import keras
from keras_resnet.models import ResNet50
from tensorflow.contrib.learn.python.learn.estimators._sklearn import accuracy_score
import face_recognition
print(keras.__version__)

import os
import cv2 as cv
import numpy as np

shape = (200, 200)

def get_face(img, name):
    t = face_recognition.face_locations(img, model='cnn')
    (top, right, bottom, left) = t[0]
    face = img[top:bottom, left:right]
    cv.imwrite("faces/"+name+"face.jpg",face)
    return face

def get_test_data(path_to_test_data):
    classes = os.listdir(path_to_test_data)
    x_train = []
    y_train = []
    for c in classes:
        imgs = os.listdir(os.path.join(path_to_test_data, c))
        for img_name in imgs:
            img = cv.imread(os.path.join(path_to_test_data, c, img_name))
            face = get_face(img, c+"_"+img_name)
            face = cv.resize(face, shape)
            x_train.append(face)
            y_train.append(int(c))
    y_train = keras.utils.to_categorical(y_train, num_classes=2)
    return np.asarray(x_train), y_train


def get_train_data(path_to_train_data):
    meta = open("data/meta.txt", "r")
    x_train = []
    y_train = []
    for v in meta.readlines():
        img_name, c = v.split(" ")
        img = cv.imread(os.path.join(path_to_train_data, img_name))
        img = cv.resize(img, shape)
        x_train.append(img)
        y_train.append(int(c))
    y_train = keras.utils.to_categorical(y_train, num_classes=2)

    return np.asarray(x_train), y_train


def get_model(shape, classes):
    x = keras.layers.Input(shape)

    model = ResNet50(x, classes=classes)
    model.compile(optimizer='adadelta',
                  loss='binary_crossentropy',
                  metrics=['accuracy'])
    # ResnetBuilder.build_resnet_50((shape[2], shape[1], shape[0]), classes)
    return model


def get_pretrained_model():
    from keras_retinanet import models
    model = models.load_model('faceglasses.h5', backbone_name="resnet50")
    return model
    # f = open("model_json.json", "r")
    # json = f.readline()
    # from keras.models import model_from_json
    # model = model_from_json(json, custom_objects={"Input":keras.layers.Input})
    # model.load_weights("glasses_w.h5")
    # return model


def load_data():
    return np.load("x_train.npy"), np.load("y_train.npy")

# 2.0.8
# x_train, y_train = get_train_data("/home/sergej2/MeGlass_120x120")
# np.save("x_train",x_train)
# np.save("y_train",y_train)


def fit_save_model():
    x_train, y_train = load_data()
    model = get_model(x_train[0].shape, 2)
    model.fit(x_train, y_train, epochs=1, batch_size=128, validation_split=0.2, shuffle=True)
    json = model.to_json()
    model_f = open("model_json.json", "w")
    model_f.write(json)
    model_f.close()
    model.save_weights("glasses_w.h5")


x_test, y_test = get_test_data("data/test")
model = get_pretrained_model()
#
# fit_save_model()

res = model.predict(x_test)
print("Accuracy")
print(accuracy_score(res, y_test))

# print(f1_measure(res, y_test))
# print(model.evaluate(x_test, y_test))