Tensorflow Neural Network Quantization

This is a school project for the Advanced Computer Architectures class @PoliMi, 2017-2018.

Be careful, the repository contains also the datasets and all the model files.

You can see the main results of MNIST models evaluation looking at these notebooks: Model (a), Model (b)

The evaluation of CIFAR-10 models is also available: Model (c), Model (d)

---

Tensorflow is an open source framework for machine learning applications such as neural networks. We use official python API. The project is developed using:

• Ubuntu 16.04 LTS
• Python 2.7
• Tensorflow 1.8 version with CPU only support

We used a desktop pc assempled with:

• Inter Core i7 2600K
• 8 GB DDR3 1333 Mhz RAM
• Nvidia GTX 960 4GB

To install Tensorflow with CPU only support and jupyter notebook run:

pip install tensorflow
pip install jupyter

Then navigate to your code directory and start jupyter

cd path/to/repo
jupyter notebook

---

As mentioned in the project description we are not interested in the training phase. However, we had to train some models. In order to speed up the training phase we used a Docker image with Nvidia GPU support, officially provided by Tensorflow.

To setup docker and train your model on GPU:

• Install proprietary Nvidia driver
• Install docker
• Install nvidia-docker

Then pull the latest GPU docker image:

sudo nvidia-docker pull tensorflow/tensorflow:latest-gpu

Create the container with port and data binding

sudo nvidia-docker run -it  --name tensorflow_docker -d -p 8888:8888 -p 6006:6006 -v ~/tensorflow:/notebooks tensorflow/tensorflow:latest-gpu

To start the container:

sudo nvidia-docker container start tensorflow_docker

To stop the container:

sudo nvidia-docker container stop tensorflow_docker

To access Tensorflow docker container bash:

sudo nvidia-docker exec -it tensorflow_docker /bin/bash

To start Tensorboard (in container bash):

tensorboard --logdir /notebooks/logdir

To access jupyter notebook navigate to http://localhost:8888/tree

To access Tensorboard navigate to http://localhost:6006

---

Useful links:

Tensorflow for Poets

Tensorflow Documentation