"""
# -*- coding: utf-8 -*-
#
# Copyright 2021 Michael Büsch <m@bues.ch>
#
# Licensed under the Apache License version 2.0
# or the MIT license, at your option.
# SPDX-License-Identifier: Apache-2.0 OR MIT
#
"""

from mlplib.activation import *
from mlplib.backward import *
from mlplib.forward import *
from mlplib.gradient_check import *
from mlplib.init import *
from mlplib.loss import *
from mlplib.parameters import *

def test_gradient_check_ok():
    seed(42)

    inputs = 3
    layout = (6, 15, 5, 2)
    params = Parameters(
        weights=init_layers_weights(inputs, layout),
        biases=init_layers_biases(layout),
        actvns=[
            ReLU(),
            ReLU(),
            ReLU(),
            Sigmoid(),
        ],
    )
    x = standard_normal((10, inputs))
    y = standard_normal((10, layout[-1]))
    gradients, _ = backward_prop(x, y, params, MSE())

    ok = gradient_check(x, y, params, MSE(), gradients)
    assert ok

def test_gradient_check_notok_weight():
    seed(43)

    inputs = 3
    layout = (6, 15, 5, 2)
    params = Parameters(
        weights=init_layers_weights(inputs, layout),
        biases=init_layers_biases(layout),
        actvns=[
            ReLU(),
            ReLU(),
            ReLU(),
            Sigmoid(),
        ],
    )
    x = standard_normal((10, inputs))
    y = standard_normal((10, layout[-1]))
    gradients, _ = backward_prop(x, y, params, MSE())

    # Manipulate one weight gradient.
    gradients.dw[1][4][3] += 0.01

    ok = gradient_check(x, y, params, MSE(), gradients)
    assert not ok

def test_gradient_check_notok_bias():
    seed(44)

    inputs = 3
    layout = (6, 15, 5, 2)
    params = Parameters(
        weights=init_layers_weights(inputs, layout),
        biases=init_layers_biases(layout),
        actvns=[
            ReLU(),
            ReLU(),
            ReLU(),
            Sigmoid(),
        ],
    )
    x = standard_normal((10, inputs))
    y = standard_normal((10, layout[-1]))
    gradients, _ = backward_prop(x, y, params, MSE())

    # Manipulate one bias gradient.
    gradients.db[1][0][3] += 0.01

    ok = gradient_check(x, y, params, MSE(), gradients)
    assert not ok

# vim: ts=4 sw=4 expandtab