"""
Implements LIME_Global. Verifies if sensitives features have high contributions.
"""
from collections import Counter
import pandas as pd
import numpy as np

from lime import submodular_pick
from lime.lime_text import LimeTextExplainer

clusters = 50


def features_contributions(predict_fn, train, class_names, sample_size, kernel_width=5):

    explainer = LimeTextExplainer(class_names=class_names, kernel_width=kernel_width)
    if sample_size > len(train):
        sample_size = len(train)
    indexes = np.random.choice(range(sample_size), sample_size)
    explanations = [explainer.explain_instance(train[i].lower(), predict_fn, num_features=1000) for i in indexes]
    # sp_obj = submodular_pick.SubmodularPick(explainer, train, predict_fn, sample_size=sample_size,
    #                                         num_features=1000, clusters=clusters)
    # explanations = sp_obj.sp_explanations

    return explainer, explanations


def fairness_eval(model, train, max_features, sensitive_features, feature_names, class_names, sample_size, threshold=None):
    
    explainer, explanations = features_contributions(model.prob, train, class_names, sample_size)
    
    contributions = Counter()
    for exp in explanations:
        vocab = exp.domain_mapper.indexed_string.inverse_vocab
        words_weights = {vocab[i]: weight for i, weight in exp.local_exp[1]}
        a1 = Counter(words_weights)
        contributions.update(a1)
    
    if threshold != None and threshold > 0:
        actual_sensitive, is_fair, df = fairness_valid_threshold(contributions, feature_names, sensitive_features, threshold)
    else:
        actual_sensitive, is_fair, df = fairness_valid_top(contributions, feature_names, sensitive_features, max_features)
    
    
    return actual_sensitive, is_fair, df, explainer

def fairness_valid_top(contributions, feature_names, sensitive_features, max_features):
    
    actual_sensitive = []
    ans_data = []
    sorted_dict = sorted(contributions.items(), key=lambda x: abs(x[1]), reverse=True)

    if max_features is None or max_features < 0 :
        max_features = len(sorted_dict)
    
    for i in range(max_features):
        feature, value = sorted_dict[i]
        ans_data.append([i, feature, value])
        
        if feature in sensitive_features:
            actual_sensitive.append(feature)
    
    df = pd.DataFrame(ans_data, columns=["Index", "Word", "Contribution"])
    return actual_sensitive, len(actual_sensitive) < 2, df


def fairness_valid_threshold(contributions, feature_names, sensitive_features, threshold):
    
    actual_sensitive = []
    ans_data = []
    
    n_contributions = normalize(contributions)
    
    sorted_dict = sorted(n_contributions.items(), key=lambda x: abs(x[1]), reverse=True)
    
    for key,value in sorted_dict:
        if abs(value) < threshold:
            break
        ans_data.append([key,feature_names[key],value])
    
    for pair in ans_data:
        key = pair[0]
        if key in sensitive_features: 
            actual_sensitive.append(key)
    
    df = pd.DataFrame(ans_data, columns = ["Index", "Feature", "Contribution"])
    return actual_sensitive, len(actual_sensitive) < 2, df

def normalize(b):
    
    a = b.copy()
    
    values = [abs(x[1]) for x in a.items()]
    
#     values = list(map(abs, a.items()))
    minv = np.min(values)
    maxv = np.max(values)
    
    for key in a.keys():
        v = a[key]
        normalized = (abs(v) - minv) / (maxv - minv)   
        a[key] = normalized if v >= 0 else -normalized
    
    return a