Hyper-Modulation/self_alignment.py at main · hecoding/Hyper-Modulation · 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
import argparse
import random
import torch
from pathlib import Path
from tqdm import trange
from torch import nn, optim
from torchvision.utils import make_grid
from torch.utils.tensorboard import SummaryWriter

from model.stylegan import StyledGenerator as OriginalStyledGenerator
from utils import save_hparams, get_run_id


def get_z(batch_size, code_size, mixing, device, fixed_seed=False):
    rng = None
    if fixed_seed:
        rng = torch.Generator(device=device)
        rng.manual_seed(2147483647)
    if mixing and random.random() < 0.9:
        return [torch.randn(batch_size, code_size, generator=rng, device=device),
                torch.randn(batch_size, code_size, generator=rng, device=device)]
    else:
        return torch.randn(batch_size, code_size, generator=rng, device=device)


def get_noise(step, batch_size, device, fixed_seed=False):
    rng = None
    if fixed_seed:
        rng = torch.Generator(device=device)
        rng.manual_seed(2147483647)
    noise = []
    for i in range(step + 1):
        size = 4 * 2 ** i
        noise.append(torch.randn(batch_size, 1, size, size, generator=rng, device=device))

    return noise


def log_images(writer, i, generator, task, generator_original, code_size, step, device, batch_size=9):
    with torch.no_grad():
        z = get_z(batch_size=batch_size, code_size=code_size, mixing=False, device=device, fixed_seed=True)
        noise = get_noise(step, batch_size, device, fixed_seed=True)

        images_orig = generator_original(z, noise=noise, step=step, alpha=1, return_hierarchical=False)
        writer.add_image('sample_orig', make_grid(images_orig, nrow=3, normalize=True, range=(-1, 1)), i)

        task_in = task(torch.zeros(batch_size, device=device, dtype=torch.long))
        images = generator(z, noise=noise, step=step, alpha=1, task=task_in, return_hierarchical=False)
        writer.add_image('sample_new', make_grid(images, nrow=3, normalize=True, range=(-1, 1)), i)


def train(generator, task, generator_original, g_optimizer, criterion, iterations, args,
          batch_size=1, code_size=512, step=6, mixing=False, name=None, device='cpu', checkpoint_interval=5_000):
    name = 'unnamed' if name is None else name
    run_dir = f'{args.run_dir}/{get_run_id(args.run_dir):05d}-{name}'
    Path(run_dir).mkdir(parents=True)
    save_hparams(run_dir, args)
    writer = SummaryWriter(run_dir)
    alpha = 1
    for i in trange(iterations):
        g_optimizer.zero_grad()
        z = get_z(batch_size, code_size, mixing, device)
        noise = get_noise(step, batch_size, device)
        with torch.no_grad():
            _, g_l_orig = generator_original(z, noise=noise, step=step, alpha=alpha, return_hierarchical=True)
        task_in = task(torch.zeros(batch_size, device=device, dtype=torch.long))
        _, g_l = generator(z, noise=noise, step=step, alpha=alpha, task=task_in, return_hierarchical=True)

        assert len(g_l) == len(g_l_orig)
        g_loss = 0
        for input, target in zip(g_l, g_l_orig):
            g_loss += criterion(input, target)

        g_loss.backward()
        g_optimizer.step()

        writer.add_scalar('loss/g', g_loss, i)
        if i % args.log_interval == 0:
            log_images(writer, i, generator, task, generator_original, code_size, step, device)

        if i % checkpoint_interval == 0:
            torch.save({
                'generator': generator.state_dict(),
                'task': task.state_dict(),
            }, f'{run_dir}/{i:06d}.model')


def get_args():
    parser = argparse.ArgumentParser(description='Self-alignment')

    parser.add_argument('checkpoint', type=str)
    parser.add_argument('--step', default=6, type=int, help='6 = 256px')
    parser.add_argument('--iterations', type=int, default=50_000, help='number of samples used')
    parser.add_argument('--lr', default=0.002, type=float, help='learning rate')
    parser.add_argument('--code_size', default=512, type=int)
    parser.add_argument('--n_mlp_style', default=8, type=int)
    parser.add_argument('--n_mlp_task', default=4, type=int)
    parser.add_argument('--task_size', default=64, type=int)
    parser.add_argument('--batch_size', default=32, type=int, help='max image size')
    parser.add_argument('--mixing', action='store_true', help='use mixing regularization')
    parser.add_argument('--device', default='cuda', type=str)
    parser.add_argument('--name', type=str, default=None)
    parser.add_argument('--run_dir', type=str, default='data/training-runs/self_align')
    parser.add_argument('--checkpoint_interval', type=int, default=5_000, help='number of samples used')
    parser.add_argument('--log_interval', type=int, default=500)

    return parser.parse_args()


def main(args):
    from model.hyper_mod import StyledGenerator, Task
    from pretrained_converter import convert_generator, assert_loaded_keys, freeze_layers

    args.origin = 'self_align'
    g_running_orig = OriginalStyledGenerator(args.code_size).to(args.device)
    g_running_orig.train(False)
    generator = StyledGenerator(code_dim=args.code_size, task_dim=args.task_size, n_mlp=args.n_mlp_style).to(args.device)
    freeze_layers(generator)
    task = Task(args.task_size, n_mlp=args.n_mlp_task, num_labels=1).to(args.device)

    g_optimizer = optim.Adam(generator.generator.parameters(), lr=args.lr, betas=(0.0, 0.99))
    g_optimizer.add_param_group(
        {
            'params': generator.style.parameters(),
            'lr': args.lr * 0.01,
            'mult': 0.01,
        }
    )
    g_optimizer.add_param_group(
        {
            'params': task.parameters(),
            'lr': args.lr * 0.01,
            'mult': 0.001,
        }
    )

    ckpt = torch.load(args.checkpoint)
    g_running_orig.load_state_dict(ckpt['g_running'])
    missing, unexpected = generator.load_state_dict(convert_generator(ckpt['g_running']), strict=False)
    assert_loaded_keys(missing, unexpected)
    del ckpt

    criterion = nn.L1Loss()

    train(generator=generator, task=task, generator_original=g_running_orig, g_optimizer=g_optimizer,
          criterion=criterion, iterations=args.iterations, args=args, batch_size=args.batch_size,
          code_size=args.code_size, step=args.step, mixing=args.mixing, name=args.name, device=args.device,
          checkpoint_interval=args.checkpoint_interval)


if __name__ == '__main__':
    main(get_args())