Skip to content

Latest commit

 

History

History
129 lines (82 loc) · 4.91 KB

README.md

File metadata and controls

129 lines (82 loc) · 4.91 KB
Raw

Various Simple Examples of Deep Learning / Vision

Deep Learning

Vision

For the edition of code, we will use Visual Code which is light, efficient, configurable and works on Linux, Windows, Mac ! It comes with Anaconda as well. See https://code.visualstudio.com/docs/languages/python

Anaconda install python and Manage environments https://www.anaconda.com/distribution/

Installation

Install Anaconda: take the one with python 3 and include the installation of Visual Code which works on Linux, Mac, Windows ! One installed, launch an Anaconda prompt and create your working env nammed 'p37' with Python 3.7 (you can choose another version).

  • conda create -n p37 python=3.7
  • activate p37

Install every packages with only 'conda install' or with 'pip install' but try to not mix both, specially for basic packages like numpy, matplotlib, etc. The '-c channel' is used to select the source channel of the packages. Standard are 'anaconda' or 'conda-forge'.

NumPy

  • conda install -c anaconda numpy

MatPlotLib, OpenCv, SciPy, SciKit Image, Pillow, PyQt, pygame, etc. (many of then are optional, depending what you want to do)

  • conda install -c anaconda numpy pylint jupyter scikit-learn scikit-image pillow pyqt
  • conda install -c conda-forge matplotlib opencv
  • conda install -c cogsci pygame

DLib

  • conda install -c menpo dlib or
  • pip install cmake
  • pip install dlib

Next chose which framework you want to use: TensorFlow with Keras or PyTorch. We recommand PyTorch.

PyTorch

  • To install PyTorch use the pip install commands provided in the web site of PyTorch

TensorFlow

  • pip install --ignore-installed --upgrade tensorflow or
  • pip install --ignore-installed --upgrade tensorflow-gpu

Keras

  • conda install -c conda-forge keras

Visual Code

Install and configure Visual Code: Open the Command Pallete (Ctrl/Cmd + Shift + P) and type “Python: Select Interpreter” See here: https://medium.com/@udiyosovzon/how-to-activate-conda-environment-in-vs-code-ce599497f20d

Classifier in src/classifier

#Classify 2D points (x,y).

Start to play with the playground

All examples are in 'src/classifier'. In these examples, the points upper than sin(x) are from class 1 (y>sin(x)), points lower than sin(x) are from class 2 (y<sin(x)).

  • classifier_empty.py: just generate the sample batches and draw the points cloud. It is the starting empty program to use to learn to write your own classifier.

  • classifier_keras.py: a simple classifier with keras: simple network that can be improve

  • classifier_keras_v2-functionnal.py: more elaborate network, functionnal representation of keras

  • classifier_keras_v2-sequential.py: more elaborate network, sequential representation of keras

  • classifier_pytorch.py

#Images

  • classifier_image_CNN_keras: Convolution neural network (ConvNet) with keras
  • classifier_generator_tf_mnist.py (BEGONIN Florian 11400915 / GRANDJEAN Valentin 11402835) Classification ET génération d'image (une sorte d'auto-encoder sur des images)

L'execution du fichier mnist.py permet d'entrainer un réseau neuronal convolutif sur la base de données MNIST. Après la phase d'entrainement des réseaux neuronaux il est possible de vérifier les résultats obtenus. Taper 1 dans la console testera le classifier en lui transmettant une image. Le résultat affiché correspond à l'image ayant été choisie dans la base de donnée, son label ainsi que les prédictions du réseau. Taper 2 dans la console testera le réseau inverse en lui demandant de créer une représentation de chaque chiffre puis passera cette image dans le classifier pour valider la reconaissance de chaque image. Taper 3 mettra fin à l'execution du programme.

GAN in src/gan

  • classifier_generator_tf_mnist: learn a classifier + a generator of images from the number as input
  • gan.py: GAN learn to generate 2D points (x,y) positionned above sin(x). => does not work yet !

Image Processing in src/image

  • dlib_facial_landmarks.py
  • dlib_video_facial_landmarks.py
  • image_warping.py => test of DLib to get landmarks on faces from photos and video

Optimization samples in src/optimization

  • cma_test: an simple example of optimization of the function f(x,y) = x^2+y^2 => it finds x=0 and y=0 as minimum

Alpha0 on a starship mini game

  • starship.py: the mini game. The score of each particles depend on the distance to the target. Action is 2D forces applyed to the particle.
  • mcts.py: the continuous mcst algo (it is a really poor implementation of mcts)
  • alpha0.ps: use mcts.py to compute couple (observation, action), the mcts provides several config, each node of the tree. And train a network to act according to the osbservation.