from tensorflow import keras
from tensorflow.keras import layers
import keras.backend as K
import json
import pickle
import matplotlib.pyplot as plt
import tensorflow_probability as tfp
import Datasets
IMG_SHAPE = (32, 32, 3)
BATCH_SIZE = 64
# Size of the noise vector
noise_dim = 100
dataset = Datasets.CelebA(zip_file_path='files/32x32.pickle')
train_images = dataset.X
def conv_block(
x,
filters,
activation,
kernel_size=(3, 3),
strides=(1, 1),
padding="same",
use_bias=True,
use_bn=False,
use_dropout=False,
drop_value=0.5,
):
self.samples = samples
self.batch_size = batch_size
self.name = 'FID'
def calculate(self, model, dataset):
noise = np.random.normal(0, 1, (self.samples, model.latent_dim))
images = model.generator.predict(noise)
gan_fid = self.fd(images, batch_size=self.batch_size)
return float(gan_fid)
if __name__ == '__main__':
import sys
sys.path.append('../')
sys.path.append('./')
import Datasets
import pickle
dataset = Datasets.CelebA(size=64)
fd = FrechetInceptionDistance(None, None, (-1, 1), 75)
mean, cov = fd.stats(dataset.X, batch_size=32)
print(mean, cov)
save = {'mean': mean, 'cov': cov}
with open('celeba_64_stats.pickle', 'wb') as handle:
pickle.dump(save, handle)