# -*- coding: utf-8 -*-
"""
Created on Wed Dec 28 14:46:30 2016

@author: valerie
"""
from sklearn.cross_validation import train_test_split
import matplotlib.pyplot as plt
#from matrix_factorization_soln import MatrixFactorizationRec
#from matrix_factorization_soln import MatrixFactorizationALS
from sklearn.linear_model import BayesianRidge, LinearRegression
from sklearn import cross_validation as cv

# On utilise les données lue dans rec_runner.py
n_users = max(ratings_data_contents.user)
n_movies = max(ratings_data_contents.movie)
print 'Number of users = ' + str(n_users) + ' | Number of movies = ' + str(n_movies)  


train_data1, test_data = cv.train_test_split(ratings_data_contents,test_size=0.5)
train_data, valid_data = cv.train_test_split(train_data1,test_size=0.25)


train_data = pd.DataFrame(train_data)
valid_data = pd.DataFrame(valid_data)
test_data = pd.DataFrame(test_data)

# Create training and test matrix
R = np.zeros((n_users, n_movies))
for line in train_data.itertuples():
    R[line[1]-1, line[2]-1] = line[3]  
Rval = np.zeros((n_users, n_movies))
for line in valid_data.itertuples():
    Rval[line[1]-1, line[2]-1] = line[3]  

T = np.zeros((n_users, n_movies))
for line in test_data.itertuples():
    T[line[1]-1, line[2]-1] = line[3]
    
    
# Index matrix for training data
I = R.copy()
I[I > 0] = 1
I[I == 0] = 0

 # Index matrix for validation data
Ival = Rval.copy()
Ival[Ival > 0] = 1
Ival[Ival == 0] = 0
 
# Index matrix for test data
I2 = T.copy()
I2[I2 > 0] = 1
I2[I2 == 0] = 0
  
### ===========================================================================
### Matrix Factorization, SGD
MF = MatrixFactorizationRec_v1(10,0.007,0.02,.5,max_iter=1000)
MF.fit(R,Rval)
MF.rmse
MF.rmse_val

# Si on veut reprendre l'estimation là où on est arrivé  
P = MF.user_mat
Q = MF.movie_mat
MF = MatrixFactorizationRec_v1(10,0.007,0.02,.3,max_iter=4000)
MF.fit(R,Rval,P_init=P,Q_init=Q)
MF.fit(R,Rval,P_init=MF.user_mat,Q_init=MF.movie_mat)

user_test = 3200
p1 = np.squeeze(np.asarray(Rpred1[user_test,:]))
plt.plot(p_test,p1,'.')

p_test = T[user_test,:]
p_pred = np.multiply(p1,I2[user_test,])
plt.plot(p_test,p_pred,'.')

np.sqrt(np.mean(np.power(p1-p_test,2)))


### ===========================================================================
### BPMF
bpmf1 = BPMF(n_users+1, n_movies+1, 10,
             max_rating=5,
             min_rating=1,
             seed=None)
#filename = '/Users/valerie/Dropbox/ENSEIGNEMENT/EXPOSES/2016-2017/MOVIES/BPMF1.npy'
#np.save(filename,bpmf1)
ratings = np.array(train_data.drop("id",axis=1))
ratings_val = np.array(valid_data.drop("id",axis=1))
bpmf1.fit(ratings, ratings_val,n_iters=30)
plt.plot(range(30),bpmf1.all_rmse,range(30),bpmf1.all_rmse_val)
rmse_1 = RMSE(bpmf1.predict(ratings[:, :2]), ratings[:, 2])
train_preds = bpmf1.predict(ratings[:, :2])
train_rmse = RMSE(train_preds, ratings[:, 2])
train_rmse

rat_test =  np.array(test_data.drop("id",axis=1))
test_preds = bpmf1.predict(rat_test[:, :2])
test_rmse = RMSE(test_preds, rat_test[:, 2])
test_rmse

rat_test[:,2] = test_preds
rat_test_df = pd.DataFrame(rat_test,columns=["user","movie","rating"])

pref = test_data.rating[test_data.user==user_test]
ppfm = rat_test_df.rating[rat_test_df.user==user_test]
np.sqrt(np.mean((np.array(pref)-np.array(ppfm))**2))
plt.plot(range(len(ppfm)),pref,range(len(ppfm)),ppfm)

plt.plot(pref,ppfm,'.')
         
         
### ===========================================================================
## Linear regression
indices = train_data['user']
age = user_df['age'][indices-1]
sex = user_df['sex'][indices-1]
csp = pd.Series(user_df['csp'][indices-1], dtype="category")
X_user = np.array([age,sex])
indices = valid_data['user']
age = user_df['age'][indices-1]
sex = user_df['sex'][indices-1]
X_user_val = np.array([age,sex])

y = train_data['rating']
ols_user = LinearRegression()
ols_user.fit(X_user.T, y)
Rtmp = ols_user.intercept_+np.dot(X_user.T,ols_user.coef_)
rmse = np.sqrt(np.mean(np.power(y-Rtmp,2)))
y1 = y-Rtmp
y_val = valid_data['rating']
 


indices = train_data['movie']
X_movie = movie_df.as_matrix()
X_movie = X_movie[indices-1,1:20]
indices = valid_data['movie']
X_movie_val = movie_df.as_matrix()
X_movie_val = X_movie_val[indices-1,1:20]
clf_movie = BayesianRidge(compute_score=True)
clf_movie.fit(X_movie, y1)
Rtmp = np.dot(X_movie,clf_movie.coef_)
rmse = np.sqrt(np.mean(np.power(y1-Rtmp,2)))

ols_movie = LinearRegression()
ols_movie.fit(X, y1) # même résultat
Rtmp = ols_movie.intercept_+np.dot(X,ols_movie.coef_)
rmse = np.sqrt(np.mean(np.power(y1-Rtmp,2)))
ratings = np.array(train_data.drop("id",axis=1))
ratings_val = np.array(valid_data.drop("id",axis=1))
ratings_c= ratings
ratings_c[:,2]  = y1-Rtmp
ratings_c_val =  ratings_val 
correct = (ols_user.intercept_+np.dot(X_user.T,ols_user.coef_))  \
                 +(ols_movie.intercept_+np.dot(X_movie,ols_movie.coef_))     
 
correct_val = (ols_user.intercept_+np.dot(X_user_val.T,ols_user.coef_))  \
                 +(ols_movie.intercept_+np.dot(X_movie_val,ols_movie.coef_))     
ratings_c_val[:,2] = valid_data['rating'] \
                 -(ols_user.intercept_+np.dot(X_user_val.T,ols_user.coef_))  \
                 -(ols_movie.intercept_+np.dot(X_movie_val,ols_movie.coef_))