import pandas as pd
from pyod.models.auto_encoder import AutoEncoder
import time
from alibi.explainers import KernelShap

random_state = np.random.RandomState(42)



# Generating fictional data
N_TRAIN_OBS = 19248
N_TEST_OBS = 60
N_FEATURES = 12
SUMMARY_SIZE = 15
SUMMARY_TYPE = 'KMEANS'

def generate_fake_data(N_TRAIN_OBS, N_TEST_OBS, N_FEATURES,
                      SUMMARY_SIZE, SUMMARY_TYPE):
    feature_dct = {}
    for i in range(N_FEATURES):
        if i == 5:
            feature_dct["F{}".format(i)] = feature_dct["F1"] + random_state.normal(0, 0.1, N_TRAIN_OBS)
        elif i == 10:
            feature_dct["F{}".format(i)] = feature_dct["F8"] + random_state.normal(0, 0.1, N_TRAIN_OBS)
        else:
            feature_dct["F{}".format(i)] = list(random_state.normal(0, 0.2, N_TRAIN_OBS))

    df_train = pd.DataFrame.from_dict(feature_dct)

    feature_dct = {}
    for i in range(N_FEATURES):
        if i == 0 or i == 7:
            print("{} is different".format(i))
            feature_dct["F{}".format(i)] = list(random_state.normal(1, 0.2, N_TEST_OBS))
        elif i == 5:
            feature_dct["F{}".format(i)] = feature_dct["F1"] + random_state.normal(0, 0.1, N_TEST_OBS)
        elif i == 10:
            feature_dct["F{}".format(i)] = feature_dct["F8"] + random_state.normal(0, 0.1, N_TEST_OBS)
        else:
            feature_dct["F{}".format(i)] = list(random_state.normal(0, 0.2, N_TEST_OBS))

    df_test = pd.DataFrame.from_dict(feature_dct)
    
    return df_train, df_test

df_train, df_test = generate_fake_data(N_TRAIN_OBS, N_TEST_OBS, N_FEATURES, SUMMARY_SIZE, SUMMARY_TYPE)

outliers_fraction = 0.001
hidden_neurons=[len(df_train.columns)-1]


# Training the classifier
def train_ANN(df_train, outliers_fraction, hidden_neurons):
    cls_ann = AutoEncoder(epochs=30, hidden_neurons=hidden_neurons, contamination=outliers_fraction, verbose = 0)
    cls_ann.fit(df_train)
    return cls_ann

cls_ann = train_ANN(df_train, outliers_fraction, hidden_neurons)

# Summarise SHAP for faster computation
if SUMMARY_TYPE == 'KMEANS':
    df_train_summary = shap.kmeans(df_train, SUMMARY_SIZE)
else:
    df_train_summary = shap.sample(df_train, nsamples=SUMMARY_SIZE, random_state=random_state)



# Compute SHAP values
explainer_ann = KernelShap(cls_ann.predict_proba, df_train, distributed_opts={'n_cpus': 6}) 
explainer_ann.fit(df_train_summary)
shapvals_ann = explainer_ann.explain(df_test.values)