ImageVAE/vae_callback.py at master · schaugf/ImageVAE · 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
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
180
import os
import imageio
import glob
from PIL import Image
import numpy as np
from scipy.stats import norm
from keras.callbacks import Callback

class VAEcallback(Callback):
    """ This callback saves sample input images, their reconstructions, and a
    latent space walk at the end of each epoch for the ImageVAE model.
    This callback accepts an ImageVAE model as its input, including both
    encoder and decoder networks.

    #   Example
        '''python
            vaecb = VAEcallback(self)
        '''

    #   Arguments
        num_save: number of individual images (nxn) to save to grid
        vae: variational autoencoding model to generate reconstructions
        decoder: VAE decoder network to generate latent space walk visual

    """

    def __init__(self, model):

        self.latent_dim     = model.latent_dim
        self.latent_samp    = model.latent_samp
        self.batch_size     = model.batch_size
        self.image_size     = model.image_size
        self.num_save       = model.num_save
        self.nchannel       = model.nchannel
        self.image_res      = model.image_res
        self.data_dir       = model.data_dir
        self.save_dir       = model.save_dir
        self.vae            = model.vae
        self.decoder        = model.decoder
        self.show_channels  = model.show_channels
        self.do_vaecb_each  = model.do_vaecb_each  # do each epoch

    def save_input_images(self):
        """ save input images
        """
        input_figure = np.zeros((self.image_size * self.num_save,
                         self.image_size * self.num_save,
                         min(3, self.nchannel)))

        to_load = glob.glob(os.path.join(self.data_dir, 'train', '*'))[:(self.num_save * self.num_save)]


        if self.nchannel > 3:
            input_images = np.array([np.load(fname) for fname in to_load])
            input_images = ((2**8 - 1) * (input_images[...,self.show_channels]/(2**self.image_res - 1))).astype(np.uint8)

        else:
            input_images = np.array([np.array(Image.open(fname)) for fname in to_load])  # needs to generalize to 2-channel?

        if self.nchannel == 1:
            input_images = input_images[..., None]  # add extra index dimension

        idx = 0
        for i in range(self.num_save):
            for j in range(self.num_save):
                input_figure[i * self.image_size : (i+1) * self.image_size,
                             j * self.image_size : (j+1) * self.image_size, :] = input_images[idx,:,:,:]
                idx += 1

        imageio.imwrite(os.path.join(self.save_dir, 'input_images.png'),
                        input_figure.astype(np.uint8))


    def save_input_reconstruction(self, epoch=0, is_final=False):
        """ save grid of both input and reconstructed images side by side
        """

        recon_figure = np.zeros((self.image_size * self.num_save,
                                 self.image_size * self.num_save,
                                 min(3, self.nchannel)))

        to_load = glob.glob(os.path.join(self.data_dir, 'train', '*'))[:(self.num_save * self.num_save)]

        # again, whether input is npy or png
        if self.nchannel > 3:
            input_images = np.array([np.load(fname) for fname in to_load])

        else:
            input_images = np.array([np.array(Image.open(fname)) for fname in to_load])

        scaled_input = input_images / (2**self.image_res - 1)

        if self.nchannel == 1:
            scaled_input = scaled_input[..., None]


        recon_images = self.vae.predict(scaled_input, batch_size = self.batch_size)
        scaled_recon = recon_images * (2**8 - 1)
        scaled_recon = scaled_recon[...,self.show_channels]

        idx = 0
        for i in range(self.num_save):
            for j in range(self.num_save):
                #print('reconstruction max:', scaled_recon.max())
                #print('reconstructed dtype:', scaled_recon.dtype)

                recon_figure[i * self.image_size : (i+1) * self.image_size,
                             j * self.image_size : (j+1) * self.image_size, :] = scaled_recon[idx,...]
                idx += 1

        if not(is_final):
            imageio.imwrite(os.path.join(self.save_dir,
                                         'reconstructed',
                                         'recon_images_epoch_{0:03d}.png'.format(epoch)),
                                         recon_figure.astype(np.uint8))
        else:
            imageio.imwrite(os.path.join(self.save_dir,
                                         'recon_images_final.png'),
                                         recon_figure.astype(np.uint8))


    def latent_walk(self, epoch=0, is_final=False):
        """ latent space walking
        """

        figure = np.zeros((self.image_size * self.latent_dim, self.image_size * self.latent_samp, min(3, self.nchannel)))
        grid_x = norm.ppf(np.linspace(0.05, 0.95, self.latent_samp))

        for i in range(self.latent_dim):
            for j, xi in enumerate(grid_x):
                z_sample = np.zeros(self.latent_dim)
                z_sample[i] = xi

                z_sample = np.tile(z_sample, self.batch_size).reshape(self.batch_size, self.latent_dim)

                x_decoded = self.decoder.predict(z_sample, batch_size=self.batch_size)
                x_decoded = x_decoded * float(2**8 - 1)

                sample = x_decoded[0].reshape(self.image_size, self.image_size, self.nchannel)

                # need show_channels input as list
                if self.nchannel > 3:
                    sample = sample[:,:,self.show_channels]

                figure[i * self.image_size: (i + 1) * self.image_size,
                       j * self.image_size: (j + 1) * self.image_size, :] = sample

        if not(is_final):
            imageio.imwrite(os.path.join(self.save_dir, 'latent_walk', 'latent_walk_epoch_{0:03d}.png'.format(epoch)),
                            figure.astype(np.uint8))
        else:
            imageio.imwrite(os.path.join(self.save_dir, 'latent_walk_final.png'),
                            figure.astype(np.uint8))


    def on_epoch_end(self, epoch, logs={}):
        if self.do_vaecb_each:
            self.save_input_reconstruction(epoch)
            self.latent_walk(epoch)


    def on_train_begin(self, logs={}):
        os.makedirs(os.path.join(self.save_dir, 'latent_walk'), exist_ok=True)
        os.makedirs(os.path.join(self.save_dir, 'reconstructed'), exist_ok=True)
        os.makedirs(os.path.join(self.save_dir, 'animated'), exist_ok=True)
        self.save_input_images()


    def on_train_end(self, logs={}):
        self.save_input_reconstruction(is_final=True)
        self.latent_walk(is_final=True)

        if self.do_vaecb_each:
            print('animating training...')
            cmd = 'ffmpeg -i ' + self.save_dir + '/latent_walk/latent_walk_epoch_%03d.png -vcodec libx264 -crf 25 ' + self.save_dir + '/animated/latent_walk_animated.mp4'
            os.system(cmd)

            cmd = 'ffmpeg -i ' + self.save_dir + '/reconstructed/recon_images_epoch_%03d.png -vcodec libx264 -crf 25 ' + self.save_dir + '/animated/reconstructed_animated.mp4'
            os.system(cmd)