numpy-diffusion/diffusion/layers/batch_norm.py at master · AkiRusProd/numpy-diffusion · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
try:
    import cupy as np
    is_cupy_available = True
except:
    import numpy as np
    is_cupy_available = False


class BatchNormalization():
    """
    Applies batch normalization to the input data
    ---------------------------------------------
        Args:
            `momentum` (float): the momentum parameter of the moving mean
            `epsilon` (float): the epsilon parameter of the algorithm
        Returns:
            output: the normalized input data with same shape
        References:
            https://kevinzakka.github.io/2016/09/14/batch_normalization/

            https://agustinus.kristia.de/techblog/2016/07/04/batchnorm/
    """

    def __init__(self, momentum = 0.99, epsilon = 0.001, input_shape = None, data_type = np.float32):

        self.momentum = momentum
        self.epsilon  = epsilon

        self.gamma = None
        self.beta = None

        self.mean = None
        self.var = None

        self.moving_mean = None
        self.moving_var = None

        self.optimizer = None

        self.input_shape = input_shape
        self.data_type = data_type

        self.build()

    def set_optimizer(self, optimizer):
        self.optimizer = optimizer


    def build(self):
        self.gamma = np.ones(self.input_shape).astype(self.data_type)
        self.beta = np.zeros(self.input_shape).astype(self.data_type)


        self.vg, self.mg         = np.zeros_like(self.gamma).astype(self.data_type), np.zeros_like(self.gamma).astype(self.data_type)
        self.vg_hat, self.mg_hat = np.zeros_like(self.gamma).astype(self.data_type), np.zeros_like(self.gamma).astype(self.data_type)

        self.vb, self.mb         = np.zeros_like(self.gamma).astype(self.data_type), np.zeros_like(self.gamma).astype(self.data_type)
        self.vb_hat, self.mb_hat = np.zeros_like(self.gamma).astype(self.data_type), np.zeros_like(self.gamma).astype(self.data_type)

        self.output_shape = self.input_shape


    def forward(self, X, training = True):
        self.input_data = X
        self.batch_size = X.shape[0]

        if self.input_shape is None:
            self.input_shape = self.input_data.shape[1:]

            self.build()

        if self.moving_mean is None: self.moving_mean = np.mean(self.input_data, axis = 0)
        if self.moving_var is None: self.moving_var = np.var(self.input_data, axis = 0)

        if training == True:
            self.mean = np.mean(self.input_data, axis = 0)
            self.var = np.var(self.input_data, axis = 0)

            self.moving_mean = self.momentum * self.moving_mean + (1.0 - self.momentum) * self.mean
            self.moving_var = self.momentum * self.moving_var + (1.0 - self.momentum) * self.var
        else:
            self.mean = self.moving_mean
            self.var = self.moving_var


        self.X_centered = (self.input_data - self.mean)
        self.stddev_inv = 1 / np.sqrt(self.var + self.epsilon)

        X_hat = self.X_centered * self.stddev_inv

        self.output_data = self.gamma * X_hat + self.beta

        return self.output_data


    def backward(self, error):

        X_hat = self.X_centered * self.stddev_inv

        #first variant
        output_error = (1 / self.batch_size) * self.gamma * self.stddev_inv * (
            self.batch_size * error
            - np.sum(error, axis = 0)
            - self.X_centered * np.power(self.stddev_inv, 2) * np.sum(error * self.X_centered, axis = 0)
            )

        #second variant
        # dX_hat = error * self.gamma
        # output_error = (1 / self.batch_size) * self.stddev_inv * (
        #     self.batch_size * dX_hat
        #     - np.sum(dX_hat, axis = 0)
        #     - X_hat * np.sum(dX_hat * X_hat, axis = 0)
        # )

        #third (naive slow) variant
        # dX_hat = error * self.gamma
        # dvar = np.sum(dX_hat * self.X_centered, axis=0) * -.5 * self.stddev_inv**3
        # dmu = np.sum(dX_hat * -self.stddev_inv, axis=0) + dvar * np.mean(-2. * self.X_centered, axis=0)

        # output_error = (dX_hat * self.stddev_inv) + (dvar * 2 * self.X_centered / self.batch_size) + (dmu / self.batch_size)

        self.grad_gamma = np.sum(error * X_hat, axis = 0)
        self.grad_beta = np.sum(error, axis = 0)


        return output_error

    def update_weights(self, layer_num):
        self.gamma, self.vg, self.mg, self.vg_hat, self.mg_hat  = self.optimizer.update(self.grad_gamma, self.gamma, self.vg, self.mg, self.vg_hat, self.mg_hat, layer_num)
        self.beta, self.vb, self.mb, self.vb_hat, self.mb_hat = self.optimizer.update(self.grad_beta, self.beta, self.vb, self.mb, self.vb_hat, self.mb_hat, layer_num)

    def get_grads(self):
        return self.grad_gamma, self.grad_beta

    def set_grads(self, grads):
        self.grad_gamma, self.grad_beta = grads