style-vae/style_vae/train/test.py at master · orgoro/style-vae · 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
138
139
from __future__ import print_function
from __future__ import absolute_import
from __future__ import division

# 3rd party:
import fire
import tensorflow as tf
import numpy as np
from matplotlib import pyplot as plt
from os import path
import imageio
from tqdm import tqdm

# different category:
from style_vae.model import StyleVae, Config
from style_vae.train_output import OUT

# same category:
from style_vae.train.style_vae_trainer import StyleVaeTrainer
from style_vae.train.vae_trainer_config import VaeTrainerConfig


def plot_manifold():
    seed = np.random.randint(0, 100)
    print(f'seed - {seed}')
    np.random.seed(seed)

    gs = 8
    trainer, sess = _build_trainer(grid_size=gs)
    model = trainer.get_model()
    code_size = model.config.code_size
    img_dim = model.config.img_dim

    manifold_0 = _generate_manifold_random(code_size, gs)
    manifold_1 = _generate_manifold_random(code_size, gs)
    manifold_shape = manifold_0.shape

    img_names = []
    num_frames = 31
    for t in tqdm(range(num_frames)):
        s1 = (num_frames-1-t)
        s2 = t
        s = np.sqrt(s1**2 + s2**2)
        manifold = (manifold_0 * s1 + manifold_1 * s2) / s
        manifold_img = np.empty((gs, gs, img_dim, img_dim, 3))
        code_ph, decoded = trainer.get_decode_stubs()
        for i in range(gs):
            code = manifold[i, :, :]
            manifold_img[i] = decoded.eval(session=sess, feed_dict={code_ph: code})
        img = np.empty((gs * img_dim, gs * img_dim, 3))
        for i in range(manifold_shape[0]):
            for j in range(manifold_shape[1]):
                offset_x = i * img_dim
                offset_y = j * img_dim
                img[offset_x: offset_x + img_dim, offset_y: offset_y + img_dim] = manifold_img[i, j]

        fname = path.join(OUT, f'grid-{seed}-{t}.jpg')
        plt.imsave(fname, np.clip(img, 0, 1))
        img_names.append(fname)

    images = []
    for fname in img_names:
        images.append(imageio.imread(fname))
    images.extend(images[::-1])
    imageio.mimsave(path.join(OUT, f'interpolation-{seed}.gif'), images)


def plot_recon():
    gs = 8
    trainer, sess = _build_trainer(grid_size=gs)
    model = trainer.get_model()
    img_dim = model.config.img_dim
    manifold_shape = (gs, gs, img_dim, img_dim, 3)
    manifold_img = np.empty(manifold_shape)
    img_tensor, recon_tensor = trainer.get_encode_stubs()
    for i in range(gs//2):
        img, recon = sess.run([img_tensor, recon_tensor])
        manifold_img[i*2] = img
        manifold_img[i*2+1] = recon

    img = np.empty((gs * img_dim, gs * img_dim, 3))
    for i in range(manifold_shape[0]):
        for j in range(manifold_shape[1]):
            offset_x = i * img_dim
            offset_y = j * img_dim
            img[offset_x: offset_x + img_dim, offset_y: offset_y + img_dim] = manifold_img[i, j]

    fname = path.join(OUT, f'recon.jpg')
    plt.imsave(fname, np.clip(img, 0, 1))


def _generate_manifold(code_size: int, grid_size: int) -> np.ndarray:
    vecs = np.random.randn(2, code_size) + 1
    std = 0.5
    manifold = np.zeros((grid_size, grid_size, code_size))
    s = np.linspace(-std, std, grid_size, endpoint=True)
    xx, yy = np.meshgrid(s, s)
    for i in range(grid_size):
        for j in range(grid_size):
            manifold[i, j] = xx[i, j] * vecs[0] + yy[i, j] * vecs[1]

    return manifold


def _generate_manifold_random(code_size: int, grid_size: int) -> np.ndarray:
    manifold = np.zeros((grid_size, grid_size, code_size))
    for i in range(grid_size):
        for j in range(grid_size):
            manifold[i, j] = np.random.randn(1, code_size)
    return manifold


def _build_trainer(reload_model=True, grid_size=8) -> (StyleVaeTrainer, tf.Session):
    # model
    model_config = Config()
    model_config.batch_size = grid_size

    print(model_config)
    model = StyleVae(model_config)

    # trainer
    trainer_config = VaeTrainerConfig()
    trainer_config.batch_size = grid_size
    trainer_config.reload_model = reload_model
    print(trainer_config)

    sess = tf.Session()
    trainer = StyleVaeTrainer(model, trainer_config, sess)
    if reload_model:
        trainer.load()
        trainer.init_iter()
    else:
        init_op = tf.global_variables_initializer()
        sess.run(init_op)
    return trainer, sess


if __name__ == '__main__':
    fire.Fire()