File size: 5,584 Bytes
d6c7221
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
import tensorflow as tf
import tensorflow.keras.backend as K
from tensorflow.keras.layers import Conv2D, Conv2DTranspose, Dense
from tensorflow.python.eager import context
from tensorflow.python.framework import tensor_shape
from tensorflow.python.keras import activations, constraints, initializers, regularizers
from tensorflow.python.keras.layers.convolutional import SeparableConv
from tensorflow.python.ops import (
    array_ops,
    gen_math_ops,
    math_ops,
    sparse_ops,
    standard_ops,
)


def l2normalize(v, eps=1e-12):
    return v / (tf.norm(v) + eps)


class ConvSN2D(tf.keras.layers.Conv2D):
    def __init__(self, filters, kernel_size, power_iterations=1, **kwargs):
        super(ConvSN2D, self).__init__(filters, kernel_size, **kwargs)
        self.power_iterations = power_iterations

    def build(self, input_shape):
        super(ConvSN2D, self).build(input_shape)

        if self.data_format == "channels_first":
            channel_axis = 1
        else:
            channel_axis = -1

        self.u = self.add_weight(
            self.name + "_u",
            shape=tuple([1, self.kernel.shape.as_list()[-1]]),
            initializer=tf.initializers.RandomNormal(0, 1),
            trainable=False,
            dtype=self.dtype,
        )

    def compute_spectral_norm(self, W, new_u, W_shape):
        for _ in range(self.power_iterations):

            new_v = l2normalize(tf.matmul(new_u, tf.transpose(W)))
            new_u = l2normalize(tf.matmul(new_v, W))

        sigma = tf.matmul(tf.matmul(new_v, W), tf.transpose(new_u))
        W_bar = W / sigma

        with tf.control_dependencies([self.u.assign(new_u)]):
            W_bar = tf.reshape(W_bar, W_shape)

        return W_bar

    def call(self, inputs):
        W_shape = self.kernel.shape.as_list()
        W_reshaped = tf.reshape(self.kernel, (-1, W_shape[-1]))
        new_kernel = self.compute_spectral_norm(W_reshaped, self.u, W_shape)
        outputs = self._convolution_op(inputs, new_kernel)

        if self.use_bias:
            if self.data_format == "channels_first":
                outputs = tf.nn.bias_add(outputs, self.bias, data_format="NCHW")
            else:
                outputs = tf.nn.bias_add(outputs, self.bias, data_format="NHWC")
        if self.activation is not None:
            return self.activation(outputs)

        return outputs


class DenseSN(Dense):
    def build(self, input_shape):
        super(DenseSN, self).build(input_shape)

        self.u = self.add_weight(
            self.name + "_u",
            shape=tuple([1, self.kernel.shape.as_list()[-1]]),
            initializer=tf.initializers.RandomNormal(0, 1),
            trainable=False,
            dtype=self.dtype,
        )

    def compute_spectral_norm(self, W, new_u, W_shape):
        new_v = l2normalize(tf.matmul(new_u, tf.transpose(W)))
        new_u = l2normalize(tf.matmul(new_v, W))
        sigma = tf.matmul(tf.matmul(new_v, W), tf.transpose(new_u))
        W_bar = W / sigma
        with tf.control_dependencies([self.u.assign(new_u)]):
            W_bar = tf.reshape(W_bar, W_shape)
        return W_bar

    def call(self, inputs):
        W_shape = self.kernel.shape.as_list()
        W_reshaped = tf.reshape(self.kernel, (-1, W_shape[-1]))
        new_kernel = self.compute_spectral_norm(W_reshaped, self.u, W_shape)
        rank = len(inputs.shape)
        if rank > 2:
            outputs = standard_ops.tensordot(inputs, new_kernel, [[rank - 1], [0]])
            if not context.executing_eagerly():
                shape = inputs.shape.as_list()
                output_shape = shape[:-1] + [self.units]
                outputs.set_shape(output_shape)
        else:
            inputs = math_ops.cast(inputs, self._compute_dtype)
            if K.is_sparse(inputs):
                outputs = sparse_ops.sparse_tensor_dense_matmul(inputs, new_kernel)
            else:
                outputs = gen_math_ops.mat_mul(inputs, new_kernel)
        if self.use_bias:
            outputs = tf.nn.bias_add(outputs, self.bias)
        if self.activation is not None:
            return self.activation(outputs)
        return outputs


class AddNoise(tf.keras.layers.Layer):
    def build(self, input_shape):
        self.b = self.add_weight(
            shape=[
                1,
            ],
            initializer=tf.keras.initializers.zeros(),
            trainable=True,
            name="noise_weight",
        )

    def call(self, inputs):
        rand = tf.random.normal(
            [tf.shape(inputs)[0], inputs.shape[1], inputs.shape[2], 1],
            mean=0.0,
            stddev=1.0,
            dtype=self.dtype,
        )
        output = inputs + self.b * rand
        return output


class PosEnc(tf.keras.layers.Layer):
    def __init__(self, **kwargs):
        super(PosEnc, self).__init__(**kwargs)

    def call(self, inputs):
        # inputs shape: [bs,mel_bins,shape,1]
        pos = tf.repeat(
            tf.reshape(tf.range(inputs.shape[-3], dtype=tf.int32), [1, -1, 1, 1]),
            inputs.shape[-2],
            -2,
        )
        pos = tf.cast(tf.repeat(pos, tf.shape(inputs)[0], 0), self.dtype) / tf.cast(
            inputs.shape[-3], self.dtype
        )
        return tf.concat([inputs, pos], -1)  # [bs,1,hop,2]


def flatten_hw(x, data_format="channels_last"):
    if data_format == "channels_last":
        x = tf.transpose(x, perm=[0, 3, 1, 2])  # Convert to `channels_first`

    old_shape = tf.shape(x)
    new_shape = [old_shape[0], old_shape[2] * old_shape[3], old_shape[1]]

    return tf.reshape(x, new_shape)