-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_utils.py
180 lines (149 loc) · 7.83 KB
/
train_utils.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
import argparse
import glob
import os
from copy import deepcopy
import torch
from torch import optim as optim
from models import get_models
def parse_train_args(arguments_string=None):
parser = argparse.ArgumentParser()
# Data
parser.add_argument('--data_path', default="/mnt/storage_ssd/datasets/FFHQ/FFHQ_1000/FFHQ_1000",
help="Path to train images")
parser.add_argument('--limit_data', default=None, type=int, help="limit the size of the dataset")
parser.add_argument('--center_crop', default=None, help='center_crop_data to specified size', type=int)
parser.add_argument('--gray_scale', action='store_true', default=False, help="Load data as grayscale")
# Model
parser.add_argument('--gen_arch', default='DCGAN')
parser.add_argument('--disc_arch', default='DCGAN')
parser.add_argument('--im_size', default=64, type=int)
parser.add_argument('--z_dim', default=64, type=int)
parser.add_argument('--z_prior', default="normal", type=str, help="[normal, binary, uniform]")
parser.add_argument('--spectral_normalization', action='store_true', default=False)
parser.add_argument('--weight_clipping', type=float, default=None)
parser.add_argument('--gp_weight', default=0, type=float)
# Training
parser.add_argument('--r_bs', default=64, type=int, help="Real data batch size: -1 for automaticly set as full size batch size")
parser.add_argument('--f_bs', default=64, type=int, help="Fake data batch size")
parser.add_argument('--loss_function', default="NonSaturatingGANLoss", type=str)
parser.add_argument('--lrG', default=0.0001, type=float)
parser.add_argument('--lrD', default=0.0001, type=float)
parser.add_argument('--avg_update_factor', default=1, type=float,
help='moving average factor weight of updating generator (1 means none)')
parser.add_argument('--D_step_every', default=1, type=int, help="D G only evry 'D_step_every' iterations")
parser.add_argument('--G_step_every', default=1, type=int, help="Update G only evry 'G_step_every' iterations")
parser.add_argument('--n_iterations', default=1000000, type=int)
parser.add_argument('--no_fake_resample', default=False, action='store_true')
# Evaluation
parser.add_argument('--wandb', action='store_true', default=False, help="Otherwise use PLT localy")
parser.add_argument('--log_freq', default=1000, type=int)
parser.add_argument('--save_every', action='store_true', default=False)
parser.add_argument('--full_batch_metrics', nargs='*', default=[
'MiniBatchLoss-dist=w1',
'MiniBatchLoss-dist=swd',
# 'MiniBatchPatchLoss-dist=swd-p=16-s=8',
'MiniBatchLocalPatchLoss-dist=swd-p=16-s=8',
'MiniBatchLocalPatchLoss-dist=full_dim_swd-p=16-s=8',
# 'MiniBatchPatchLoss-dist=fd-p=8-s=8',
# 'MiniBatchNeuralLoss-dist=fd-device=cuda:0-b=64-layer_idx=9',
# 'MiniBatchNeuralLoss-dist=fd-device=cuda:0-b=64-layer_idx=18',
# 'MiniBatchNeuralPatchLoss-dist=fd-device=cuda:0-b=64-layer_idx=9',
# 'MiniBatchNeuralPatchLoss-dist=fd-device=cuda:0-b=64-layer_idx=18',
])
# Other
parser.add_argument('--project_name', default='GANs')
parser.add_argument('--train_name', default=None)
parser.add_argument('--n_workers', default=4, type=int)
parser.add_argument('--loadG', default=None, type=str)
parser.add_argument('--resume_last_ckpt', action='store_true', default=False,
help="Search for the latest ckpt in the same folder to resume training")
parser.add_argument('--load_data_to_memory', action='store_true', default=False)
parser.add_argument('--device', default="cuda:0")
if arguments_string is not None:
arguments_string = arguments_string.split()
return parser.parse_args(arguments_string)
def copy_G_params(model):
flatten = deepcopy(list(p.data for p in model.parameters()))
return flatten
def load_params(model, new_param):
for p, new_p in zip(model.parameters(), new_param):
p.data.copy_(new_p)
class Prior:
def __init__(self, prior_type, z_dim):
self.prior_type = prior_type
self.z_dim = z_dim
self.z = None
if "const" in self.prior_type:
self.b = int(self.prior_type.split("=")[1])
def sample(self, b):
if "const" in self.prior_type:
if self.z is None:
self.z = torch.randn((self.b, self.z_dim))
if b != self.b:
z = self.z[torch.randint(self.b, (b,))]
else:
z = self.z
elif self.prior_type == "binary":
z = torch.sign(torch.randn((b, self.z_dim)))
elif self.prior_type == "uniform":
z = torch.rand((b, self.z_dim))
else:
z = torch.randn((b, self.z_dim))
return z
def save_model(prior, netG, netD, optimizerG, optimizerD, saved_model_folder, iteration, args):
fname = f"{saved_model_folder}/{'last' if not args.save_every else iteration}.pth"
torch.save({"iteration": iteration,
'prior': prior.z,
'netG': netG.state_dict(),
'netD': netD.state_dict(),
"optimizerG": optimizerG.state_dict(),
"optimizerD": optimizerD.state_dict()
},
fname)
def get_models_and_optimizers(args, device, saved_model_folder):
prior = Prior(args.z_prior, args.z_dim)
netG, netD = get_models(args, device)
netG.train()
netD.train()
optimizerG = optim.Adam(netG.parameters(), lr=args.lrG, betas=(0.5, 0.9))
optimizerD = optim.Adam(netD.parameters(), lr=args.lrD, betas=(0.5, 0.9))
if args.loadG is not None:
ckpt = torch.load(args.loadG, map_location=args.device)
netG.load_state_dict(ckpt['netG'])
prior.z = ckpt['prior']
start_iteration = 0
if args.resume_last_ckpt:
ckpts = glob.glob(f'{saved_model_folder}/*.pth')
if ckpts:
latest_ckpt = max(ckpts, key = os.path.getctime)
ckpt = torch.load(latest_ckpt, map_location=args.device)
prior.z = ckpt['prior']
netG.load_state_dict(ckpt['netG'])
netD.load_state_dict(ckpt['netD'])
optimizerG.load_state_dict(ckpt['optimizerG'])
optimizerD.load_state_dict(ckpt['optimizerD'])
start_iteration = ckpt['iteration']
print(f"Loaded ckpt of iteration: {start_iteration}")
return prior, netG, netD, optimizerG, optimizerD, start_iteration
def calc_gradient_penalty(netD, real_data, fake_data, one_sided=False):
"""Ensure the netD is smooth by forcing the gradient between real and fake data to ahve norm of 1"""
device = real_data.device
alpha = torch.rand(1, 1)
alpha = alpha.expand(real_data.size())
alpha = alpha.to(device)
interpolates = alpha * real_data + ((1 - alpha) * fake_data)
interpolates = interpolates.to(device)
interpolates = torch.autograd.Variable(interpolates, requires_grad=True)
disc_interpolates = netD(interpolates)
gradients = torch.autograd.grad(outputs=disc_interpolates,
inputs=interpolates,
grad_outputs=torch.ones(disc_interpolates.size()).to(device),
create_graph=True, retain_graph=True,
only_inputs=True)[0]
gradients = gradients.view(gradients.shape[0], -1)
gradient_norm = gradients.norm(2, dim=1)
diff = (gradient_norm - 1)
if one_sided:
diff = torch.clamp(diff, min=0)
gradient_penalty = (diff ** 2).mean()
return gradient_penalty, gradient_norm.mean().item()