import pandas as pd
import numpy as np

%load_ext rmagic

%%R
download.file("https://dl.dropbox.com/u/7710864/courseraPublic/samsungData.rda",destfile="../data/samsungData.rda",method="curl")
load("../data/samsungData.rda")
write.csv(samsungData,file="../data/samsungData.csv")

samsungData = pd.read_csv('../data/samsungData.csv')
samsungData = samsungData.drop(['Unnamed: 0'], axis=1)

samsungData.columns[:12]

samsungData['activity'].value_counts()

subj1 = samsungData[samsungData['subject'] == 1]

numericActivity = subj1.groupby('activity')

cols = {'laying' : 'b',
        'sitting' : 'g',
        'standing' : 'r',
        'walk' : 'c',
        'walkdown' : 'm',
        'walkup' : 'y'}

f, (ax1, ax2) = subplots(ncols=2)
f.set_size_inches(10, 5)

for act, df in numericActivity:
    ax1.scatter(df.index, df.ix[:,0], c=cols[act], label=act)
    ax2.scatter(df.index, df.ix[:,1], c=cols[act], label=act)

ax1.set_ylabel(samsungData.columns[0])
ax2.set_ylabel(samsungData.columns[1])
ax2.legend(loc='lower right')

f.tight_layout();

subj1.ix[:,:3].columns

from scipy.spatial.distance import pdist, squareform
from scipy.cluster.hierarchy import linkage, dendrogram

actlabels = pd.Categorical.from_array(subj1['activity'])

distanceMatrix = pdist(subj1.ix[:,:3])
dendrogram(linkage(distanceMatrix, method='complete'), 
           color_threshold=0.3, 
           leaf_label_func=lambda x: 'O' * (actlabels.labels[x] + 1),
           leaf_font_size=6)

f = gcf()
f.set_size_inches(8, 4);

subj1.ix[:,[9, 10]].columns

f, (ax1, ax2) = subplots(ncols=2)
f.set_size_inches(10, 5)

for act, df in numericActivity:
    ax1.scatter(df.index, df.ix[:,9], c=cols[act], label=act)
    ax2.scatter(df.index, df.ix[:,10], c=cols[act], label=act)

ax1.set_ylabel(samsungData.columns[9])
ax2.set_ylabel(samsungData.columns[10])
ax2.legend(loc='lower right')

f.tight_layout();

distanceMatrix = pdist(subj1.ix[:,9:12])
dendrogram(linkage(distanceMatrix, method='complete'), 
           color_threshold=0.3, 
           leaf_label_func=lambda x: 'O' * (actlabels.labels[x] + 1),
           leaf_font_size=6)

f = gcf()
f.set_size_inches(8, 4);

# a simple scale function to normalize a matrix
def scale(matrix):
    from numpy import mean, std
    return (matrix - mean(matrix, axis=0)) / std(matrix, axis=0)

U, D, Vt = np.linalg.svd(subj1.ix[:,:-2].apply(scale), full_matrices=False)

f, (ax1, ax2) = subplots(ncols=2)
f.set_size_inches(10, 5)

for lb, cl in zip(list(actlabels.levels), 'b g r c m y k'.split()):
    idx = subj1['activity'] == lb
    ax1.scatter(subj1.index[idx], U[idx,0], c=cl, label=lb)
    ax2.scatter(subj1.index[idx], U[idx,1], c=cl, label=lb)
    
ax1.set_ylabel('U[:,0]')
ax2.set_ylabel('U[:,1]')
ax2.legend(loc='lower right')

f.tight_layout();

plot(Vt[1,:], 'ok');

maxContrib = np.argmax(Vt[1,:])
maxContrib

distanceMatrix = pdist(subj1.take(range(9,12) + [maxContrib], axis=1))
dendrogram(linkage(distanceMatrix, method='complete'), 
           color_threshold=0.3, 
           leaf_label_func=lambda x: 'O' * (actlabels.labels[x] + 1),
           leaf_font_size=6)

f = gcf()
f.set_size_inches(8, 4);

samsungData.columns[maxContrib]

from scipy.cluster.vq import kmeans, vq

data = np.matrix(subj1.ix[:,:-2])

centers, _ = kmeans(data, 6, iter=100)
cluster, _ = vq(data, centers)

df = pd.crosstab(cluster, subj1['activity'])
df

idmax = np.argmax(df['laying'])

plot(centers[idmax,:10], 'ok')
ylabel('Cluster Center');

idmax = np.argmax(df['walk'])

plot(centers[idmax,:10], 'ok')
ylabel('Cluster Center');