#!/usr/bin/env python
# coding: utf-8

# # ロジステック回帰 - pymc3

# In[8]:


from __future__ import division
import os
import sys
import glob
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
get_ipython().run_line_magic('matplotlib', 'inline')
get_ipython().run_line_magic('precision', '4')
#plt.style.use('ggplot')
import seaborn as sns
sns.set_style('white')
sns.set_context('paper')

np.random.seed(1234)
import pymc3 as pm
import scipy.stats as stats

import logging
_logger = logging.getLogger("theano.gof.compilelock")
_logger.setLevel(logging.ERROR)


# ## ロジステック回帰による女性・男性決定問題
# 
# - 身長，体重データを利用して女性・男性を決定する問題を扱う．
# - pymc3のexample/dataフォルダにあるデータを利用する．
# - データはpandasのデータフレームとして渡す(データの単位は不明…．foot/poundでない)

# In[9]:


# observed data
df = pd.read_csv('data/HtWt.csv')
df.head()


# In[10]:


get_ipython().run_line_magic('pinfo', 'pm.glm.glm')


# In[24]:


niter = 1000
with pm.Model() as model:
    pm.glm.glm('male ~ height + weight', df, family=pm.glm.families.Binomial())
    trace = pm.sample(niter, step=pm.Slice(), random_seed=123, progressbar=True)
    pm.traceplot(trace)


# In[25]:


# note that height and weigth in trace refer to the coefficients
df_trace = pm.trace_to_dataframe(trace)
pd.scatter_matrix(df_trace[-1000:], diagonal='kde');


# In[26]:


plt.figure(figsize=(12, 4))
plt.subplot(121)
plt.plot(df_trace.ix[-1000:, 'height'], linewidth=0.7)
plt.subplot(122)
plt.plot(df_trace.ix[-1000:, 'weight'], linewidth=0.7);


# In[14]:


pm.summary(trace);


# In[27]:


#import seaborn as sn
sns.kdeplot(trace['weight'], trace['height'])
plt.xlabel('Weight', fontsize=20)
plt.ylabel('Height', fontsize=20)


# In[16]:


intercept, height, p, p_logodds, weight = df_trace[-niter//2:].mean(0)

def predict(w, h, height=height, weight=weight):
    """Predict gender given weight (w) and height (h) values."""
    v = intercept + height*h + weight*w
    return np.exp(v)/(1+np.exp(v))

# calculate predictions on grid
xs = np.linspace(df.weight.min(), df.weight.max(), 100)
ys = np.linspace(df.height.min(), df.height.max(), 100)
X, Y = np.meshgrid(xs, ys)
Z = predict(X, Y)

plt.figure(figsize=(6,6))

# plot 0.5 contour line - classify as male if above this line
plt.contour(X, Y, Z, levels=[0.5])

# classify all subjects
colors = ['lime' if i else 'yellow' for i in df.male]
ps = predict(df.weight, df.height)
errs = ((ps < 0.5) & df.male) |((ps >= 0.5) & (1-df.male))
plt.scatter(df.weight[errs], df.height[errs], facecolors='red', s=150)
plt.scatter(df.weight, df.height, facecolors=colors, edgecolors='k', s=50, alpha=1);
plt.xlabel('Weight', fontsize=16)
plt.ylabel('Height', fontsize=16)
plt.title('Gender classification by weight and height', fontsize=16)
plt.tight_layout();