-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcdcgan.py
170 lines (139 loc) · 6.12 KB
/
cdcgan.py
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
import numpy as np
import torch
import os
import matplotlib.pyplot as plt
from torchvision import datasets
import torchvision.transforms as transforms
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from PIL import Image
from mpl_toolkits.axes_grid1 import ImageGrid
import sys
import pickle
from collections import defaultdict
from funcs import *
from models import *
from pathlib import Path
latent_vector_size = 100
class_size=7
batch_size = 64
lr=0.0002
generator_scale=128
discriminator_scale=128
save_file=2
gen_steps=1
desc_steps=1
device = None
import torchvision.utils as vutils
if torch.cuda.is_available():
print("Using GPU.")
device = torch.device('cuda:0')
else:
print("Using CPU.")
device = torch.device("cpu")
def gan_loss(discriminator_preds, labels):
criterion = nn.BCELoss()
loss = criterion(discriminator_preds.squeeze(), labels)
return loss
def train_cgan(generator, discriminator, train_loader, lr=0.0001, latent_vector_size=100, nepochs=100, print_freq=400, device="cuda",batch_size=32,file_name="models",probs=False,losses=[]):
'''
The function trains a gan model.
'''
# Create optimizers for the discriminator and generator
# each optimizes different model
d_optimizer = optim.Adam(discriminator.parameters(), lr, betas=(0.5, 0.999))
g_optimizer = optim.Adam(generator.parameters(), lr, betas=(0.5, 0.999))
# start train loop
mydesc_step=0
global desc_steps
for epoch in range(nepochs):
# set both networks to train mode
discriminator.train()
generator.train()
for batch_i, (real_images, real_classes) in enumerate(train_loader):
# rescale images before training
real_images = (real_images*2 - 1).to(device)
# get batch size
batch_size = real_images.size(0)
real_classes = real_classes.to(device)
######### A. TRAIN THE DISCRIMINATOR #########
d_optimizer.zero_grad()
mydesc_step=mydesc_step+1
# 1. Compute the discriminator loss on real images
discriminator_batch_labels = create_labels(batch_size, device, real_data=True, for_discriminator=True)
disc_output=discriminator(real_images,real_classes)
loss1 = gan_loss(disc_output,discriminator_batch_labels)
# 2. Generate fake images using the generator (use ```create_latent_batch_vectors```)
latent_batch,classes = create_latent_batch_vectors(batch_size=batch_size, latent_vector_size=latent_vector_size, device=device,class_size=class_size,probs=probs)
generator_images = generator(latent_batch,classes)
# 3. Compute the discriminator loss on fake images
generator_batch_labels = create_labels(batch_size, device, real_data=False, for_discriminator=True)
disc_output=discriminator(generator_images,classes.type(torch.LongTensor).to(device))
loss2 = gan_loss(disc_output,generator_batch_labels)
# 4. Calculate discriminator total loss and do backprop
discriminator_loss = loss1 + loss2
discriminator_loss.backward()
######################
d_optimizer.step()
if mydesc_step%desc_steps!=0:
continue
######### B. TRAIN THE GENERATOR #########
for i in range(gen_steps):
g_optimizer.zero_grad()
# 1. Generate fake images
latent_batch,classes = create_latent_batch_vectors(batch_size=batch_size, latent_vector_size=latent_vector_size, device=device,class_size=class_size,probs=probs)
generator_images = generator(latent_batch,classes)
# 2. Compute the discriminator loss on them but with flipped labels.
flipped_labels = create_labels(batch_size, device, real_data=False, for_discriminator=False)
disc_output=discriminator(generator_images,classes.type(torch.LongTensor).to(device))
generator_loss = gan_loss(disc_output,flipped_labels)
# 3. perform backprop
generator_loss.backward()
######################
g_optimizer.step()
print(f'Epoch {epoch}, Batch {batch_i}, Disc_loss: {discriminator_loss.item()}, Gen_loss: {generator_loss.item()}')
if((epoch+1)%save_file==0):
file_save(file_name,generator, discriminator,losses,device)
# keep track of losses
losses.append((discriminator_loss.item(), generator_loss.item()))
generator.eval()
return losses
def main():
file_name="models"
if(len(sys.argv)>1):
file_name=sys.argv[1]
file_name=file_name+".pkl"
print(file_name)
my_file = Path(file_name)
if my_file.is_file() and len(sys.argv)>1:
with open(file_name,"rb") as f:
gen, disc,losses= pickle.load(f)
gen.to(device)
disc.to(device)
if(len(sys.argv)>2 and sys.argv[2]=="create"):
return create(gen)
# file exists
else:
print("new file")
gen=Generator(100,generator_scale).to(device)
disc=Discriminator_cgan(discriminator_scale).to(device)
gen.apply(weights_init)
disc.apply(weights_init)
losses=[]
mydata=load_data("list_patition_label.txt")
classess_options=['Surprised',"Fear","Disgusted","happy","sad","Anger","Neutral"]
train_tfms = transforms.Compose([transforms.ToTensor()])
my_dataset = CustomDataSet("aligned/",mydata,classess_options, transform=train_tfms)
train_loader = DataLoader(my_dataset, batch_size=batch_size,
shuffle=True, num_workers=0, drop_last=True)
print(my_dataset.sizes)
total=my_dataset.total_size
sizes=my_dataset.sizes
probs=[]
for i in range(len(sizes)):
probs.append(sizes[i]/total)
losess=train_cgan(gen, disc, train_loader, lr=lr, latent_vector_size=latent_vector_size, nepochs=300, print_freq=5, device=device,file_name=file_name,probs=probs,losses=losses)
if __name__ == '__main__':
main()