Spaces:
Running
Running
Upload image_wgan.py
Browse files- image_wgan.py +119 -0
image_wgan.py
ADDED
@@ -0,0 +1,119 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from os import mkdir
|
2 |
+
from os.path import exists
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
|
6 |
+
import torch
|
7 |
+
from torch.autograd import Variable
|
8 |
+
from torch.utils.data import DataLoader
|
9 |
+
from torchvision.utils import save_image
|
10 |
+
|
11 |
+
from ml.pytorch.image_dataset import ImageDataset
|
12 |
+
from ml.pytorch.wgan.discriminator import Discriminator
|
13 |
+
from ml.pytorch.wgan.generator import Generator
|
14 |
+
|
15 |
+
|
16 |
+
class ImageWgan:
|
17 |
+
def __init__(
|
18 |
+
self,
|
19 |
+
image_shape: (int, int, int),
|
20 |
+
latent_space_dimension: int = 100,
|
21 |
+
use_cuda: bool = False,
|
22 |
+
generator_saved_model: str or None = None,
|
23 |
+
discriminator_saved_model: str or None = None
|
24 |
+
):
|
25 |
+
self.generator = Generator(image_shape, latent_space_dimension, use_cuda, generator_saved_model)
|
26 |
+
self.discriminator = Discriminator(image_shape, use_cuda, discriminator_saved_model)
|
27 |
+
|
28 |
+
self.image_shape = image_shape
|
29 |
+
self.latent_space_dimension = latent_space_dimension
|
30 |
+
self.use_cuda = use_cuda
|
31 |
+
if use_cuda:
|
32 |
+
self.generator.cuda()
|
33 |
+
self.discriminator.cuda()
|
34 |
+
|
35 |
+
def train(
|
36 |
+
self,
|
37 |
+
image_dataset: ImageDataset,
|
38 |
+
learning_rate: float = 0.00005,
|
39 |
+
batch_size: int = 64,
|
40 |
+
workers: int = 8,
|
41 |
+
epochs: int = 100,
|
42 |
+
clip_value: float = 0.01,
|
43 |
+
discriminator_steps: int = 5,
|
44 |
+
sample_interval: int = 1000,
|
45 |
+
sample_folder: str = 'samples',
|
46 |
+
generator_save_file: str = 'generator.model',
|
47 |
+
discriminator_save_file: str = 'discriminator.model'
|
48 |
+
):
|
49 |
+
if not exists(sample_folder):
|
50 |
+
mkdir(sample_folder)
|
51 |
+
|
52 |
+
generator_optimizer = torch.optim.RMSprop(self.generator.parameters(), lr=learning_rate)
|
53 |
+
discriminator_optimizer = torch.optim.RMSprop(self.discriminator.parameters(), lr=learning_rate)
|
54 |
+
|
55 |
+
Tensor = torch.cuda.FloatTensor if self.use_cuda else torch.FloatTensor
|
56 |
+
|
57 |
+
data_loader = torch.utils.data.DataLoader(
|
58 |
+
image_dataset,
|
59 |
+
batch_size=batch_size,
|
60 |
+
shuffle=True,
|
61 |
+
num_workers=workers
|
62 |
+
)
|
63 |
+
batches_done = 0
|
64 |
+
for epoch in range(epochs):
|
65 |
+
for i, imgs in enumerate(data_loader):
|
66 |
+
real_imgs = Variable(imgs.type(Tensor))
|
67 |
+
|
68 |
+
discriminator_optimizer.zero_grad()
|
69 |
+
|
70 |
+
# Sample noise as generator input
|
71 |
+
z = Variable(Tensor(np.random.normal(0, 1, (imgs.shape[0], self.latent_space_dimension))))
|
72 |
+
|
73 |
+
fake_imgs = self.generator(z).detach()
|
74 |
+
# Adversarial loss
|
75 |
+
discriminator_loss = -torch.mean(self.discriminator(real_imgs)) + torch.mean(self.discriminator(fake_imgs))
|
76 |
+
|
77 |
+
discriminator_loss.backward()
|
78 |
+
discriminator_optimizer.step()
|
79 |
+
|
80 |
+
# Clip weights of discriminator
|
81 |
+
for p in self.discriminator.parameters():
|
82 |
+
p.data.clamp_(-clip_value, clip_value)
|
83 |
+
|
84 |
+
# Train the generator every n_critic iterations
|
85 |
+
if i % discriminator_steps == 0:
|
86 |
+
generator_optimizer.zero_grad()
|
87 |
+
|
88 |
+
# Generate a batch of images
|
89 |
+
gen_imgs = self.generator(z)
|
90 |
+
# Adversarial loss
|
91 |
+
generator_loss = -torch.mean(self.discriminator(gen_imgs))
|
92 |
+
|
93 |
+
generator_loss.backward()
|
94 |
+
generator_optimizer.step()
|
95 |
+
|
96 |
+
print(
|
97 |
+
f'[Epoch {epoch}/{epochs}] [Batch {batches_done % len(data_loader)}/{len(data_loader)}] ' +
|
98 |
+
f'[D loss: {discriminator_loss.item()}] [G loss: {generator_loss.item()}]'
|
99 |
+
)
|
100 |
+
|
101 |
+
if batches_done % sample_interval == 0:
|
102 |
+
save_image(gen_imgs.data[:25], f'{sample_folder}/{batches_done}.png', nrow=5, normalize=True)
|
103 |
+
batches_done += 1
|
104 |
+
self.discriminator.save(discriminator_save_file)
|
105 |
+
self.generator.save(generator_save_file)
|
106 |
+
|
107 |
+
def generate(
|
108 |
+
self,
|
109 |
+
sample_folder: str = 'samples'
|
110 |
+
):
|
111 |
+
if not exists(sample_folder):
|
112 |
+
mkdir(sample_folder)
|
113 |
+
|
114 |
+
Tensor = torch.cuda.FloatTensor if self.use_cuda else torch.FloatTensor
|
115 |
+
z = Variable(Tensor(np.random.normal(0, 1, (self.image_shape[0], self.latent_space_dimension))))
|
116 |
+
gen_imgs = self.generator(z)
|
117 |
+
generator_loss = -torch.mean(self.discriminator(gen_imgs))
|
118 |
+
generator_loss.backward()
|
119 |
+
save_image(gen_imgs.data[:25], f'{sample_folder}/generated.png', nrow=5, normalize=True)
|