1000s of images. In our case, we will use around 100 images, but each image often contains more than 300 instances of the object we wish to detect, cells.

Try to select images that are sharp and that show the object in the environment in which it will be detected by the model. Make sure that various type of the object are evenly represented throughout the image dataset. That is, if the objects you wish to detect are dogs, make sure to include images of dogs from various breeds so that the model learns the general concept of what a dog looks like and not only what a specific breed looks like.

These images may be placed in a folder for now. They will be processed further down the tutorial.

Python

First, let’s install Python. Python is the programming language we are going to use throughout this tutorial and is widely used in the scientific community for its various and incredibly useful packages. Although Python is pre-installed on many systems, this pre-installed version is often outdated. Hence, we will install a more recent version. The simplest way to do so is by going to the Python.org download page to download the latest version. You may click the following link and follow the install process for your OS (Operating System):

1
https://www.python.org/downloads/

Anaconda & Conda

We will install the necessary packages for our project in a compartmentalized conda environment using Anaconda. This makes it easy to keep track of the packages we install and keep them at a specific version. It also allows us to use a specific Python version for our project; certain Python packages are not yet compatible with the most recent versions of Python.

To install Anaconda (and conda), you may click the following link and follow the install process based on your OS:

1
https://docs.anaconda.com/anaconda/install/

Virtual Environment

Once Anaconda is installed, we will create an environment for the BYNN project. The following command creates the environment using conda¹:

1
conda create -n <name> python=<version>

The argument <name> specifies the name of the environment and the <version> tag specifies the version of Python we wish to use.

We will create one with the name “BYNN” and the latest 3.9.x Python version as it is the latest version compatible with the packages we will use as of now.

1
conda create -n BYNN python=3.9

Before continuing, let’s activate the environment:

1
conda activate <name>

where <name> is the name you specified using the -n flag when creating the environment. If you followed the tutorial, the command should be:

1
conda activate BYNN

IMPORTANT: For the rest of the tutorial, and of your project, the conda environment you created should be activated.

YOLOv5

With the conda environment activated, run the following commands to download the YOLOv5 project from github and install its requirements²:

1
2
3
git clone https://github.com/ultralytics/yolov5  # clone
cd yolov5
pip install -r requirements.txt  # install

The first line will clone (download) the github repository of the project on your computer. The second one will move you inside the folder you just downloaded. The third line will install the requirements for the YOLOv5 project.

Labelling/Annotation

Similarly to children, computers have to learn what an object is to gain the ability to identify it in any environment. In the case of this tutorial, that means teaching the computer what a cell in each stage of mitosis looks like. That is, we need to provide ground truth to our neural network so that it knows what each cell should look like. In another project, ground truth could be what a dog looks like versus a cat.

This teaching will be achieved through the annotation of images. Put simply, with a computer, we will have to manually identify the objects we wish to detect with the neural network. Two tools to identify the objects are presented here. Both of them are simple and free to use which is why they were chosen amongst others.

1. The first one, Roboflow, is an online platform that allows the user to label objects on their images and format their labels.
2. The second one, LabelMe, is an open-source software that is downloaded and may be used offline.

That said, LabelMe requires some tinkering to get the labels in the right format and does not provide the ability to perform other actions on the dataset. This is why, although LabelMe could be used to follow this tutorial, Roboflow will be used for its simplicity of use. A link to LabelMe will still be provided.

Roboflow

Here are the steps required to use Roboflow with a free account:

1. Go to app.roboflow.com and login or create an account.
2. On the projects tab, click on the “+ Add Workspace” button and follow the setup steps.
   • A Community workspace type is fine for this project. It creates a free but public project.
   • In the case of this tutorial, the workspace will be named BYNN.
3. Create a new project inside the workspace by clicking on “Create New Project”. Follow the steps for the creation process:
   1. Click the “Skip Tutorial” button.
   2. Select a project name (BYNN in our case).
   3. Select a licence for the Roboflow project (MIT in our case).
   4. For the project type, select “Object Detection (Bounding Boxes)”.
   5. Decide what your model will predict (cells in our case).

That’s it! Your Roboflow account is setup and ready for your project. Continue with the next steps!

LabelMe

Here is a link to the LabelMe Github repository for the ones whishing to use it throughout this tutorial: https://github.com/wkentaro/labelme#installation