Repository navigation
Expand file tree
/
Copy pathtrain.py
More file actions
187 lines (159 loc) · 6.64 KB
/
Copy pathtrain.py
File metadata and controls
187 lines (159 loc) · 6.64 KB
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
8000
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
import math
import torch
import torch.nn.functional as F
from matplotlib import gridspec
from matplotlib import pyplot as plt
from torch import nn
from torch.utils.data import sampler
from torchvision import datasets, transforms
import os
from models import *
import numpy as np
dtype = torch.FloatTensor
IMG_DIM=28
NOISE_DIM=50
plt.rcParams['figure.figsize'] = (10.0, 8.0) # set default size of plots
plt.rcParams['image.interpolation'] = 'nearest'
plt.rcParams['image.cmap'] = 'gray'
def g_bceloss(scores_fake):
return nn.BCEWithLogitsLoss()(scores_fake, torch.ones(scores_fake.shape[0]).float().unsqueeze(-1))
def d_bceloss(D, realx, fakex, y):
N = realx.shape[0]
real_scores = D(realx, y)
fake_scores = D(fakex, y)
return nn.BCEWithLogitsLoss()(real_scores, torch.ones(N).float().unsqueeze(-1)) \
+ nn.BCEWithLogitsLoss()(fake_scores, torch.zeros(N).float().unsqueeze(-1))
def noise(batch_size, noise_dim):
return torch.rand((batch_size, noise_dim)) * 2 - 1
class GAN:
def __init__(self, generator, discriminator, d_loss, g_loss, **kwargs):
self.generator = generator
self.discriminator = discriminator
self.num_epochs = 10
self.every = 100
self.d_overtrain = kwargs.get("d_overtrain", 1)
self.batch_size = kwargs.get("batch_size", 128)
self.img_size = kwargs.get("img_size", 28)
self.noise_dim = kwargs.get("noise_dim", 50)
self.n_classes = kwargs.get("n_classes", None)
self.d_optim = kwargs.get("d_optim",
lambda params: torch.optim.Adam(params, lr=0.0002, betas=(0.5, 1 - 1e-3)))(
self.discriminator.parameters())
self.g_optim = kwargs.get("g_optim",
lambda params: torch.optim.Adam(params, lr=0.0002, betas=(0.5,
1 -
1e-3)))(self.generator.parameters())
self.d_loss = d_loss
self.g_loss = g_loss
self.load_data()
def load_data(self):
transform = transforms.Compose(
[
# transforms.Resize(20),
transforms.ToTensor(),
])
dataset1 = datasets.MNIST('../data', train=True, download=True, transform=transform)
self.train_loader = torch.utils.data.DataLoader(dataset1, batch_size=self.batch_size)
def sample_generator(self, model_path):
net = torch.load(model_path)
net.eval()
N = 49
labels = (torch.ones(7,7)*(torch.arange(7).unsqueeze(-1))).type(torch.long)
with torch.no_grad():
fake_imgs = net(noise(N,NOISE_DIM), labels.view(-1))
self.show_images(fake_imgs.squeeze(), "samples")
def train(self, name):
if not hasattr('self', 'train_loader'):
print("loading data")
self.load_data()
print("Starting training")
counter = 0
for epoch in range(self.num_epochs):
for x, y in self.train_loader:
if len(x) != self.batch_size:
continue
for _ in range(self.d_overtrain):
d_loss, fake_images = self.train_discriminator(x,y)
g_loss = self.train_generator()
if counter % self.every == self.every - 1:
self.log_status(counter, d_loss, g_loss, fake_images, name)
counter += 1
def train_discriminator(self, realx, realy):
# train on real images
self.discriminator.zero_grad()
realx = 2 * (realx - 0.5)
# train on fake images
with torch.no_grad():
z = noise(self.batch_size, self.noise_dim)
fakex = self.generator(z, realy) # (batch_size, noise_dim)
loss = self.d_loss(self.discriminator, realx, fakex, realy)
loss.backward()
self.d_optim.step()
return loss, fakex
def train_generator(self):
self.generator.zero_grad()
fakey = torch.LongTensor(np.random.randint(0, self.n_classes, self.batch_size))
z = noise(self.batch_size, self.noise_dim)
fake_images = self.generator(z, fakey)
d_scores = self.discriminator(fake_images, fakey)
g_loss = self.g_loss(d_scores)
g_loss.backward()
self.g_optim.step()
return g_loss
def log_status(self, counter, d_loss, g_loss, fake_images, name):
print(f"Epoch {counter}: D loss: {d_loss} g loss: {g_loss}")
fake_images = fake_images.data.cpu().numpy().reshape(-1, self.img_size, self.img_size)
self.show_images(fake_images, f"{name}_imgs_step_{counter}.png")
torch.save(self.discriminator, f"models/discriminator_{counter}")
torch.save(self.generator, f"models/generator_{counter}")
def show_images(self, images,filename=None, max_imgs=50):
# images = images[:max_imgs].view(-1,28,28)
gridlen = math.floor(math.sqrt(max_imgs))
fig = plt.figure(figsize=(gridlen, gridlen))
gs = gridspec.GridSpec(gridlen, gridlen)
gs.update(wspace=0.05, hspace=0.05)
for i, img in enumerate(images):
ax = plt.subplot(gs[i])
plt.axis('off')
ax.set_xticklabels([])
ax.set_yticklabels([])
ax.set_aspect('equal')
plt.imshow(img)
if i == gridlen ** 2 - 1:
break
if filename:
plt.savefig(os.path.join("fake_imgs", filename))
else:
plt.show()
if __name__ == "__main__":
g_params = {"noise_dim": NOISE_DIM,
"img_dim": IMG_DIM,
"hidden_dim": 1024,
"num_conv1": 128,
"size_conv1": 4,
"num_conv2": 64,
"size_conv2": 4,
"n_classes": 10,
"label_emb_dim": 50
}
d_params = {"img_dim": IMG_DIM,
"num_conv1": 32,
"size_conv1": 5,
"num_conv2": 64,
"size_conv2": 5,
"n_classes": 10,
"label_emb_dim": 50
}
train_params = {"noise_dim": NOISE_DIM,
"img_size": IMG_DIM,
"batch_size": 128,
"d_overtrain": 3,
"n_classes": 10,
"d_loss": d_bceloss,
"g_loss": g_bceloss
}
G = cDCGenerator(**g_params)
D = cDCDiscriminator(**d_params)
gan = GAN(G, D, **train_params)
# gan.train("cGAN")
gan.sample_generator("models/generator_4299")