# encoding: utf-8
"""
@author: monitor1379 
@contact: yy4f5da2@hotmail.com
@site: www.monitor1379.com

@version: 1.0
@license: GNU General Public License(Version 3)
@file: np_utils.py
@time: 2016/9/11 14:35


"""

import numpy as np


def to_one_hot(y, nb_class):
    n = np.shape(y)[0]
    new_y = np.zeros((n, nb_class), dtype=np.float64)
    new_y[range(n), y] = 1
    return new_y


def to_categorical(y):
    new_y = np.argmax(y, axis=1)
    return new_y


def split_training_data(data, nb_validation=None, split_ratio=None):
    x, y = data
    n = np.shape(x)[0]

    if nb_validation:
        if nb_validation >= n:
            raise Exception('Error: nb_validation is bigger than x.shape[0]!')
        else:
            idx = n - nb_validation
    else:
        if split_ratio:
            idx = int(n * split_ratio)
        else:
            raise Exception('Error: nb_validation and split_ratio can not '
                            'be None both!')
    training_x = x[:idx]
    training_y = y[:idx]
    validation_x = x[idx:]
    validation_y = y[idx:]
    training_data = (training_x, training_y)
    validation_data = (validation_x, validation_y)
    return training_data, validation_data


def padding(images, pad_size):
    """
    对每一个图片的每一个通道上的二维图像进行padding填充操作，
    支持HW/CHW/NCHW三种格式的images。
    """
    top, bottom, left, right = pad_size
    pad_width = ((0, 0),) * (images.ndim - 2) + ((top, bottom), (left, right))
    return np.pad(images, pad_width=pad_width, mode='constant', constant_values=0)


def rot180(images):
    """
    旋转图片180度。
    支持HW/CHW/NCHW三种格式的images。
    """
    out = np.empty(shape=images.shape, dtype=images.dtype)
    if images.ndim == 2:
        out = np.rot90(images, k=2)
    elif images.ndim == 3:
        for c in xrange(images.shape[0]):
            out[c] = np.rot90(images[c], k=2)
    elif images.ndim == 4:
        for n in xrange(images.shape[0]):
            for c in xrange(images.shape[1]):
                out[n][c] = np.rot90(images[n][c], k=2)
    else:
        raise Exception('Invalid ndim: ' + str(images.ndim) +
                        ', only support ndim between 2 and 4.')
    return out


# 给定函数f，自变量x以及对f的梯度df，求对x的梯度
def eval_numerical_gradient_array(f, x, df=None, verbose=False, h=1e-4):
    it = np.nditer(x, flags=['multi_index'])
    grad = np.zeros_like(x)
    while not it.finished:
        xi = it.multi_index
        old_val = x[xi]

        x[xi] = old_val + h
        y_top = f(x)

        x[xi] = old_val - h
        y_bottom = f(x)

        if df is not None:
            grad[xi] = np.sum((y_top - y_bottom) * df) / (2 * h)
        else:
            grad[xi] = np.sum(y_top - y_bottom) / (2 * h)

        if verbose:
            print it.multi_index, grad[xi]
        x[xi] = old_val
        it.iternext()

    return grad


def sum_abs_err(u, v):
    return np.sum(np.abs(u - v))