import re
import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.tree import export_graphviz
from io import StringIO 
from IPython.display import Image
from sklearn.tree import plot_tree
import pydotplus
from IPython.display import Image
from sklearn.pipeline import Pipeline
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score
from sklearn.metrics import classification_report
from sklearn.metrics import confusion_matrix
from sklearn.model_selection import cross_val_score
from sklearn.preprocessing import OrdinalEncoder, LabelEncoder
from sklearn.compose import make_column_transformer
from sklearn.impute import SimpleImputer
pd.options.display.max_columns = None
pd.options.display.max_rows = None
import sklearn

def displayNumberOfNaNValues(df):
    # Create an empty list to store tuples of column index and number of NaN values
    na = []
    # Loop through each column in the DataFrame
    for index, col in enumerate(df):
        # Count the number of NaN values in each column and append the index and count to 'na'
        na.append((index, df[col].isna().sum())) 
    # Make a copy of 'na' and sort it based on the count of NaN values in descending order
    na.sort(key=lambda x: x[1], reverse=True) 
    # Iterate through the sorted list of columns
    for i in range(len(df.columns)):
        # Check if the count of NaN values for the current column is not zero
        if na[i][1] != 0:
            # Print the column name, count of NaN values, and "NaN"
            print(df.columns[na[i][0]], ":", na[i][1], "NaN")
    # Calculate and print the total number of features with NaN values
    print('Number of features with NaN values:', len([x[1] for x in na if x[1] > 0]))
    print("Total NaN in dataframe :" , df.isna().sum().sum())


df = pd.read_csv('archive/data.csv')

# Before April 2001, there were almost no rules in UFC (no judges, no time limits, no rounds, etc.). 
#It's up to this precise date that UFC started to implement a set of rules known as 
#"Unified Rules of Mixed Martial Arts".
#Therefore, we delete all fights before this major update in UFC's rules history.

# Using this old data would not be representative of current fights, especially since this 
#sport has become one of the most regulated due to its mixity and complexity.

limit_date = '2001-04-01'
df = df[(df['date'] > limit_date)]

# Display NaN values
displayNumberOfNaNValues(df)

# Define the list of important features to impute
imp_features = ['R_Weight_lbs', 'R_Height_cms', 'B_Height_cms', 'R_age', 'B_age', 'R_Reach_cms', 'B_Reach_cms']

# Initialize a SimpleImputer to impute missing values with median
imp_median = SimpleImputer(missing_values=np.nan, strategy='median')

# Iterate over each feature to impute missing values
for feature in imp_features:
    # Fit and transform the feature using median imputation
    imp_feature = imp_median.fit_transform(df[feature].values.reshape(-1,1))
    # Assign the imputed values back to the DataFrame
    df[feature] = imp_feature

# Impute missing values for 'R_Stance' using most frequent strategy
imp_stance_R = SimpleImputer(missing_values=np.nan, strategy='most_frequent')
imp_R_stance = imp_stance_R.fit_transform(df['R_Stance'].values.reshape(-1,1))

# Impute missing values for 'B_Stance' using most frequent strategy
imp_stance_B = SimpleImputer(missing_values=np.nan, strategy='most_frequent')
imp_B_stance = imp_stance_B.fit_transform(df['B_Stance'].values.reshape(-1,1))

# Create DataFrames for imputed stances
df['R_Stance'] = pd.DataFrame(imp_R_stance, columns=['R_Stance'])
df['B_Stance'] = pd.DataFrame(imp_B_stance, columns=['B_Stance'])

# drop B_avg_BODY_att values in the dataframe
    # List of features with NaN values to drop
    #na_features = ['B_avg_BODY_att', 'R_avg_BODY_att']

    # Drop rows with NaN values in specified features
    #df.dropna(subset=na_features, inplace=True)

# Drop columns 'Referee' and 'location' from the DataFrame
# The value of references and location has a low impact in battles, which makes it irrelevant to keep
df.drop(['Referee', 'location'], axis=1, inplace=True)

# Drop column 'B_draw' and 'R_draw' and 'Draw' fight and 'Catch Weight' fight
df.drop(['B_draw', 'R_draw'], axis=1, inplace=True)
df = df[df['Winner'] != 'Draw']
df = df[df['weight_class'] != 'Catch Weight']

# Remove column when data type is not float or int
dfWithoutString = df.select_dtypes(include=['float64', 'int64'])

plt.figure(figsize=(50, 40))
corr_matrix = dfWithoutString.corr(method='pearson').abs()
sns.heatmap(corr_matrix, annot=True)

# Show the correlation matrix of the dataframe
# Very laggy feature

# plt.show()

#  i = index of the fighter's fight, 0 means the last fight, -1 means first fight
def select_fight_row(df, name, i): 
    df_temp = df[(df['R_fighter'] == name) | (df['B_fighter'] == name)]  # filter df on fighter's name
    df_temp.reset_index(drop=True, inplace=True) #  as we created a new temporary dataframe, we have to reset indexes
    idx = max(df_temp.index)  #  get the index of the oldest fight
    if i > idx:  #  if we are looking for a fight that didn't exist, we return nothing
        return 
    arr = df_temp.iloc[i,:].values
    return arr
    
#  we get the last fight of Khabib :'(
print(select_fight_row(df, 'Khabib Nurmagomedov', 0))


# get all active UFC fighters (according to the limit_date parameter)
def list_fighters(df, limit_date):
    df_temp = df[df['date'] > limit_date]
    set_R = set(df_temp['R_fighter'])
    set_B = set(df_temp['B_fighter'])
    fighters = list(set_R.union(set_B))
    print("Number of fighter: "+str(len(fighters)))
    return fighters

# Last year when data fight was not full and correct
fighters = list_fighters(df,'2016-01-01')

def build_df(df, fighters, i):      
    arr = [select_fight_row(df, fighters[f], i) for f in range(len(fighters)) if select_fight_row(df, fighters[f], i) is not None]
    cols = [col for col in df] 
    df_fights = pd.DataFrame(data=arr, columns=cols)
    df_fights.drop_duplicates(inplace=True)
    df_fights['title_bout'] = df_fights['title_bout'].map({True: 1, False: 0})
    df_fights.drop(['R_fighter', 'B_fighter', 'date'], axis=1, inplace=True)
    return df_fights

df_train = build_df(df, fighters, 0)
df_test = build_df(df, fighters, 1)

# print(df_train.head(5))

preprocessor = make_column_transformer((OrdinalEncoder(), ['weight_class', 'B_Stance', 'R_Stance']), remainder='passthrough')

# If the winner is from the Red corner, Winner label will be encoded as 1, otherwise it will be 0 (Blue corner)
label_encoder = LabelEncoder()
y_train = label_encoder.fit_transform(df_train['Winner'])
y_test = label_encoder.transform(df_test['Winner'])

X_train, X_test = df_train.drop(['Winner'], axis=1), df_test.drop(['Winner'], axis=1)

# Random Forest composed of 100 decision trees. We optimized parameters using cross-validation and GridSearch tool paired together
random_forest = RandomForestClassifier(n_estimators=100, 
                                       criterion='entropy', 
                                       max_depth=10, 
                                       min_samples_split=2,
                                       min_samples_leaf=1, 
                                       random_state=0)

model = Pipeline([('encoding', preprocessor), ('random_forest', random_forest)])
model.fit(X_train, y_train)

# We use cross-validation with 5-folds to have a more precise accuracy (reduce variation)
accuracies = cross_val_score(estimator=model, X=X_train, y=y_train, cv=5)
print('Accuracy mean : ', accuracies.mean())
print('Accuracy standard deviation : ', accuracies.std())

y_pred = model.predict(X_test)
print('Testing accuracy : ', accuracy_score(y_test, y_pred), '\n')

target_names = ["Blue","Red"]
print(classification_report(y_test, y_pred, labels=[0,1], target_names=target_names))

# cm = confusion_matrix(y_test, y_pred) 
# ax = plt.subplot()
# sns.heatmap(cm, annot = True, ax = ax, fmt = "d")
# ax.set_xlabel('Actual')
# ax.set_ylabel('Predicted')
# ax.set_title("Confusion Matrix")
# ax.xaxis.set_ticklabels(['Blue', 'Red'])
# ax.yaxis.set_ticklabels(['Blue', 'Red'])
# plt.show()

feature_names = [col for col in X_train]
feature_importances = model['random_forest'].feature_importances_
indices = np.argsort(feature_importances)[::-1]
n = 30 # maximum feature importances displayed
idx = indices[0:n] 
std = np.std([tree.feature_importances_ for tree in model['random_forest'].estimators_], axis=0)

#for f in range(n):
#    print("%d. feature %s (%f)" % (f + 1, feature_names[idx[f]], feature_importances[idx[f]])) 

# plt.figure(figsize=(30, 8))
# plt.title("Feature importances")
# plt.bar(range(n), feature_importances[idx], color="r", yerr=std[idx], align="center")
# plt.xticks(range(n), [feature_names[id] for id in idx], rotation = 45) 
# plt.xlim([-1, n]) 
# plt.show()

# Sélectionnez un arbre de votre modèle
tree_estimator = model['random_forest'].estimators_[10]

# Tracez l'arbre
# plt.figure(figsize=(1, 1))
# plot_tree(tree_estimator, feature_names=df_train.columns, filled=True, rounded=True, fontsize=10)
# plt.savefig('tree.png', dpi=600)  # Enregistrez l'image au format PNG
# plt.show()

def predict(df, pipeline, blue_fighter, red_fighter, weightclass, rounds, title_bout=False): 
    try:
        #We build two dataframes, one for each figther 
        f1 = df[(df['R_fighter'] == blue_fighter) | (df['B_fighter'] == blue_fighter)].copy()
        f1.reset_index(drop=True, inplace=True)
        f1 = f1[:1]
        f2 = df[(df['R_fighter'] == red_fighter) | (df['B_fighter'] == red_fighter)].copy()
        f2.reset_index(drop=True, inplace=True)
        f2 = f2[:1]

        # if the fighter was red/blue corner on his last fight, we filter columns to only keep his statistics (and not the other fighter)
        # then we rename columns according to the color of  the corner in the parameters using re.sub()
        if (f1.loc[0, ['R_fighter']].values[0]) == blue_fighter:
            result1 = f1.filter(regex='^R', axis=1).copy() #here we keep the red corner stats
            result1.rename(columns = lambda x: re.sub('^R','B', x), inplace=True)  #we rename it with "B_" prefix because he's in the blue_corner
        else: 
            result1 = f1.filter(regex='^B', axis=1).copy()
        if (f2.loc[0, ['R_fighter']].values[0]) == red_fighter:
            result2 = f2.filter(regex='^R', axis=1).copy()
        else:
            result2 = f2.filter(regex='^B', axis=1).copy()
            result2.rename(columns = lambda x: re.sub('^B','R', x), inplace=True)
            
        fight = pd.concat([result1, result2], axis = 1) # we concatenate the red and blue fighter dataframes (in columns)
        fight.drop(['R_fighter','B_fighter'], axis = 1, inplace = True) # we remove fighter names
        fight.insert(0, 'title_bout', title_bout) # we add tittle_bout, weight class and number of rounds data to the dataframe
        fight.insert(1, 'weight_class', weightclass)
        fight.insert(2, 'no_of_rounds', rounds)
        fight['title_bout'] = fight['title_bout'].map({True: 1, False: 0})
        
        pred = pipeline.predict(fight)
        proba = pipeline.predict_proba(fight)
        if (pred == 1.0): 
            print("The predicted winner is", red_fighter, 'with a probability of', round(proba[0][1] * 100, 2), "%")
        else:
            print("The predicted winner is", blue_fighter, 'with a probability of ', round(proba[0][0] * 100, 2), "%")
        return proba
    except:
        print("One of fighter doesn't exist in the dataframe")
        return


predict(df, model, 'Kamaru Usman', 'Colby Covington', 'Welterweight', 3, True) 
predict(df, model, 'Leon Edwards', 'Belal Muhammad', 'Welterweight', 3, True)
predict(df, model, 'Conor McGregor', 'Khabib Nurmagomedov', 'Lightweight', 5, True)
predict(df, model, 'Conor McGregor', 'Tai Tuivasa', 'Heavyweight', 5, True)

predict(df,model,'Charles Oliveira','Conor McGregor','Lightweight',5,True)
predict(df,model,'Charles Oliveira','Khabib Nurmagomedov','Lightweight',5,True)
predict(df, model, 'Leon Edwards', 'Kamaru Usman', 'Welterweight', 5, True)