Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_pggan.py 7.19 KB
Newer Older
limm's avatar
add tests code  
limm committed
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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195 
# Copyright (c) OpenMMLab. All rights reserved.
from copy import deepcopy

import numpy as np
import pytest
import torch
import torch.nn as nn

from mmgen.models.gans import ProgressiveGrowingGAN


class TestPGGAN:

    @classmethod
    def setup_class(cls):
        cls.generator_cfg = dict(
            type='PGGANGenerator',
            noise_size=8,
            out_scale=16,
            base_channels=32,
            max_channels=32)

        cls.discriminator_cfg = dict(
            type='PGGANDiscriminator', in_scale=16, label_size=0)

        cls.gan_loss = dict(type='GANLoss', gan_type='vanilla')
        cls.disc_auxiliary_loss = [
            dict(
                type='DiscShiftLoss',
                loss_weight=0.5,
                data_info=dict(pred='disc_pred_fake')),
            dict(
                type='DiscShiftLoss',
                loss_weight=0.5,
                data_info=dict(pred='disc_pred_real'))
        ]
        cls.train_cfg = dict(
            use_ema=True,
            nkimgs_per_scale={
                '4': 0.004,
                '8': 0.008,
                '16': 0.016
            },
            optimizer_cfg=dict(
                generator=dict(type='Adam', lr=0.0002, betas=(0.5, 0.999)),
                discriminator=dict(type='Adam', lr=0.0002,
                                   betas=(0.5, 0.999))),
            g_lr_base=0.0001,
            d_lr_base=0.0001,
            g_lr_schedule={'16': 0.00005})

    def test_pggan_cpu(self):
        # test default config
        pggan = ProgressiveGrowingGAN(
            self.generator_cfg,
            self.discriminator_cfg,
            self.gan_loss,
            disc_auxiliary_loss=self.disc_auxiliary_loss,
            train_cfg=self.train_cfg)

        data_batch = dict(real_img=torch.randn(3, 3, 16, 16))

        for iter_num in range(6):
            outputs = pggan.train_step(
                data_batch,
                None,
                running_status=dict(iteration=iter_num, batch_size=3))
            results = outputs['results']
            if iter_num == 1:
                assert results['fake_imgs'].shape == (3, 3, 4, 4)
            elif iter_num == 2:
                assert results['fake_imgs'].shape == (3, 3, 8, 8)
                assert np.isclose(pggan._actual_nkimgs[0], 0.006, atol=1e-8)
            elif iter_num == 3:
                assert results['fake_imgs'].shape == (3, 3, 8, 8)
                assert np.isclose(pggan._actual_nkimgs[0], 0.006, atol=1e-8)
                assert np.isclose(
                    pggan.optimizer['generator'].defaults['lr'],
                    0.0001,
                    atol=1e-8)
            elif iter_num == 5:
                assert results['fake_imgs'].shape == (3, 3, 16, 16)
                assert np.isclose(pggan._actual_nkimgs[-1], 0.012, atol=1e-8)
                assert np.isclose(
                    pggan.optimizer['generator'].defaults['lr'],
                    0.00005,
                    atol=1e-8)

        # test sample from noise
        outputs = pggan.sample_from_noise(None, num_batches=2)
        assert outputs.shape == (4, 3, 16, 16)

        outputs = pggan.sample_from_noise(
            None,
            num_batches=2,
            return_noise=True,
            transition_weight=0.2,
            sample_model='ema')
        assert outputs['fake_img'].shape == (2, 3, 16, 16)

        outputs = pggan.sample_from_noise(
            None, num_batches=2, return_noise=True, sample_model='orig')
        assert outputs['fake_img'].shape == (2, 3, 16, 16)

        with pytest.raises(RuntimeError):
            data_batch = dict(real_img=torch.randn(3, 3, 4, 32))
            _ = pggan.train_step(
                data_batch, None, running_status=dict(iteration=5))

        # test customized config
        train_cfg_ = deepcopy(self.train_cfg)
        train_cfg_['use_ema'] = False
        train_cfg_['interp_real_cfg'] = dict(
            mode='bilinear', align_corners=False)
        train_cfg_['interp_real_cfg'] = dict(
            mode='bilinear', align_corners=False)
        train_cfg_['reset_optim_for_new_scale'] = False
        pggan = ProgressiveGrowingGAN(
            self.generator_cfg,
            self.discriminator_cfg,
            self.gan_loss,
            disc_auxiliary_loss=None,
            train_cfg=train_cfg_)

        data_batch = dict(real_img=torch.randn(3, 3, 16, 16))

        outputs = pggan.train_step(
            data_batch, None, running_status=dict(iteration=0, batch_size=3))
        results = outputs['results']
        assert results['fake_imgs'].shape == (3, 3, 4, 4)
        assert not pggan.with_gen_auxiliary_loss
        assert not pggan.with_disc_auxiliary_loss
        assert not pggan.use_ema

        data_batch = dict(real_img=torch.randn(3, 3, 16, 16))

        for iter_num in range(1, 3):
            outputs = pggan.train_step(
                data_batch,
                None,
                running_status=dict(iteration=iter_num, batch_size=3))
            results = outputs['results']
            if iter_num == 1:
                assert results['fake_imgs'].shape == (3, 3, 4, 4)
            elif iter_num == 2:
                assert results['fake_imgs'].shape == (3, 3, 8, 8)
                assert np.isclose(pggan._actual_nkimgs[0], 0.006, atol=1e-8)

        train_cfg_ = deepcopy(self.train_cfg)
        train_cfg_['optimizer_cfg'] = dict(
            generator=dict(type='Adam', lr=0.0002, betas=(0.5, 0.999)))
        pggan = ProgressiveGrowingGAN(
            self.generator_cfg,
            None,
            None,
            disc_auxiliary_loss=dict(
                type='DiscShiftLoss',
                loss_weight=0.5,
                data_info=dict(pred='disc_pred_fake')),
            gen_auxiliary_loss=dict(type='GeneratorPathRegularizer'),
            train_cfg=train_cfg_)
        assert pggan.with_gen_auxiliary_loss
        assert isinstance(pggan.disc_auxiliary_losses, nn.ModuleList)
        assert pggan.gan_loss is None
        assert pggan.discriminator is None

    @pytest.mark.skipif(not torch.cuda.is_available(), reason='requires cuda')
    def test_pggan_cuda(self):
        # test default config
        pggan = ProgressiveGrowingGAN(
            self.generator_cfg,
            self.discriminator_cfg,
            self.gan_loss,
            disc_auxiliary_loss=self.disc_auxiliary_loss,
            train_cfg=self.train_cfg).cuda()

        data_batch = dict(real_img=torch.randn(3, 3, 32, 32).cuda())

        for iter_num in range(6):
            outputs = pggan.train_step(
                data_batch,
                None,
                running_status=dict(iteration=iter_num, batch_size=3))
            results = outputs['results']
            if iter_num == 1:
                assert results['fake_imgs'].shape == (3, 3, 4, 4)
            elif iter_num == 2:
                assert results['fake_imgs'].shape == (3, 3, 8, 8)
                assert np.isclose(pggan._actual_nkimgs[0], 0.006, atol=1e-8)
            elif iter_num == 3:
                assert results['fake_imgs'].shape == (3, 3, 8, 8)
                assert np.isclose(pggan._actual_nkimgs[0], 0.006, atol=1e-8)
            elif iter_num == 5:
                assert results['fake_imgs'].shape == (3, 3, 16, 16)
                assert np.isclose(pggan._actual_nkimgs[-1], 0.012, atol=1e-8)