armgilles

Les IRIS

• Les communes d'au moins 10 000 habitants et une forte proportion des communes de 5 000 à 10000 habitants sont découpées en Iris. Ce découpage constitue une partition de leur territoire. La France compte environ 16 100 Iris dont 650 dans les DOM.

• 3 types d'IRIS :

  • Les Iris d'habitat : leur population se situe en général entre 1 800 et 5 000 habitants. Ils sont homogènes quant au type d'habitat et leurs limites s'appuient sur les grandes coupures du tissu urbain (voies principales, voies ferrées, cours d'eau...).

armgilles

armgilles

git init
git add -all
git commit -m "First commit"
git remote add Name Adresse
git push -u Name master

armgilles

Popularization :

Simple example

Sport example

Date with a woman & Bayes

Introduction of of-bayes-theorem

armgilles

sturges = lambda n: int(log2(n) + 1) square_root = lambda n: int(sqrt(n)) from scipy.stats import kurtosis doanes = lambda data: int(1 + log(len(data)) + log(1 + kurtosis(data) * (len(data) / 6.) ** 0.5))

n = len(titanic) sturges(n), square_root(n), doanes(titanic.fare.dropna())

titanic.fare.hist(bins=doanes(titanic.fare.dropna()))

armgilles

from sklearn.cluster import DBSCAN
import pandas as pd
# You already have your feature in X

dbscan_eps = []
for i in [x / 10.0 for x in range(1, 20, 1)]:
    db = DBSCAN(eps=i).fit(X)
    n_clusters_ = len(set(db.labels_)) - (1 if -1 in db.labels_ else 0)
    print "eps = " +" "+ str(i) +" "+ " cluster = " + str(n_clusters_)
    dbscan_eps.append({'eps' : i,

armgilles

from sklearn.cluster import AffinityPropagation
import pandas as pd
import sys
import cStringIO
# You already have your feature in X

aff_eps = []
for i in [x for x in range(-50, 0, 5)]:
    # To know caputre the output of verbose
    tdout_ = sys.stdout #Keep track of the previous value.

armgilles

Here comes Deep Learning !

Some intro

• what is deep learning ?

Books

• Neural Networks and Deep Learning is a free online book
• MIT Press book DRAFT

armgilles

List of partern for neuronal network :

1. Autoencoders are simplest ones. They are intuitively understandable, easy to implement and to reason about (e.g. it's much easier to find good meta-parameters for them than for RBMs).
2. RBMs are generative. That is, unlike autoencoders that only discriminate some data vectors in favour of others, RBMs can also generate new data with given joined distribution. They are also considered more feature-rich and flexible.
3. CNNs are very specific model that is mostly used for very specific task (though pretty popular task). Most of the top-level algorithms in image recognition are somehow based on CNNs today, but outside that niche they are hardly applicable (e.g. what's the reason to use convolution for film review analysis?).

Autoencoder

Autoencoder is a simple 3-layer neural network where output units are directly connected back to input units. E.g. in a network like this

armgilles