Siamese Network for Face Verification and recognition

Here you will find an Implementation of a Siamese Network for face verification and face recognition in tensorflow. next I'll provide some explanation for you, So you can easily use this project for yourself. if you ever end up using this project That would be so cool if you mentioned my name :)

Folder Structure For The Data

your images must be in three folders named anchor, positive, negative all in one data folder:

some_path
-------------/data
--------------------/anchor
--------------------/negative
--------------------/positive

for example in my case some_path was: '/content/drive/MyDrive/bachelor_project'

data_setup.py

In this python script you will find the code that can create and return tensorflow datasets for train and test. you'll have to specify the necessary paths to your data according to the arguments. all the required preprocessings will be taken care of. the model needs two images as input and a '0' or '1' as output. here is some visualization of my data:

model.py

This script contains the siamese network model that will be trained. please take note that the model output returns both a distance vector and one single number tensor between [0,1].

predict.py

This one as its name suggests is for prediction part. it takes the model and a tuple of two images and returns two things :

• First the probability of those images being the same .
• Second returns 1 if it is more than 50% sure they are the same and 0 otherwise.

train.py

This script contains functions that can be used to train the model. you will have to provide the model, loss funnction, optimizer number of epochs, train data, validation data as input to start the training.

notebook folder

In this folder, you will find a .ipynb file that I initially used to train the model. everything is there in one notebook. if you want to check only one thing to understand all I've done, check this notebook.

How to get things done right away...

First import the create_dataset, siamese_network and train functions from the python scripts then you can modify the following code to instantly train this siamese network model:

train_data, val_data = create_dataset(anchor_dir = '/content/drive/MyDrive/bachelor_project/data/anchor',
               positive_dir = '/content/drive/MyDrive/bachelor_project/data/positive',
               negative_dir = '/content/drive/MyDrive/bachelor_project/data/negative',
               train_split = 0.8,
               batch_size = 8)

sn = siamese_network(input_shape = (224,224,3))

loss_fn = tf.losses.BinaryCrossentropy()
optimizer = tf.keras.optimizers.Adam()
history = train(sn,train_data,val_data,3,loss_fn,optimizer)