benchmark-tensorflow

Little tool written in Python to try different configurations of TensorFlow Machine-learning models, to see what parameters can improve the most the accuracy of the model.

The model uses the well-known MNIST dataset (hand written digits https://yann.lecun.com/exdb/mnist/).

Usage

Install python3, and add modules tensorflow and matplotlib pip install tensorflow matplotlib

You can then run the benchmark.py file python3 benchmark.py

You can try different configuration, changing these paremeters:

• number of training epochs
• number of layers
• number of neurons in each layer

Edit configuration by modifiying configurations.py:

# So we see if results are different for small amount of training and big amount of training
number_of_training_epochs = [5, 11] # All configurations will be benchmarked for all number of epochs


configurations_to_try: list[Configuration] = [
        Configuration([
            Layer(10)
        ]),
        Configuration([
            Layer(50),
        ]),
        Configuration([
            Layer(100),
        ]),
        Configuration([
            Layer(10),
            Layer(10),
        ]),
        Configuration([
            Layer(50),
            Layer(50),
        ]),
        Configuration([
            Layer(100),
            Layer(100),
        ]),
        Configuration([
            Layer(10),
            Layer(10),
            Layer(10)
        ]),
]

The data is then displayed in a bar chart using the matplotlib library