forked from CACTuS-AI/GlowIP
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_glow.py
197 lines (177 loc) · 9.05 KB
/
train_glow.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
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
import torch
from torchvision import datasets
import torchvision.transforms as transforms
from torchvision.utils import make_grid
from glow.glow import Glow
import numpy as np
import skimage.io as sio
import matplotlib.pyplot as plt
import os
import json
import argparse
def trainGlow(args):
save_path = "./trained_models/%s/glow"%args.dataset
training_folder = "./data/%s_preprocessed/train"%args.dataset
# setting up configs as json
config_path = save_path+"/configs.json"
configs = {"K":args.K,
"L":args.L,
"coupling":args.coupling,
"last_zeros":args.last_zeros,
"batchsize":args.batchsize,
"size":args.size,
"lr": args.lr,
"n_bits_x":args.n_bits_x,
"warmup_iter":args.warmup_iter}
if not os.path.exists(save_path):
print("creating directory to save model weights")
os.makedirs(save_path)
# loading pre-trained model to resume training
if os.path.exists(save_path+"/glowmodel.pt"):
print("loading previous model and saved configs to resume training ...")
with open(config_path, 'r') as f:
configs = json.load(f)
glow = Glow((3,configs["size"],configs["size"]),
K=configs["K"],L=configs["L"],
coupling=configs["coupling"],
n_bits_x=configs["n_bits_x"],
nn_init_last_zeros=configs["last_zeros"],
device=args.device)
glow.load_state_dict(torch.load(save_path+"/glowmodel.pt"))
print("pre-trained model and configs loaded successfully")
glow.set_actnorm_init()
print("actnorm initialization flag set to True to avoid data dependant re-initialization")
glow.train()
else:
# creating and initializing glow model
print("creating and initializing model for training")
glow = Glow((3,args.size,args.size),
K=args.K,L=args.L,coupling=args.coupling,n_bits_x=args.n_bits_x,
nn_init_last_zeros=args.last_zeros,
device=args.device)
glow.train()
print("saving configs as json file")
with open(config_path, 'w') as f:
json.dump(configs, f, sort_keys=True, indent=4, ensure_ascii=False)
# setting up dataloader
print("setting up dataloader for the training data")
trans = transforms.Compose([transforms.Resize((args.size,args.size)),
transforms.ToTensor()])
dataset = datasets.ImageFolder(training_folder, transform=trans)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batchsize,
drop_last=True, shuffle=True)
# setting up optimizer and learning rate scheduler
opt = torch.optim.Adam(glow.parameters(), lr=args.lr)
lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(opt,mode="min",
factor=0.5,
patience=1000,
verbose=True,
min_lr=1e-8)
# starting training code here
print("+-"*10,"starting training","-+"*10)
global_step = 0
global_loss = []
warmup_completed = False
for i in range(args.epochs):
Loss_epoch = []
for j, data in enumerate(dataloader):
opt.zero_grad()
glow.zero_grad()
# loading batch
x = data[0].to(device=args.device)*255
# pre-processing data
x = glow.preprocess(x)
# computing loss: "nll"
n,c,h,w = x.size()
nll,logdet,logpz,z_mu,z_std = glow.nll_loss(x)
# skipping first batch due to data dependant initialization (if not initialized)
if global_step == 0:
global_step += 1
continue
# backpropogating loss and gradient clipping
nll.backward()
torch.nn.utils.clip_grad_value_(glow.parameters(), 5)
grad_norm = torch.nn.utils.clip_grad_norm_(glow.parameters(), 100)
# linearly increase learning rate till warmup_iter upto args.lr
if global_step <= args.warmup_iter:
warmup_lr = args.lr / args.warmup_iter * global_step
for params in opt.param_groups:
params["lr"] = warmup_lr
# taking optimizer step
opt.step()
# learning rate scheduling after warm up iterations
if global_step > args.warmup_iter:
lr_scheduler.step(nll)
if not warmup_completed:
if args.warmup_iter == 0:
print("no model warming...")
else:
print("\nwarm up completed")
warmup_completed = True
# printing training metrics
print("\repoch=%0.2d..nll=%0.2f..logdet=%0.2f..logpz=%0.2f..mu=%0.2f..std=%0.2f..gradnorm=%0.2f"
%(i,nll.item(),logdet,logpz,z_mu,z_std,grad_norm),end="\r")
# saving generated samples during training
try:
if j % args.sample_freq == 0:
plt.plot(global_loss)
plt.xlabel("iterations",size=15)
plt.ylabel("nll",size=15)
plt.savefig(save_path+"/nll_training_curve.jpg")
plt.close()
with torch.no_grad():
z_sample, z_sample_t = glow.generate_z(n=10,mu=0,std=0.7,to_torch=True)
x_gen = glow(z_sample_t, reverse=True)
x_gen = glow.postprocess(x_gen)
x_gen = make_grid(x_gen,nrow=int(np.sqrt(len(x_gen))))
x_gen = x_gen.data.cpu().numpy()
x_gen = x_gen.transpose([1,2,0])
if x_gen.shape[-1] == 1:
x_gen = x_gen[...,0]
if not os.path.exists(save_path+"/samples_training"):
os.makedirs(save_path+"/samples_training")
x_gen = (x_gen * 255).astype("uint8")
sio.imsave(save_path+"/samples_training/%0.6d.jpg"%global_step, x_gen )
except:
print("\n failed to sample from glow at global step = %d"%global_step)
global_step = global_step + 1
global_loss.append(nll.item())
if global_step % args.save_freq == 0:
torch.save(glow.state_dict(), save_path+"/glowmodel.pt")
# # model visualization
# temperature = [0.1,0.3,0.4,0.5,0.7,0.8, 0.9]
# for temp in temperature:
# with torch.no_grad():
# glow.eval()
# z_sample, z_sample_t = glow.generate_z(n=10,mu=0,std=temp,to_torch=True)
# x_gen = glow(z_sample_t, reverse=True)
# x_gen = glow.postprocess(x_gen)
# x_gen = make_grid(x_gen,nrow=int(np.sqrt(len(x_gen))))
# x_gen = x_gen.data.cpu().numpy()
# x_gen = x_gen.transpose([1,2,0])
# if x_gen.shape[-1] == 1:
# x_gen = x_gen[...,0]
# plt.figure()
# plt.title("temperature = %0.1f"%temp,fontsize=15)
# plt.axis("off")
# plt.imshow(x_gen)
# saving model weights
torch.save(glow.state_dict(), save_path+"/glowmodel.pt")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='train glow network')
parser.add_argument('-dataset',type=str,help='the dataset to train the model on', default='celeba')
parser.add_argument('-K',type=int,help='no. of steps of flow',default=48)
parser.add_argument('-L',type=int,help='no. of time squeezing is performed',default=4)
parser.add_argument('-coupling',type=str,help='type of coupling layer to use',default='affine')
parser.add_argument('-last_zeros',type=bool,help='whether to initialize last layer ot NN with zeros',default=True)
parser.add_argument('-batchsize',type=int,help='batch size for training',default=6)
parser.add_argument('-size',type=int,help='images will be resized to this dimension',default=64)
parser.add_argument('-lr',type=float,help='learning rate for training',default=1e-4)
parser.add_argument('-n_bits_x',type=int,help='requantization of training images',default=5)
parser.add_argument('-epochs',type=int,help='epochs to train for',default=1000)
parser.add_argument('-warmup_iter',type=int,help='no. of warmup iterations',default=10000)
parser.add_argument('-sample_freq',type=int,help='sample after every save_freq',default=50)
parser.add_argument('-save_freq',type=int,help='save after every save_freq',default=1000)
parser.add_argument('-device',type=str,help='whether to use',default="cuda")
args = parser.parse_args()
trainGlow(args)