Train a multilayer perceptron (MLP) and a convolutional neural network (CNN) to classify handwritten digits (MNIST). Implementation PyTorch and Tensorflow with eager execution.

Requirements (example)

Originally based on Python 3.7.x. Works with Python 3.11.4 (July 2023)

• Pytorch 1.1.0 to 2.0.1 (see warning below for older version)

You should install using conda to avoid installing CUDA and CuDNN by yourself. However, you must install the NVIDIA drivers:

ubuntu-drivers devices # --> liste des drivers disponibles
sudo apt install nvidia-driver-535

Some Python packages are required as well:

• matplotlib
• NumPy (installed with tf/pytorch)

Warning with old PyTorch (January 2020, about version 1.3.1), torchvision is not compatible with PILLOW 7 ('PILLOW_VERSION' removed from 'PIL'). You need to downgrade it to e.g. version 6.2.1. With PyTorch 1.7.0, it works directly.

To only remove warnings, you can add this in the Python program:

import warnings
import matplotlib.cbook
warnings.filterwarnings("ignore",category=matplotlib.cbook.mplDeprecation)

Exercise

To work on the exercice use program mnist_MLP_CNN_pytorch_exercice.py that contains TODOs.

Run the program (correction)

For PyTorch, choose the model into mnist_MLP_CNN_pytorch.py (line 117--119) then run:

python mnist_MLP_CNN_pytorch.py

For those interested

For Tensorflow (2.0.0) and the MLP model:

cd code-tf2/src/
python train.py

For Tensorflow and the CNN model:

cd code-tf2/src/
python train.py --model cnn