-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy patheval.py
163 lines (148 loc) · 7.24 KB
/
eval.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
import os
import argparse
import logging
import torch
import torch.distributed as dist
from torchvision.utils import save_image
from sde import Follmer
from misc import add_dict_to_argparser, create_network, create_karras_grid, create_dataset, create_dataloader
from defaults import data_defaults, model_defaults, eval_defaults
from dist_util import setup_dist, get_device
import warnings
warnings.filterwarnings('ignore', 'Grad strides do not match bucket view strides') # False warning printed by PyTorch 1.12.
torch._logging.set_logs(all=logging.ERROR)
def create_parser():
parser = argparse.ArgumentParser()
defaults = dict(
data_name="mnist",
start_seed=0,
end_seed=9999,
mode="denoiser",
workdir="tmp/",
checkpoint="",
)
defaults.update(data_defaults(defaults["data_name"]))
defaults.update(model_defaults(defaults["data_name"]))
defaults.update(eval_defaults(defaults["data_name"]))
add_dict_to_argparser(parser, defaults)
return parser
class StackedRandomGenerator:
def __init__(self, device, seeds):
super().__init__()
self.generators = [torch.Generator(device).manual_seed(int(seed) % (1 << 32)) for seed in seeds]
def randn(self, size, **kwargs):
assert size[0] == len(self.generators)
return torch.stack([torch.randn(size[1:], generator=gen, **kwargs) for gen in self.generators])
def randn_like(self, input):
return self.randn(input.shape, dtype=input.dtype, layout=input.layout, device=input.device)
def randint(self, *args, size, **kwargs):
assert size[0] == len(self.generators)
return torch.stack([torch.randint(*args, size=size[1:], generator=gen, **kwargs) for gen in self.generators])
@torch.no_grad()
def ode_sampler(sde, model, latents, x_mean=None, label=None, solver='euler', rng=None, num_steps=18, rho=7, t_min=1e-4, t_max=0.9999, churn=False, s_min=0.05, s_max=0.6, gamma=0.4):
"""Enhanced ODE sampler, solves from latents at 1 to 0.
(1, t_max) is solved with x_mean(exact velocity at 1 is data mean, we use empirical mean here).
(t_max, t_min) is solved using `solver`.
Inject noises over interval (s_min, s_max), with noise level (1+gamma)*t if churn=True.
(t_min, 0) is solved with one euler step if using heun solver as it faces singularity at zero.
(t_max, t_min) is discretized by Karras grid."""
bsz = latents.shape[0]
device = latents.device
xt = latents - (1-t_max) * x_mean if x_mean is not None else latents
step_fn = sde._get_step_fn(solver)
if num_steps == 1:
t_cur = torch.ones([bsz, ], device=device)*t_max
dt = t_min-t_max
xt = sde.deis_step(model, xt, t_cur, label, dt)
else:
grid = create_karras_grid(t_min, t_max, num_steps, rho).to(device)
for i in range(num_steps-1):
t_cur = grid[i]
if churn and (s_min <= t_cur <= s_max):
t_churn = t_cur*(1+gamma)
xt = sde.churn_noise(xt, t_cur, t_churn, rng)
t_cur = t_churn
dt = grid[i+1] - t_cur
t_cur = torch.ones([bsz, ], device=device)*t_cur
xt = step_fn(model, xt, t_cur, label, dt)
t_cur = torch.ones([bsz, ], device=device)*t_min
dt = -t_min
xt = sde.deis_step(model, xt, t_cur, label, dt) if solver.lower() == 'deis' else sde.euler_step(model, xt, t_cur, label, dt)
return xt
@torch.no_grad()
def characteristic_sampler(sde, model, latents, x_mean, label=None, rng=None, num_steps=18, rho=7, t_min=1e-4, t_max=0.9999, churn=False, s_min=0.05, s_max=0.6, gamma=0.4):
"""Enhanced ODE sampler, solves from latents at t_max to t_min.
(t_max, t_min) is discretized by Karras grid.
"""
bsz = latents.shape[0]
device = latents.device
xt = latents - (1-t_max) * x_mean
if num_steps == 1:
t_cur = torch.ones(bsz, device=device)*t_max
t_next = torch.zeros(bsz, device=device)
_, xt = sde.get_denoiser_and_traj(model, xt, t_cur, t_next, label)
else:
grid = create_karras_grid(t_min, t_max, num_steps, rho).to(device)
for i in range(num_steps-1):
t_cur = grid[i]
if churn and (s_min <= t_cur <= s_max):
t_churn = t_cur*(1+gamma)
xt = sde.churn_noise(xt, t_cur, t_churn, rng)
t_cur = t_churn
t_cur = torch.ones(bsz, device=device)*t_cur
t_next = torch.ones(bsz, device=device)*grid[i+1]
_, xt = sde.get_denoiser_and_traj(model, xt, t_cur, t_next, label)
t_cur = torch.ones(bsz, device=device)*t_min
t_next = torch.zeros(bsz, device=device)
_, xt = sde.get_denoiser_and_traj(model, xt, t_cur, t_next, label)
return xt
@torch.no_grad()
def main(args):
setup_dist()
os.makedirs(args.workdir, exist_ok=True)
device = get_device()
seeds = range(args.start_seed, args.end_seed+1)
num_batches = ((len(seeds) - 1) // (args.eval_bsz * dist.get_world_size()) + 1) * dist.get_world_size()
all_batches = torch.as_tensor(seeds).tensor_split(num_batches)
rank_batches = all_batches[dist.get_rank() :: dist.get_world_size()]
model = create_network(args, traj_mode=True if args.mode.lower() == 'characteristic' else False).eval().to(device)
model.load_state_dict(torch.load(args.checkpoint, map_location=device)['ema'])
sde = Follmer(args)
x_mean = _compute_x_mean(args).to(device)
for batch_seeds in rank_batches:
batch_size = len(batch_seeds)
if batch_size == 0:
continue
rng = StackedRandomGenerator(device, batch_seeds)
latents = rng.randn([batch_size, args.in_channels, args.img_resolution, args.img_resolution], device=device)
labels = None
if args.label_dim:
labels = torch.eye(args.label_dim, device=device)[rng.randint(args.label_dim, size=[batch_size], device=device)]
if args.class_idx is not None:
labels[:, :] = 0
labels[:, args.class_idx] = 1
if args.mode == 'denoiser':
images = ode_sampler(sde, model, latents, x_mean, labels, args.ode_solver, rng, args.num_steps, args.rho, args.eval_eps0, 1-args.eval_eps1, args.churn, args.s_min, args.s_max, args.gamma)
elif args.mode == 'characteristic':
images = characteristic_sampler(sde, model, latents, x_mean, labels, rng, args.num_steps, args.rho, args.eval_eps0, 1-args.eval_eps1, args.churn, args.s_min, args.s_max, args.gamma)
else:
raise ValueError(f"unsupported eval mode {args.mode}")
images = (images+1.)/2 if args.data_name.lower() != 'mnist' else images
for seed, image in zip(batch_seeds, images):
save_image(image, f'{args.workdir}/{seed:06d}.png')
def _compute_x_mean(args):
dataset = create_dataset(args)
# compute target mean
loader = create_dataloader(dataset, args.bsz, args.num_workers, infinite=False)
img_shape = [args.in_channels, args.img_resolution, args.img_resolution]
x_mean = torch.zeros(img_shape)
n_batch = 0
for batch, _ in loader:
x_mean += batch.mean(dim=0)
n_batch += 1
x_mean /= n_batch
x_mean = x_mean*2.-1 if args.data_name.lower() != 'mnist' else x_mean
return x_mean
if __name__ == '__main__':
args = create_parser().parse_args()
main(args)