-
Notifications
You must be signed in to change notification settings - Fork 0
/
losses.py
106 lines (88 loc) · 4.33 KB
/
losses.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
import torch
import torch.optim as optim
from torch.autograd import Variable
import torch.nn.functional as F
def get_optimizer(config, model):
if config.model.name == 'unet':
optimizer = optim.Adam(model.parameters(), lr=config.optim.lr, betas=(config.optim.beta1, 0.999),
eps=config.optim.eps, weight_decay=config.optim.weight_decay)
if config.model.name == 'fcdense':
optimizer = optim.RMSprop(model.parameters(), lr=config.optim.lr, weight_decay=config.optim.weight_decay)
elif config.model.name == 'fcn':
optimizer = optim.RMSprop(model.parameters(), lr=config.optim.lr, momentum=config.optim.momentum,
weight_decay=config.optim.weight_decay)
return optimizer
def get_cross_entropy_loss(config):
def cross_entropy_one_hot_cityscapes(pred, targets):
targets = torch.argmax(targets, dim=1)
return F.cross_entropy(pred, Variable(targets), ignore_index=0)
def cross_entropy_one_hot_flickr(pred, targets):
targets = torch.argmax(targets, dim=1)
return F.cross_entropy(pred, Variable(targets))
if config.data.dataset == 'cityscapes256':
return cross_entropy_one_hot_cityscapes
elif config.data.dataset == 'flickr':
return cross_entropy_one_hot_flickr
def get_nll_loss(config):
def nll_loss_cityscapes(pred, targets):
targets = torch.argmax(targets, dim=1)
weights = torch.Tensor([0, 0.8373, 0.918, 0.866, 1.0345,
1.0166, 0.9969, 0.9754, 1.0489,
0.8786, 1.0023, 0.9539, 0.9843,
1.1116, 0.9037, 1.0865, 1.0955,
1.0865, 1.1529, 1.0507])
weights = weights.to(config.device, dtype=torch.float32)
return F.nll_loss(pred, Variable(targets), weight=weights, ignore_index=0)
def nll_loss_flickr(pred, targets):
targets = torch.argmax(targets, dim=1)
return F.nll_loss(pred, Variable(targets))
if config.data.dataset == 'cityscapes256':
return nll_loss_cityscapes
elif config.data.dataset == 'flickr':
return nll_loss_flickr
def get_loss_fn(config):
if config.model.name == 'unet':
return get_cross_entropy_loss(config)
if config.model.name == 'fcdense':
return get_nll_loss(config)
elif config.model.name == 'fcn':
return F.binary_cross_entropy_with_logits
def get_step_fn(config, optimizer, model, loss_fn, sde=None, scaler=None, train=True):
def step_fn(img, target):
# Conditioning on noise scales
if config.model.conditional:
eps = 1e-5
# t = (0.4 - 1) * torch.rand(int(img.shape[0]), device=config.device) + 1
if train:
t = torch.rand(int(img.shape[0]), device=config.device) * (1 - eps) + eps
else:
t = torch.linspace(1, eps, img.shape[0], device=config.device)
z = torch.randn_like(img)
mean, std = sde.marginal_prob(img, t)
perturbed_img = mean + std[:, None, None, None] * z
max = torch.ones(perturbed_img.shape[0], device=config.device)
min = torch.ones(perturbed_img.shape[0], device=config.device)
for N in range(perturbed_img.shape[0]):
max[N] = torch.max(perturbed_img[N, :, :, :])
min[N] = torch.min(perturbed_img[N, :, :, :])
perturbed_img = perturbed_img - min[:, None, None, None] * torch.ones_like(img, device=config.device)
perturbed_img = torch.div(perturbed_img, (max - min)[:, None, None, None])
# Training step
if train:
optimizer.zero_grad()
if not config.optim.mixed_prec:
pred = model(img) if not config.model.conditional else model(perturbed_img, std)
loss = loss_fn(pred, target)
if train:
loss.backward()
optimizer.step()
else:
with torch.cuda.amp.autocast():
pred = model(img) if not config.model.conditional else model(perturbed_img, std)
loss = loss_fn(pred, target)
if train:
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
return loss.detach().item(), pred.detach().cpu()
return step_fn