File size: 5,981 Bytes
4dd2c81 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 |
import numpy as np
import tensorflow as tf
import tensorflow_addons as tfa
import keras
from keras import Model, Sequential, initializers
from keras.layers import Layer, Conv2D, LeakyReLU, Dropout
class SPADE(Layer):
def __init__(self, filters: int, epsilon=1e-5, **kwargs):
super().__init__(**kwargs)
self.epsilon = epsilon
self.conv = Conv2D(128, 3, padding="same", activation="relu")
self.conv_gamma = Conv2D(filters, 3, padding="same")
self.conv_beta = Conv2D(filters, 3, padding="same")
def build(self, input_shape):
self.resize_shape = input_shape[1:3]
def call(self, input_tensor, raw_mask):
mask = tf.image.resize(raw_mask, self.resize_shape, method="nearest")
x = self.conv(mask)
gamma = self.conv_gamma(x)
beta = self.conv_beta(x)
mean, var = tf.nn.moments(input_tensor, axes=(0, 1, 2), keepdims=True)
std = tf.sqrt(var + self.epsilon)
normalized = (input_tensor - mean) / std
output = gamma * normalized + beta
return output
def get_config(self):
return {
"epsilon": self.epsilon,
"conv": self.conv,
"conv_gamma": self.conv_gamma,
"conv_beta": self.conv_beta
}
class ResBlock(Layer):
def __init__(self, filters: int, **kwargs):
super().__init__(**kwargs)
self.filters = filters
def build(self, input_shape):
input_filter = input_shape[-1]
self.spade_1 = SPADE(input_filter)
self.spade_2 = SPADE(self.filters)
self.conv_1 = Conv2D(self.filters, 3, padding="same")
self.conv_2 = Conv2D(self.filters, 3, padding="same")
self.leaky_relu = LeakyReLU(0.2)
self.learned_skip = False
if self.filters != input_filter:
self.learned_skip = True
self.spade_3 = SPADE(input_filter)
self.conv_3 = Conv2D(self.filters, 3, padding="same")
def call(self, input_tensor, mask):
x = self.spade_1(input_tensor, mask)
x = self.conv_1(self.leaky_relu(x))
x = self.spade_2(x, mask)
x = self.conv_2(self.leaky_relu(x))
skip = (
self.conv_3(self.leaky_relu(self.spade_3(input_tensor, mask)))
if self.learned_skip
else input_tensor
)
output = skip + x
return output
def get_config(self):
return {"filters": self.filters}
class Downsample(Layer):
def __init__(self,
channels: int,
kernels: int,
strides: int = 2,
apply_norm=True,
apply_activation=True,
apply_dropout=False,
**kwargs
):
super().__init__(**kwargs)
self.channels = channels
self.kernels = kernels
self.strides = strides
self.apply_norm = apply_norm
self.apply_activation = apply_activation
self.apply_dropout = apply_dropout
def build(self, input_shape):
self.block = Sequential([
Conv2D(
self.channels,
self.kernels,
strides=self.strides,
padding="same",
use_bias=False,
kernel_initializer=initializers.GlorotNormal(),
)])
if self.apply_norm:
self.block.add(tfa.layers.InstanceNormalization())
if self.apply_activation:
self.block.add(LeakyReLU(0.2))
if self.apply_dropout:
self.block.add(Dropout(0.5))
def call(self, inputs):
return self.block(inputs)
def get_config(self):
return {
"channels": self.channels,
"kernels": self.kernels,
"strides": self.strides,
"apply_norm": self.apply_norm,
"apply_activation": self.apply_activation,
"apply_dropout": self.apply_dropout,
}
class GaussianSampler(Layer):
def __init__(self, latent_dim: int, **kwargs):
super().__init__(**kwargs)
self.latent_dim = latent_dim
def call(self, inputs):
means, variance = inputs
epsilon = tf.random.normal(
shape=(tf.shape(means)[0], self.latent_dim), mean=0.0, stddev=1.0
)
samples = means + tf.exp(0.5 * variance) * epsilon
return samples
def get_config(self):
return {"latent_dim": self.latent_dim}
class GauganPredictor():
CLASSES = (
'unknown','wall', 'sky', 'tree', 'road', 'grass', 'earth',
'mountain', 'plant', 'water', 'sea', 'field', 'fence', 'rock',
'sand', 'path', 'river', 'flower', 'hill', 'palm', 'tower',
'dirt', 'land', 'waterfall', 'lake'
)
def __init__(self, model_g_path: str, model_e_path: str = None) -> None:
custom_objects = {
'ResBlock': ResBlock,
'Downsample': Downsample,
}
if model_e_path is not None:
self.encoder: Model = keras.models.load_model(model_e_path, custom_objects=custom_objects)
self.sampler = GaussianSampler(256)
self.gen: Model = keras.models.load_model(
model_g_path, custom_objects=custom_objects)
def __call__(self, im: np.ndarray, z=None) -> np.ndarray:
if len(im.shape) == 3:
im = im[np.newaxis]
if z is None:
z = tf.random.normal((im.shape[0], 256))
tmp = self.gen.predict_on_batch([z, im])
x = np.array((tmp + 1) * 127.5, np.uint8)
return x
def predict_reference(self, im: np.ndarray, reference_im: np.ndarray) -> np.ndarray:
if len(im.shape) == 3:
im = im[np.newaxis]
reference_im = reference_im[np.newaxis]
mean, variance = self.encoder(reference_im)
z = self.sampler([mean, variance])
x = np.array((self.gen.predict_on_batch([z, im]) + 1) * 127.5, np.uint8)
return x
|