def read_cifar_batch(file):
    with open(file, 'rb') as fo:
        dict = pickle.load(fo, encoding='bytes')
    return (np.array(dict[b'data']).astype('float32'), np.array(dict[b'labels']).astype('int64') )

def read_cifar(path):
    data = []
    labels = []

    #Add the 5 batches
    for i in range(1,6):
        data_temp, labels_temps = read_cifar_batch(f'{path}/data_batch_{i}')
        data.append(data_temp)
        labels.append(labels_temps)

    #Add the test batches
    data_temp, labels_temps = read_cifar_batch(f'{path}/test_batch')
    data.append(data_temp)
    labels.append(labels_temps)

    #Concatenate all the batches to create a big one
    data = np.concatenate(data, axis = 0)
    labels = np.concatenate(labels, axis = 0)

    return(data, labels)

def split_dataset(data, labels, split):
    X_train, X_test, y_train, y_test = train_test_split(
    data, labels, test_size=(1 - split), random_state=0)

    return(X_train, X_test, y_train, y_test)

def distance_matrix(matrix1, matrix2):
    #X_test then X_train in this order
    sum_of_squares_matrix1 = np.sum(np.square(matrix1), axis=1, keepdims=True) #A^2
    sum_of_squares_matrix2 = np.sum(np.square(matrix2), axis=1, keepdims=True) #B^2

    dot_product = np.dot(matrix1, matrix2.T) # A * B (matrix mutliplication)
    
    dists = np.sqrt(sum_of_squares_matrix1 + sum_of_squares_matrix2.T - 2 * dot_product) # Compute the product
    return dists

def knn_predict(dists, labels_train, k):
    output = []
    # Loop on all the images_test
    for i in range(len(dists)):
        # Innitialize table to store the neighbors
        res = [0] * 10
        # Get the closest neighbors
        labels_close = np.argsort(dists[i])[:k]
        for label in labels_close:
            #add a label to the table of result
            res[labels_train[label]] += 1
        # Get the class with the maximum neighbors
        label_temp = np.argmax(res) #Careful to the logic here, if there is two or more maximum, the function the first maximum encountered
        output.append(label_temp)
    return(np.array(output))

def evaluate_knn(data_train, labels_train, data_test, labels_tests, k):
    dist = distance_matrix(data_test, data_train)
    result_test = knn_predict(dist, labels_train, k)

    #accuracy 
    N = labels_tests.shape[0]
    accuracy = (labels_tests == result_test).sum() / N
    return(accuracy)

def bench_knn():

    k_indices = [i for i in range(20) if i % 2 != 0]
    accuracies = []

    # Load data
    data, labels = read_cifar.read_cifar('image-classification/data/cifar-10-batches-py')
    X_train, X_test, y_train, y_test = read_cifar.split_dataset(data, labels, 0.9)

    # Loop on the k_indices to get all the accuracies
    for k in k_indices:
        accuracy = evaluate_knn(X_train, y_train, X_test, y_test, k)
        accuracies.append(accuracy)
    
    # Save and show the graph of accuracies
    fig = plt.figure()
    plt.plot(k_indices, accuracies)
    plt.title("Accuracy as function of k")
    plt.show()
    plt.savefig('image-classification/results/knn_batch_1.png')
    plt.close(fig)

def learn_once_mse(w1, b1, w2, b2, data, targets, learning_rate):

    N_out = len(targets) #number of training examples

    # Forward pass
    a0 = data # the data are the input of the first layer
    z1 = np.matmul(a0, w1) + b1  # input of the hidden layer
    a1 = sigmoid(z1)  # output of the hidden layer (sigmoid activation function)
    z2 = np.matmul(a1, w2) + b2  # input of the output layer
    a2 = sigmoid(z2)  # output of the output layer (sigmoid activation function)
    predictions = a2  # the predicted values are the outputs of the output layer

    # Compute loss (MSE)
    loss = np.mean(np.square(predictions - targets))
    print(f'loss: {loss}')
    # print('shape a1', a1.shape)
    # print('shape w1', w1.shape)
    # print('shape b1', b1.shape)

    # print('shape a2', a2.shape)
    # print('shape w2', w2.shape)
    # print('shape b2', b2.shape)
   
    # Backpropagation
    delta_a2 = 2 / N_out * (a2 - targets)
    # print('shape delta_a2', delta_a2.shape)
    delta_z2 = delta_a2 * (a2 * (1 - a2)) 
    # print('shape delta_z2', delta_z2.shape)
    delta_w2 = np.dot(a1.T, delta_z2)
    # print('shape delta_w2', delta_w2.shape)
    delta_b2 = delta_z2

    delta_a1 = np.dot(delta_z2, w2.T)
    # print('shape delta_a1', delta_a1.shape)
    delta_z1 = delta_a1 * (a1 * (1- a1))
    # print('shape delta_z1', delta_z1.shape)
    delta_w1 = np.dot(a0.T, delta_z1)
    # print('shape delta_w1', delta_w2.shape)
    delta_b1 = delta_z1

    # Update weights and biases
    w2 -= learning_rate * delta_w2
    b2 -= learning_rate * np.sum(delta_b2, axis = 0, keepdims = True)

    w1 -= learning_rate * delta_w1
    b1 -= learning_rate * np.sum(delta_b1, axis = 0, keepdims = True)

    return w1, b1, w2, b2, loss

def one_hot(labels):
    #num_classes = np.max(labels) + 1 on va le hardcoder ici
    num_classes = 10
    one_hot_matrix = np.eye(num_classes)[labels]
    return one_hot_matrix

def cross_entropy_loss(y_pred, y_true):
    loss = -np.sum(y_true * np.log(y_pred)) / len(y_pred)
    return loss

def learn_once_cross_entropy(w1, b1, w2, b2, data, labels_train, learning_rate):

    N_out = len(labels_train) #number of training examples

    # Forward pass
    a0 = data # the data are the input of the first layer
    z1 = np.matmul(a0, w1) + b1  # input of the hidden layer
    a1 = sigmoid(z1)  # output of the hidden layer (sigmoid activation function)
    z2 = np.matmul(a1, w2) + b2  # input of the output layer
    a2 = softmax_stable(z2)  # output of the output layer (sigmoid activation function)
    predictions = a2  # the predicted values are the outputs of the output layer


    # Compute loss (cross-entropy loss)
    y_true_one_hot = one_hot(labels_train)
    loss = cross_entropy_loss(predictions, y_true_one_hot)


    # Backpropagation
    delta_z2 = (a2 - y_true_one_hot) 
    delta_w2 = np.dot(a1.T, delta_z2) / N_out # We divide by the sample size to have an average on the error and avoid big gradient jumps
    delta_b2 = delta_z2 / N_out


    delta_a1 = np.dot(delta_z2, w2.T)
    delta_z1 = delta_a1 * (a1 * (1 - a1))
    delta_w1 = np.dot(a0.T, delta_z1) / N_out
    delta_b1 = delta_z1 / N_out

    # Update weights and biases
    w2 -= learning_rate * delta_w2
    b2 -= learning_rate * np.sum(delta_b2, axis = 0, keepdims = True)

    w1 -= learning_rate * delta_w1
    b1 -= learning_rate * np.sum(delta_b1, axis = 0, keepdims = True)

    return w1, b1, w2, b2, loss

def forward(w1, b1, w2, b2, data):
    # Forward pass
    a0 = data # the data are the input of the first layer
    z1 = np.matmul(a0, w1) + b1  # input of the hidden layer
    a1 = sigmoid(z1)  # output of the hidden layer (sigmoid activation function)
    z2 = np.matmul(a1, w2) + b2  # input of the output layer
    a2 = softmax_stable(z2)  # output of the output layer (sigmoid activation function)
    predictions = a2  # the predicted values are the outputs of the output layer
    return(predictions)

def train_mlp(w1, b1, w2, b2, data_train, labels_train, learning_rate, num_epoch):
    train_accuracies = []
    for epoch in range(num_epoch):
        w1, b1, w2, b2, loss = learn_once_cross_entropy(w1, b1, w2, b2, data_train, labels_train, learning_rate)

        # Compute accuracy
        predictions = forward(w1, b1, w2, b2, data_train)
        predicted_labels = np.argmax(predictions, axis=1)
        # print(predictions.shape)
        # print(predicted_labels.shape)
        # print(labels_train.shape)
        accuracy = np.mean(predicted_labels == labels_train)
        train_accuracies.append(accuracy)

        print(f'Epoch {epoch + 1}/{num_epoch}, Loss: {loss:.3f}, Train Accuracy: {accuracy:.2f}')

    return w1, b1, w2, b2, train_accuracies

 
    # Compute accuracy
    predictions = forward(w1, b1, w2, b2, data_test)
    predicted_labels = np.argmax(predictions, axis=1)
    print(predicted_labels)
    test_accuracy = np.mean(predicted_labels == labels_test)
    print(f'Train Accuracy: {test_accuracy:.2f}')
    return test_accuracy

def run_mlp_training(data_train, labels_train, data_test, labels_test, d_h,learning_rate, num_epoch):

    d_in = data_train.shape[1]
    d_out = 10 #we can hard code it here or len(np.unique(label_train))

    #Random initialisation of weights
    w1 = np.random.randn(d_in, d_h)
    b1 = np.random.randn(1, d_h)

    w2 = np.random.randn(d_h, d_out)
    b2 = np.random.randn(1, d_out)

    # Train MLP
    w1, b1, w2, b2, train_accuracies = train_mlp(w1, b1, w2, b2, data_train, labels_train, learning_rate, num_epoch)

    # Test MLP
    test_accuracy = test_mlp(w1, b1, w2, b2, data_test, labels_test)
    return train_accuracies, test_accuracy