Scribd
Upload
132 views

Deep Learning With PyTorch: Object Classification - Filliat Et Al

The document introduces a practical course on image classification using deep convolutional neural networks with PyTorch. It discusses using the CIFAR10 dataset containing images from 10 categories. A smaller neural network will be used due to the limited computing power. The objective is to understand convolutional network structure and parameter influences. Students will run a Jupyter notebook guiding them through dataset use and network construction. Reporting involves detailing tested networks, hyperparameters, and test accuracy in a PDF by February 2nd. Installation instructions for PyTorch and dependencies are also provided.

Uploaded by

John Smith
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
132 views

Deep Learning With PyTorch: Object Classification - Filliat Et Al

The document introduces a practical course on image classification using deep convolutional neural networks with PyTorch. It discusses using the CIFAR10 dataset containing images from 10 categories. A smaller neural network will be used due to the limited computing power. The objective is to understand convolutional network structure and parameter influences. Students will run a Jupyter notebook guiding them through dataset use and network construction. Reporting involves detailing tested networks, hyperparameters, and test accuracy in a PDF by February 2nd. Installation instructions for PyTorch and dependencies are also provided.

Uploaded by

John Smith
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Deep Learning with PyTorch: Object classification

David Filliat, Antoine Manzanera, Antonin Raffin, Natalia Díaz Rodríguez


18 janvier 2018

1 Introduction
In this practical course we will study different structures of deep convolutional neural networks to work
on image classification using the PyTorch 1 Python library. Material can be downloaded from :
http://perso.ensta-paristech.fr/~filliat/Courses/index.html

Due to the limitation of the machines power (specially the lack of dedicated GPUs), we will work with
a smaller network that is partially representative of real applications in order to test the capacity of deep
learning techniques. The main objective is to understand the structure of convolutional networks and the
influence of their parameters over the learning process.

2 Running the notebook


All required software should be installed on the school computers in a virtual machine on Python 3 that
you need to launch by running :
torchvision
If you want to use your own computer, see the installation at the end of this file.
Download the notebook from the address :
http://perso.ensta-paristech.fr/~filliat/Courses/Polys/cnn-tutorial.zip

You can then run the notebook by executing, in the main project folder :

jupyter notebook cifar10_tutorial.ipynb 2

The notebook will guide you through the use of the CIFAR10 database (Figure 1) with ten categories
of objects and the construction of an artificial neural network to classify them. Each image is 32x32 pixels
and examples of categories are planes, cars, birds, etc.
The data can be downloaded automatically by the notebook, or manually from https://drive.
google.com/file/d/1LrmhnOvuilYyJ_Gto_r0xjXlNFq8PYPM/view (put the data/ folder as a subfol-
der of the one containing cifar10_tutorial.ipynb.

3 Reporting
Send a pdf report to natalia.diaz@ensta-paristech.fr latest on 2nd Feb. detailing what you have
done in part III : explain what neural networks and hyper parameters configurations you tested, and the
resulting performances. Most importantly, report the accuracy on the image test set) and plot of losses. The
best performance will be used for a leader board. All used parameters must be reported (at least learning
rates, batch size, nr of epochs).
You can try to modify :
1. http://pytorch.org/tutorials/index.html
2. On some computers, you may need to add the –ip option : jupyter notebook -ip=127.0.0.1 cifar10_tutorial.ipynb

1
F IGURE 1 – Sample images from CIFAR 10 Database

— The neural network structure (kernel size, number of feature maps, number of convolutional layers,
size of fully connected layers, ...
— The size of the training set/validation set
— Batch size, number of epochs, learning rates

4 Installation
Everything should be installed on the school computers. If you want to use your own computer, here are
the installation instructions. The notebooks are tested in Python 3 but they should be Python 2 compatible
(if you go with Python 3, substitute all pip instructions with pip3) :

4.1 Windows
1. Download and install Anaconda (it includes Jupyter notebook)
https://www.anaconda.com/download/#download
2. Install PyTorch : run in Anaconda Prompt :
conda install -c peterjc123 pytorch
pip install torchvision

4.2 Linux / Mac Os


4.2.1 Anaconda Users
1. Download and install Anaconda (it includes Jupyter notebook)
https://www.anaconda.com/download/#download
2. Install PyTorch and Torch vision (here without CUDA support)
conda install pytorch torchvision -c soumith
3. We recommend you create one virtual environment for each of your projects to avoid different
external libraries versions break when you install new packages and dependencies. Create a conda
virtual environment with conda create env_name,
e.g. : conda create deeplearning
And activate it (install all project related libraries inside) :
source activate env_name
or deactivate when you no longer work on it :
source deactivate

2
4.2.2 Pip users (advanced)
1. Make sure Jupyter Notebook is installed
pip install -upgrade pip
pip install jupyter
2. Go to http://pytorch.org/ and select your configuration to get the installation command line.
For example, with Linux and python 3.5 :
pip3 install http://download.pytorch.org/whl/cu75/torch-0.2.0.post3-cp35-cp35m-manylinux1_
x86_64.whl
3. Install torchvision :
pip install torchvision

5 Q & A (Potential issues)


1. Q : If, when running the Jupyter notebook you get : ImportError: No module named jinja2
A : pip install jinja2
2. Q : Proxy problem installing pyTorch : ’pip._vendor.requests.exceptions.InvalidSchema:
Missing dependencies for SOCKS support.’ A : This was corrected by unsetting the ALL_PROXY :
printenv | grep -i proxy unset ALL_PROXY printenv | grep -i proxy
Q : Installing PyTorch may require you to be sudo : A : Use sudo -H to get the installations wor-
king : sudo -H pip install http://download.pytorch.org/whl/cu75/torch-0.2.0.post3-cp35-cp35m-man
-E option can also be used for commands under proxy, in which case, your /̃.bashrc file must
contain : export http_proxy=http://proxy.ensta.fr:8080 export https_proxy=http://proxy.ensta.fr:8
sudo -H pip install torchvision

Acknowledgement
This short practical is strongly based on the courses :
— Coursera DeepLearning.ai by Andrew Ng
https://www.coursera.org/learn/convolutional-neural-networks
More complete tutorials and courses can be found at :
— Python Numpy http://cs231n.github.io/python-numpy-tutorial/ and IPython tutorials
http://cs231n.github.io/ipython-tutorial/
— Andrej Karpathy course CS231n Convolutional Neural Networks for Visual Recognition http:
//cs231n.github.io/
— MILA tutorials https://github.com/mila-udem/welcome_tutorials/tree/master/pytorch
— PyTorch’s excellent tutorials : http://pytorch.org/tutorials/beginner/blitz/cifar10_
tutorial.html
— Applications with PyTorch : https://twitter.com/pytorch
— Learn PyTorch with no Deep Learning background with Sung Kim’s PyTorchZeroToAll
https://github.com/hunkim/PyTorchZeroToAll
— A guide to convolution arithmetic for deep learning, V. Dumoulin and Francesco Visin https:
//github.com/vdumoulin/conv_arithmetic.
— Deep Learning. I. Goodfellow, Y. Bengio and A. Courville, MIT Press, http://www.deeplearningbook.
org, 2016.
— Deep Learning with Python, Francois Chollet, Manning Publications. 2017, https://github.
com/fchollet/deep-learning-with-python-notebooks

You might also like