Fabrice-TIERCELIN
commited on
Commit
•
6d08643
1
Parent(s):
dd6a633
Upload 13 files
Browse files- sgm/modules/diffusionmodules/denoiser.py +73 -0
- sgm/modules/diffusionmodules/denoiser_scaling.py +31 -0
- sgm/modules/diffusionmodules/denoiser_weighting.py +24 -0
- sgm/modules/diffusionmodules/discretizer.py +69 -0
- sgm/modules/diffusionmodules/guiders.py +88 -0
- sgm/modules/diffusionmodules/loss.py +69 -0
- sgm/modules/diffusionmodules/model.py +743 -0
- sgm/modules/diffusionmodules/openaimodel.py +1272 -0
- sgm/modules/diffusionmodules/sampling.py +766 -0
- sgm/modules/diffusionmodules/sampling_utils.py +48 -0
- sgm/modules/diffusionmodules/sigma_sampling.py +40 -0
- sgm/modules/diffusionmodules/util.py +309 -0
- sgm/modules/diffusionmodules/wrappers.py +103 -0
sgm/modules/diffusionmodules/denoiser.py
ADDED
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
|
3 |
+
from ...util import append_dims, instantiate_from_config
|
4 |
+
|
5 |
+
|
6 |
+
class Denoiser(nn.Module):
|
7 |
+
def __init__(self, weighting_config, scaling_config):
|
8 |
+
super().__init__()
|
9 |
+
|
10 |
+
self.weighting = instantiate_from_config(weighting_config)
|
11 |
+
self.scaling = instantiate_from_config(scaling_config)
|
12 |
+
|
13 |
+
def possibly_quantize_sigma(self, sigma):
|
14 |
+
return sigma
|
15 |
+
|
16 |
+
def possibly_quantize_c_noise(self, c_noise):
|
17 |
+
return c_noise
|
18 |
+
|
19 |
+
def w(self, sigma):
|
20 |
+
return self.weighting(sigma)
|
21 |
+
|
22 |
+
def __call__(self, network, input, sigma, cond):
|
23 |
+
sigma = self.possibly_quantize_sigma(sigma)
|
24 |
+
sigma_shape = sigma.shape
|
25 |
+
sigma = append_dims(sigma, input.ndim)
|
26 |
+
c_skip, c_out, c_in, c_noise = self.scaling(sigma)
|
27 |
+
c_noise = self.possibly_quantize_c_noise(c_noise.reshape(sigma_shape))
|
28 |
+
return network(input * c_in, c_noise, cond) * c_out + input * c_skip
|
29 |
+
|
30 |
+
|
31 |
+
class DiscreteDenoiser(Denoiser):
|
32 |
+
def __init__(
|
33 |
+
self,
|
34 |
+
weighting_config,
|
35 |
+
scaling_config,
|
36 |
+
num_idx,
|
37 |
+
discretization_config,
|
38 |
+
do_append_zero=False,
|
39 |
+
quantize_c_noise=True,
|
40 |
+
flip=True,
|
41 |
+
):
|
42 |
+
super().__init__(weighting_config, scaling_config)
|
43 |
+
sigmas = instantiate_from_config(discretization_config)(
|
44 |
+
num_idx, do_append_zero=do_append_zero, flip=flip
|
45 |
+
)
|
46 |
+
self.register_buffer("sigmas", sigmas)
|
47 |
+
self.quantize_c_noise = quantize_c_noise
|
48 |
+
|
49 |
+
def sigma_to_idx(self, sigma):
|
50 |
+
dists = sigma - self.sigmas[:, None]
|
51 |
+
return dists.abs().argmin(dim=0).view(sigma.shape)
|
52 |
+
|
53 |
+
def idx_to_sigma(self, idx):
|
54 |
+
return self.sigmas[idx]
|
55 |
+
|
56 |
+
def possibly_quantize_sigma(self, sigma):
|
57 |
+
return self.idx_to_sigma(self.sigma_to_idx(sigma))
|
58 |
+
|
59 |
+
def possibly_quantize_c_noise(self, c_noise):
|
60 |
+
if self.quantize_c_noise:
|
61 |
+
return self.sigma_to_idx(c_noise)
|
62 |
+
else:
|
63 |
+
return c_noise
|
64 |
+
|
65 |
+
|
66 |
+
class DiscreteDenoiserWithControl(DiscreteDenoiser):
|
67 |
+
def __call__(self, network, input, sigma, cond, control_scale):
|
68 |
+
sigma = self.possibly_quantize_sigma(sigma)
|
69 |
+
sigma_shape = sigma.shape
|
70 |
+
sigma = append_dims(sigma, input.ndim)
|
71 |
+
c_skip, c_out, c_in, c_noise = self.scaling(sigma)
|
72 |
+
c_noise = self.possibly_quantize_c_noise(c_noise.reshape(sigma_shape))
|
73 |
+
return network(input * c_in, c_noise, cond, control_scale) * c_out + input * c_skip
|
sgm/modules/diffusionmodules/denoiser_scaling.py
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
|
4 |
+
class EDMScaling:
|
5 |
+
def __init__(self, sigma_data=0.5):
|
6 |
+
self.sigma_data = sigma_data
|
7 |
+
|
8 |
+
def __call__(self, sigma):
|
9 |
+
c_skip = self.sigma_data**2 / (sigma**2 + self.sigma_data**2)
|
10 |
+
c_out = sigma * self.sigma_data / (sigma**2 + self.sigma_data**2) ** 0.5
|
11 |
+
c_in = 1 / (sigma**2 + self.sigma_data**2) ** 0.5
|
12 |
+
c_noise = 0.25 * sigma.log()
|
13 |
+
return c_skip, c_out, c_in, c_noise
|
14 |
+
|
15 |
+
|
16 |
+
class EpsScaling:
|
17 |
+
def __call__(self, sigma):
|
18 |
+
c_skip = torch.ones_like(sigma, device=sigma.device)
|
19 |
+
c_out = -sigma
|
20 |
+
c_in = 1 / (sigma**2 + 1.0) ** 0.5
|
21 |
+
c_noise = sigma.clone()
|
22 |
+
return c_skip, c_out, c_in, c_noise
|
23 |
+
|
24 |
+
|
25 |
+
class VScaling:
|
26 |
+
def __call__(self, sigma):
|
27 |
+
c_skip = 1.0 / (sigma**2 + 1.0)
|
28 |
+
c_out = -sigma / (sigma**2 + 1.0) ** 0.5
|
29 |
+
c_in = 1.0 / (sigma**2 + 1.0) ** 0.5
|
30 |
+
c_noise = sigma.clone()
|
31 |
+
return c_skip, c_out, c_in, c_noise
|
sgm/modules/diffusionmodules/denoiser_weighting.py
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
class UnitWeighting:
|
4 |
+
def __call__(self, sigma):
|
5 |
+
return torch.ones_like(sigma, device=sigma.device)
|
6 |
+
|
7 |
+
|
8 |
+
class EDMWeighting:
|
9 |
+
def __init__(self, sigma_data=0.5):
|
10 |
+
self.sigma_data = sigma_data
|
11 |
+
|
12 |
+
def __call__(self, sigma):
|
13 |
+
return (sigma**2 + self.sigma_data**2) / (sigma * self.sigma_data) ** 2
|
14 |
+
|
15 |
+
|
16 |
+
class VWeighting(EDMWeighting):
|
17 |
+
def __init__(self):
|
18 |
+
super().__init__(sigma_data=1.0)
|
19 |
+
|
20 |
+
|
21 |
+
class EpsWeighting:
|
22 |
+
def __call__(self, sigma):
|
23 |
+
return sigma**-2
|
24 |
+
|
sgm/modules/diffusionmodules/discretizer.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from abc import abstractmethod
|
2 |
+
from functools import partial
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
|
7 |
+
from ...modules.diffusionmodules.util import make_beta_schedule
|
8 |
+
from ...util import append_zero
|
9 |
+
|
10 |
+
|
11 |
+
def generate_roughly_equally_spaced_steps(
|
12 |
+
num_substeps: int, max_step: int
|
13 |
+
) -> np.ndarray:
|
14 |
+
return np.linspace(max_step - 1, 0, num_substeps, endpoint=False).astype(int)[::-1]
|
15 |
+
|
16 |
+
|
17 |
+
class Discretization:
|
18 |
+
def __call__(self, n, do_append_zero=True, device="cpu", flip=False):
|
19 |
+
sigmas = self.get_sigmas(n, device=device)
|
20 |
+
sigmas = append_zero(sigmas) if do_append_zero else sigmas
|
21 |
+
return sigmas if not flip else torch.flip(sigmas, (0,))
|
22 |
+
|
23 |
+
@abstractmethod
|
24 |
+
def get_sigmas(self, n, device):
|
25 |
+
pass
|
26 |
+
|
27 |
+
|
28 |
+
class EDMDiscretization(Discretization):
|
29 |
+
def __init__(self, sigma_min=0.02, sigma_max=80.0, rho=7.0):
|
30 |
+
self.sigma_min = sigma_min
|
31 |
+
self.sigma_max = sigma_max
|
32 |
+
self.rho = rho
|
33 |
+
|
34 |
+
def get_sigmas(self, n, device="cpu"):
|
35 |
+
ramp = torch.linspace(0, 1, n, device=device)
|
36 |
+
min_inv_rho = self.sigma_min ** (1 / self.rho)
|
37 |
+
max_inv_rho = self.sigma_max ** (1 / self.rho)
|
38 |
+
sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** self.rho
|
39 |
+
return sigmas
|
40 |
+
|
41 |
+
|
42 |
+
class LegacyDDPMDiscretization(Discretization):
|
43 |
+
def __init__(
|
44 |
+
self,
|
45 |
+
linear_start=0.00085,
|
46 |
+
linear_end=0.0120,
|
47 |
+
num_timesteps=1000,
|
48 |
+
):
|
49 |
+
super().__init__()
|
50 |
+
self.num_timesteps = num_timesteps
|
51 |
+
betas = make_beta_schedule(
|
52 |
+
"linear", num_timesteps, linear_start=linear_start, linear_end=linear_end
|
53 |
+
)
|
54 |
+
alphas = 1.0 - betas
|
55 |
+
self.alphas_cumprod = np.cumprod(alphas, axis=0)
|
56 |
+
self.to_torch = partial(torch.tensor, dtype=torch.float32)
|
57 |
+
|
58 |
+
def get_sigmas(self, n, device="cpu"):
|
59 |
+
if n < self.num_timesteps:
|
60 |
+
timesteps = generate_roughly_equally_spaced_steps(n, self.num_timesteps)
|
61 |
+
alphas_cumprod = self.alphas_cumprod[timesteps]
|
62 |
+
elif n == self.num_timesteps:
|
63 |
+
alphas_cumprod = self.alphas_cumprod
|
64 |
+
else:
|
65 |
+
raise ValueError
|
66 |
+
|
67 |
+
to_torch = partial(torch.tensor, dtype=torch.float32, device=device)
|
68 |
+
sigmas = to_torch((1 - alphas_cumprod) / alphas_cumprod) ** 0.5
|
69 |
+
return torch.flip(sigmas, (0,))
|
sgm/modules/diffusionmodules/guiders.py
ADDED
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from functools import partial
|
2 |
+
|
3 |
+
import torch
|
4 |
+
|
5 |
+
from ...util import default, instantiate_from_config
|
6 |
+
|
7 |
+
|
8 |
+
class VanillaCFG:
|
9 |
+
"""
|
10 |
+
implements parallelized CFG
|
11 |
+
"""
|
12 |
+
|
13 |
+
def __init__(self, scale, dyn_thresh_config=None):
|
14 |
+
scale_schedule = lambda scale, sigma: scale # independent of step
|
15 |
+
self.scale_schedule = partial(scale_schedule, scale)
|
16 |
+
self.dyn_thresh = instantiate_from_config(
|
17 |
+
default(
|
18 |
+
dyn_thresh_config,
|
19 |
+
{
|
20 |
+
"target": "sgm.modules.diffusionmodules.sampling_utils.NoDynamicThresholding"
|
21 |
+
},
|
22 |
+
)
|
23 |
+
)
|
24 |
+
|
25 |
+
def __call__(self, x, sigma):
|
26 |
+
x_u, x_c = x.chunk(2)
|
27 |
+
scale_value = self.scale_schedule(sigma)
|
28 |
+
x_pred = self.dyn_thresh(x_u, x_c, scale_value)
|
29 |
+
return x_pred
|
30 |
+
|
31 |
+
def prepare_inputs(self, x, s, c, uc):
|
32 |
+
c_out = dict()
|
33 |
+
|
34 |
+
for k in c:
|
35 |
+
if k in ["vector", "crossattn", "concat", "control", 'control_vector', 'mask_x']:
|
36 |
+
c_out[k] = torch.cat((uc[k], c[k]), 0)
|
37 |
+
else:
|
38 |
+
assert c[k] == uc[k]
|
39 |
+
c_out[k] = c[k]
|
40 |
+
return torch.cat([x] * 2), torch.cat([s] * 2), c_out
|
41 |
+
|
42 |
+
|
43 |
+
|
44 |
+
class LinearCFG:
|
45 |
+
def __init__(self, scale, scale_min=None, dyn_thresh_config=None):
|
46 |
+
if scale_min is None:
|
47 |
+
scale_min = scale
|
48 |
+
scale_schedule = lambda scale, scale_min, sigma: (scale - scale_min) * sigma / 14.6146 + scale_min
|
49 |
+
self.scale_schedule = partial(scale_schedule, scale, scale_min)
|
50 |
+
self.dyn_thresh = instantiate_from_config(
|
51 |
+
default(
|
52 |
+
dyn_thresh_config,
|
53 |
+
{
|
54 |
+
"target": "sgm.modules.diffusionmodules.sampling_utils.NoDynamicThresholding"
|
55 |
+
},
|
56 |
+
)
|
57 |
+
)
|
58 |
+
|
59 |
+
def __call__(self, x, sigma):
|
60 |
+
x_u, x_c = x.chunk(2)
|
61 |
+
scale_value = self.scale_schedule(sigma)
|
62 |
+
x_pred = self.dyn_thresh(x_u, x_c, scale_value)
|
63 |
+
return x_pred
|
64 |
+
|
65 |
+
def prepare_inputs(self, x, s, c, uc):
|
66 |
+
c_out = dict()
|
67 |
+
|
68 |
+
for k in c:
|
69 |
+
if k in ["vector", "crossattn", "concat", "control", 'control_vector', 'mask_x']:
|
70 |
+
c_out[k] = torch.cat((uc[k], c[k]), 0)
|
71 |
+
else:
|
72 |
+
assert c[k] == uc[k]
|
73 |
+
c_out[k] = c[k]
|
74 |
+
return torch.cat([x] * 2), torch.cat([s] * 2), c_out
|
75 |
+
|
76 |
+
|
77 |
+
|
78 |
+
class IdentityGuider:
|
79 |
+
def __call__(self, x, sigma):
|
80 |
+
return x
|
81 |
+
|
82 |
+
def prepare_inputs(self, x, s, c, uc):
|
83 |
+
c_out = dict()
|
84 |
+
|
85 |
+
for k in c:
|
86 |
+
c_out[k] = c[k]
|
87 |
+
|
88 |
+
return x, s, c_out
|
sgm/modules/diffusionmodules/loss.py
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Optional, Union
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
from omegaconf import ListConfig
|
6 |
+
|
7 |
+
from ...util import append_dims, instantiate_from_config
|
8 |
+
from ...modules.autoencoding.lpips.loss.lpips import LPIPS
|
9 |
+
|
10 |
+
|
11 |
+
class StandardDiffusionLoss(nn.Module):
|
12 |
+
def __init__(
|
13 |
+
self,
|
14 |
+
sigma_sampler_config,
|
15 |
+
type="l2",
|
16 |
+
offset_noise_level=0.0,
|
17 |
+
batch2model_keys: Optional[Union[str, List[str], ListConfig]] = None,
|
18 |
+
):
|
19 |
+
super().__init__()
|
20 |
+
|
21 |
+
assert type in ["l2", "l1", "lpips"]
|
22 |
+
|
23 |
+
self.sigma_sampler = instantiate_from_config(sigma_sampler_config)
|
24 |
+
|
25 |
+
self.type = type
|
26 |
+
self.offset_noise_level = offset_noise_level
|
27 |
+
|
28 |
+
if type == "lpips":
|
29 |
+
self.lpips = LPIPS().eval()
|
30 |
+
|
31 |
+
if not batch2model_keys:
|
32 |
+
batch2model_keys = []
|
33 |
+
|
34 |
+
if isinstance(batch2model_keys, str):
|
35 |
+
batch2model_keys = [batch2model_keys]
|
36 |
+
|
37 |
+
self.batch2model_keys = set(batch2model_keys)
|
38 |
+
|
39 |
+
def __call__(self, network, denoiser, conditioner, input, batch):
|
40 |
+
cond = conditioner(batch)
|
41 |
+
additional_model_inputs = {
|
42 |
+
key: batch[key] for key in self.batch2model_keys.intersection(batch)
|
43 |
+
}
|
44 |
+
|
45 |
+
sigmas = self.sigma_sampler(input.shape[0]).to(input.device)
|
46 |
+
noise = torch.randn_like(input)
|
47 |
+
if self.offset_noise_level > 0.0:
|
48 |
+
noise = noise + self.offset_noise_level * append_dims(
|
49 |
+
torch.randn(input.shape[0], device=input.device), input.ndim
|
50 |
+
)
|
51 |
+
noised_input = input + noise * append_dims(sigmas, input.ndim)
|
52 |
+
model_output = denoiser(
|
53 |
+
network, noised_input, sigmas, cond, **additional_model_inputs
|
54 |
+
)
|
55 |
+
w = append_dims(denoiser.w(sigmas), input.ndim)
|
56 |
+
return self.get_loss(model_output, input, w)
|
57 |
+
|
58 |
+
def get_loss(self, model_output, target, w):
|
59 |
+
if self.type == "l2":
|
60 |
+
return torch.mean(
|
61 |
+
(w * (model_output - target) ** 2).reshape(target.shape[0], -1), 1
|
62 |
+
)
|
63 |
+
elif self.type == "l1":
|
64 |
+
return torch.mean(
|
65 |
+
(w * (model_output - target).abs()).reshape(target.shape[0], -1), 1
|
66 |
+
)
|
67 |
+
elif self.type == "lpips":
|
68 |
+
loss = self.lpips(model_output, target).reshape(-1)
|
69 |
+
return loss
|
sgm/modules/diffusionmodules/model.py
ADDED
@@ -0,0 +1,743 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# pytorch_diffusion + derived encoder decoder
|
2 |
+
import math
|
3 |
+
from typing import Any, Callable, Optional
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
import torch.nn as nn
|
8 |
+
from einops import rearrange
|
9 |
+
from packaging import version
|
10 |
+
|
11 |
+
try:
|
12 |
+
import xformers
|
13 |
+
import xformers.ops
|
14 |
+
|
15 |
+
XFORMERS_IS_AVAILABLE = True
|
16 |
+
except:
|
17 |
+
XFORMERS_IS_AVAILABLE = False
|
18 |
+
print("no module 'xformers'. Processing without...")
|
19 |
+
|
20 |
+
from ...modules.attention import LinearAttention, MemoryEfficientCrossAttention
|
21 |
+
|
22 |
+
|
23 |
+
def get_timestep_embedding(timesteps, embedding_dim):
|
24 |
+
"""
|
25 |
+
This matches the implementation in Denoising Diffusion Probabilistic Models:
|
26 |
+
From Fairseq.
|
27 |
+
Build sinusoidal embeddings.
|
28 |
+
This matches the implementation in tensor2tensor, but differs slightly
|
29 |
+
from the description in Section 3.5 of "Attention Is All You Need".
|
30 |
+
"""
|
31 |
+
assert len(timesteps.shape) == 1
|
32 |
+
|
33 |
+
half_dim = embedding_dim // 2
|
34 |
+
emb = math.log(10000) / (half_dim - 1)
|
35 |
+
emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
|
36 |
+
emb = emb.to(device=timesteps.device)
|
37 |
+
emb = timesteps.float()[:, None] * emb[None, :]
|
38 |
+
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
|
39 |
+
if embedding_dim % 2 == 1: # zero pad
|
40 |
+
emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
|
41 |
+
return emb
|
42 |
+
|
43 |
+
|
44 |
+
def nonlinearity(x):
|
45 |
+
# swish
|
46 |
+
return x * torch.sigmoid(x)
|
47 |
+
|
48 |
+
|
49 |
+
def Normalize(in_channels, num_groups=32):
|
50 |
+
return torch.nn.GroupNorm(
|
51 |
+
num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
|
52 |
+
)
|
53 |
+
|
54 |
+
|
55 |
+
class Upsample(nn.Module):
|
56 |
+
def __init__(self, in_channels, with_conv):
|
57 |
+
super().__init__()
|
58 |
+
self.with_conv = with_conv
|
59 |
+
if self.with_conv:
|
60 |
+
self.conv = torch.nn.Conv2d(
|
61 |
+
in_channels, in_channels, kernel_size=3, stride=1, padding=1
|
62 |
+
)
|
63 |
+
|
64 |
+
def forward(self, x):
|
65 |
+
x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
|
66 |
+
if self.with_conv:
|
67 |
+
x = self.conv(x)
|
68 |
+
return x
|
69 |
+
|
70 |
+
|
71 |
+
class Downsample(nn.Module):
|
72 |
+
def __init__(self, in_channels, with_conv):
|
73 |
+
super().__init__()
|
74 |
+
self.with_conv = with_conv
|
75 |
+
if self.with_conv:
|
76 |
+
# no asymmetric padding in torch conv, must do it ourselves
|
77 |
+
self.conv = torch.nn.Conv2d(
|
78 |
+
in_channels, in_channels, kernel_size=3, stride=2, padding=0
|
79 |
+
)
|
80 |
+
|
81 |
+
def forward(self, x):
|
82 |
+
if self.with_conv:
|
83 |
+
pad = (0, 1, 0, 1)
|
84 |
+
x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
|
85 |
+
x = self.conv(x)
|
86 |
+
else:
|
87 |
+
x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
|
88 |
+
return x
|
89 |
+
|
90 |
+
|
91 |
+
class ResnetBlock(nn.Module):
|
92 |
+
def __init__(
|
93 |
+
self,
|
94 |
+
*,
|
95 |
+
in_channels,
|
96 |
+
out_channels=None,
|
97 |
+
conv_shortcut=False,
|
98 |
+
dropout,
|
99 |
+
temb_channels=512,
|
100 |
+
):
|
101 |
+
super().__init__()
|
102 |
+
self.in_channels = in_channels
|
103 |
+
out_channels = in_channels if out_channels is None else out_channels
|
104 |
+
self.out_channels = out_channels
|
105 |
+
self.use_conv_shortcut = conv_shortcut
|
106 |
+
|
107 |
+
self.norm1 = Normalize(in_channels)
|
108 |
+
self.conv1 = torch.nn.Conv2d(
|
109 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
110 |
+
)
|
111 |
+
if temb_channels > 0:
|
112 |
+
self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
|
113 |
+
self.norm2 = Normalize(out_channels)
|
114 |
+
self.dropout = torch.nn.Dropout(dropout)
|
115 |
+
self.conv2 = torch.nn.Conv2d(
|
116 |
+
out_channels, out_channels, kernel_size=3, stride=1, padding=1
|
117 |
+
)
|
118 |
+
if self.in_channels != self.out_channels:
|
119 |
+
if self.use_conv_shortcut:
|
120 |
+
self.conv_shortcut = torch.nn.Conv2d(
|
121 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
122 |
+
)
|
123 |
+
else:
|
124 |
+
self.nin_shortcut = torch.nn.Conv2d(
|
125 |
+
in_channels, out_channels, kernel_size=1, stride=1, padding=0
|
126 |
+
)
|
127 |
+
|
128 |
+
def forward(self, x, temb):
|
129 |
+
h = x
|
130 |
+
h = self.norm1(h)
|
131 |
+
h = nonlinearity(h)
|
132 |
+
h = self.conv1(h)
|
133 |
+
|
134 |
+
if temb is not None:
|
135 |
+
h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
|
136 |
+
|
137 |
+
h = self.norm2(h)
|
138 |
+
h = nonlinearity(h)
|
139 |
+
h = self.dropout(h)
|
140 |
+
h = self.conv2(h)
|
141 |
+
|
142 |
+
if self.in_channels != self.out_channels:
|
143 |
+
if self.use_conv_shortcut:
|
144 |
+
x = self.conv_shortcut(x)
|
145 |
+
else:
|
146 |
+
x = self.nin_shortcut(x)
|
147 |
+
|
148 |
+
return x + h
|
149 |
+
|
150 |
+
|
151 |
+
class LinAttnBlock(LinearAttention):
|
152 |
+
"""to match AttnBlock usage"""
|
153 |
+
|
154 |
+
def __init__(self, in_channels):
|
155 |
+
super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
|
156 |
+
|
157 |
+
|
158 |
+
class AttnBlock(nn.Module):
|
159 |
+
def __init__(self, in_channels):
|
160 |
+
super().__init__()
|
161 |
+
self.in_channels = in_channels
|
162 |
+
|
163 |
+
self.norm = Normalize(in_channels)
|
164 |
+
self.q = torch.nn.Conv2d(
|
165 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
166 |
+
)
|
167 |
+
self.k = torch.nn.Conv2d(
|
168 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
169 |
+
)
|
170 |
+
self.v = torch.nn.Conv2d(
|
171 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
172 |
+
)
|
173 |
+
self.proj_out = torch.nn.Conv2d(
|
174 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
175 |
+
)
|
176 |
+
|
177 |
+
def attention(self, h_: torch.Tensor) -> torch.Tensor:
|
178 |
+
h_ = self.norm(h_)
|
179 |
+
q = self.q(h_)
|
180 |
+
k = self.k(h_)
|
181 |
+
v = self.v(h_)
|
182 |
+
|
183 |
+
b, c, h, w = q.shape
|
184 |
+
q, k, v = map(
|
185 |
+
lambda x: rearrange(x, "b c h w -> b 1 (h w) c").contiguous(), (q, k, v)
|
186 |
+
)
|
187 |
+
h_ = torch.nn.functional.scaled_dot_product_attention(
|
188 |
+
q, k, v
|
189 |
+
) # scale is dim ** -0.5 per default
|
190 |
+
# compute attention
|
191 |
+
|
192 |
+
return rearrange(h_, "b 1 (h w) c -> b c h w", h=h, w=w, c=c, b=b)
|
193 |
+
|
194 |
+
def forward(self, x, **kwargs):
|
195 |
+
h_ = x
|
196 |
+
h_ = self.attention(h_)
|
197 |
+
h_ = self.proj_out(h_)
|
198 |
+
return x + h_
|
199 |
+
|
200 |
+
|
201 |
+
class MemoryEfficientAttnBlock(nn.Module):
|
202 |
+
"""
|
203 |
+
Uses xformers efficient implementation,
|
204 |
+
see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
|
205 |
+
Note: this is a single-head self-attention operation
|
206 |
+
"""
|
207 |
+
|
208 |
+
#
|
209 |
+
def __init__(self, in_channels):
|
210 |
+
super().__init__()
|
211 |
+
self.in_channels = in_channels
|
212 |
+
|
213 |
+
self.norm = Normalize(in_channels)
|
214 |
+
self.q = torch.nn.Conv2d(
|
215 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
216 |
+
)
|
217 |
+
self.k = torch.nn.Conv2d(
|
218 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
219 |
+
)
|
220 |
+
self.v = torch.nn.Conv2d(
|
221 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
222 |
+
)
|
223 |
+
self.proj_out = torch.nn.Conv2d(
|
224 |
+
in_channels, in_channels, kernel_size=1, stride=1, padding=0
|
225 |
+
)
|
226 |
+
self.attention_op: Optional[Any] = None
|
227 |
+
|
228 |
+
def attention(self, h_: torch.Tensor) -> torch.Tensor:
|
229 |
+
h_ = self.norm(h_)
|
230 |
+
q = self.q(h_)
|
231 |
+
k = self.k(h_)
|
232 |
+
v = self.v(h_)
|
233 |
+
|
234 |
+
# compute attention
|
235 |
+
B, C, H, W = q.shape
|
236 |
+
q, k, v = map(lambda x: rearrange(x, "b c h w -> b (h w) c"), (q, k, v))
|
237 |
+
|
238 |
+
q, k, v = map(
|
239 |
+
lambda t: t.unsqueeze(3)
|
240 |
+
.reshape(B, t.shape[1], 1, C)
|
241 |
+
.permute(0, 2, 1, 3)
|
242 |
+
.reshape(B * 1, t.shape[1], C)
|
243 |
+
.contiguous(),
|
244 |
+
(q, k, v),
|
245 |
+
)
|
246 |
+
out = xformers.ops.memory_efficient_attention(
|
247 |
+
q, k, v, attn_bias=None, op=self.attention_op
|
248 |
+
)
|
249 |
+
|
250 |
+
out = (
|
251 |
+
out.unsqueeze(0)
|
252 |
+
.reshape(B, 1, out.shape[1], C)
|
253 |
+
.permute(0, 2, 1, 3)
|
254 |
+
.reshape(B, out.shape[1], C)
|
255 |
+
)
|
256 |
+
return rearrange(out, "b (h w) c -> b c h w", b=B, h=H, w=W, c=C)
|
257 |
+
|
258 |
+
def forward(self, x, **kwargs):
|
259 |
+
h_ = x
|
260 |
+
h_ = self.attention(h_)
|
261 |
+
h_ = self.proj_out(h_)
|
262 |
+
return x + h_
|
263 |
+
|
264 |
+
|
265 |
+
class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention):
|
266 |
+
def forward(self, x, context=None, mask=None, **unused_kwargs):
|
267 |
+
b, c, h, w = x.shape
|
268 |
+
x = rearrange(x, "b c h w -> b (h w) c")
|
269 |
+
out = super().forward(x, context=context, mask=mask)
|
270 |
+
out = rearrange(out, "b (h w) c -> b c h w", h=h, w=w, c=c)
|
271 |
+
return x + out
|
272 |
+
|
273 |
+
|
274 |
+
def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
|
275 |
+
assert attn_type in [
|
276 |
+
"vanilla",
|
277 |
+
"vanilla-xformers",
|
278 |
+
"memory-efficient-cross-attn",
|
279 |
+
"linear",
|
280 |
+
"none",
|
281 |
+
], f"attn_type {attn_type} unknown"
|
282 |
+
if (
|
283 |
+
version.parse(torch.__version__) < version.parse("2.0.0")
|
284 |
+
and attn_type != "none"
|
285 |
+
):
|
286 |
+
assert XFORMERS_IS_AVAILABLE, (
|
287 |
+
f"We do not support vanilla attention in {torch.__version__} anymore, "
|
288 |
+
f"as it is too expensive. Please install xformers via e.g. 'pip install xformers==0.0.16'"
|
289 |
+
)
|
290 |
+
attn_type = "vanilla-xformers"
|
291 |
+
print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
|
292 |
+
if attn_type == "vanilla":
|
293 |
+
assert attn_kwargs is None
|
294 |
+
return AttnBlock(in_channels)
|
295 |
+
elif attn_type == "vanilla-xformers":
|
296 |
+
print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...")
|
297 |
+
return MemoryEfficientAttnBlock(in_channels)
|
298 |
+
elif type == "memory-efficient-cross-attn":
|
299 |
+
attn_kwargs["query_dim"] = in_channels
|
300 |
+
return MemoryEfficientCrossAttentionWrapper(**attn_kwargs)
|
301 |
+
elif attn_type == "none":
|
302 |
+
return nn.Identity(in_channels)
|
303 |
+
else:
|
304 |
+
return LinAttnBlock(in_channels)
|
305 |
+
|
306 |
+
|
307 |
+
class Model(nn.Module):
|
308 |
+
def __init__(
|
309 |
+
self,
|
310 |
+
*,
|
311 |
+
ch,
|
312 |
+
out_ch,
|
313 |
+
ch_mult=(1, 2, 4, 8),
|
314 |
+
num_res_blocks,
|
315 |
+
attn_resolutions,
|
316 |
+
dropout=0.0,
|
317 |
+
resamp_with_conv=True,
|
318 |
+
in_channels,
|
319 |
+
resolution,
|
320 |
+
use_timestep=True,
|
321 |
+
use_linear_attn=False,
|
322 |
+
attn_type="vanilla",
|
323 |
+
):
|
324 |
+
super().__init__()
|
325 |
+
if use_linear_attn:
|
326 |
+
attn_type = "linear"
|
327 |
+
self.ch = ch
|
328 |
+
self.temb_ch = self.ch * 4
|
329 |
+
self.num_resolutions = len(ch_mult)
|
330 |
+
self.num_res_blocks = num_res_blocks
|
331 |
+
self.resolution = resolution
|
332 |
+
self.in_channels = in_channels
|
333 |
+
|
334 |
+
self.use_timestep = use_timestep
|
335 |
+
if self.use_timestep:
|
336 |
+
# timestep embedding
|
337 |
+
self.temb = nn.Module()
|
338 |
+
self.temb.dense = nn.ModuleList(
|
339 |
+
[
|
340 |
+
torch.nn.Linear(self.ch, self.temb_ch),
|
341 |
+
torch.nn.Linear(self.temb_ch, self.temb_ch),
|
342 |
+
]
|
343 |
+
)
|
344 |
+
|
345 |
+
# downsampling
|
346 |
+
self.conv_in = torch.nn.Conv2d(
|
347 |
+
in_channels, self.ch, kernel_size=3, stride=1, padding=1
|
348 |
+
)
|
349 |
+
|
350 |
+
curr_res = resolution
|
351 |
+
in_ch_mult = (1,) + tuple(ch_mult)
|
352 |
+
self.down = nn.ModuleList()
|
353 |
+
for i_level in range(self.num_resolutions):
|
354 |
+
block = nn.ModuleList()
|
355 |
+
attn = nn.ModuleList()
|
356 |
+
block_in = ch * in_ch_mult[i_level]
|
357 |
+
block_out = ch * ch_mult[i_level]
|
358 |
+
for i_block in range(self.num_res_blocks):
|
359 |
+
block.append(
|
360 |
+
ResnetBlock(
|
361 |
+
in_channels=block_in,
|
362 |
+
out_channels=block_out,
|
363 |
+
temb_channels=self.temb_ch,
|
364 |
+
dropout=dropout,
|
365 |
+
)
|
366 |
+
)
|
367 |
+
block_in = block_out
|
368 |
+
if curr_res in attn_resolutions:
|
369 |
+
attn.append(make_attn(block_in, attn_type=attn_type))
|
370 |
+
down = nn.Module()
|
371 |
+
down.block = block
|
372 |
+
down.attn = attn
|
373 |
+
if i_level != self.num_resolutions - 1:
|
374 |
+
down.downsample = Downsample(block_in, resamp_with_conv)
|
375 |
+
curr_res = curr_res // 2
|
376 |
+
self.down.append(down)
|
377 |
+
|
378 |
+
# middle
|
379 |
+
self.mid = nn.Module()
|
380 |
+
self.mid.block_1 = ResnetBlock(
|
381 |
+
in_channels=block_in,
|
382 |
+
out_channels=block_in,
|
383 |
+
temb_channels=self.temb_ch,
|
384 |
+
dropout=dropout,
|
385 |
+
)
|
386 |
+
self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
|
387 |
+
self.mid.block_2 = ResnetBlock(
|
388 |
+
in_channels=block_in,
|
389 |
+
out_channels=block_in,
|
390 |
+
temb_channels=self.temb_ch,
|
391 |
+
dropout=dropout,
|
392 |
+
)
|
393 |
+
|
394 |
+
# upsampling
|
395 |
+
self.up = nn.ModuleList()
|
396 |
+
for i_level in reversed(range(self.num_resolutions)):
|
397 |
+
block = nn.ModuleList()
|
398 |
+
attn = nn.ModuleList()
|
399 |
+
block_out = ch * ch_mult[i_level]
|
400 |
+
skip_in = ch * ch_mult[i_level]
|
401 |
+
for i_block in range(self.num_res_blocks + 1):
|
402 |
+
if i_block == self.num_res_blocks:
|
403 |
+
skip_in = ch * in_ch_mult[i_level]
|
404 |
+
block.append(
|
405 |
+
ResnetBlock(
|
406 |
+
in_channels=block_in + skip_in,
|
407 |
+
out_channels=block_out,
|
408 |
+
temb_channels=self.temb_ch,
|
409 |
+
dropout=dropout,
|
410 |
+
)
|
411 |
+
)
|
412 |
+
block_in = block_out
|
413 |
+
if curr_res in attn_resolutions:
|
414 |
+
attn.append(make_attn(block_in, attn_type=attn_type))
|
415 |
+
up = nn.Module()
|
416 |
+
up.block = block
|
417 |
+
up.attn = attn
|
418 |
+
if i_level != 0:
|
419 |
+
up.upsample = Upsample(block_in, resamp_with_conv)
|
420 |
+
curr_res = curr_res * 2
|
421 |
+
self.up.insert(0, up) # prepend to get consistent order
|
422 |
+
|
423 |
+
# end
|
424 |
+
self.norm_out = Normalize(block_in)
|
425 |
+
self.conv_out = torch.nn.Conv2d(
|
426 |
+
block_in, out_ch, kernel_size=3, stride=1, padding=1
|
427 |
+
)
|
428 |
+
|
429 |
+
def forward(self, x, t=None, context=None):
|
430 |
+
# assert x.shape[2] == x.shape[3] == self.resolution
|
431 |
+
if context is not None:
|
432 |
+
# assume aligned context, cat along channel axis
|
433 |
+
x = torch.cat((x, context), dim=1)
|
434 |
+
if self.use_timestep:
|
435 |
+
# timestep embedding
|
436 |
+
assert t is not None
|
437 |
+
temb = get_timestep_embedding(t, self.ch)
|
438 |
+
temb = self.temb.dense[0](temb)
|
439 |
+
temb = nonlinearity(temb)
|
440 |
+
temb = self.temb.dense[1](temb)
|
441 |
+
else:
|
442 |
+
temb = None
|
443 |
+
|
444 |
+
# downsampling
|
445 |
+
hs = [self.conv_in(x)]
|
446 |
+
for i_level in range(self.num_resolutions):
|
447 |
+
for i_block in range(self.num_res_blocks):
|
448 |
+
h = self.down[i_level].block[i_block](hs[-1], temb)
|
449 |
+
if len(self.down[i_level].attn) > 0:
|
450 |
+
h = self.down[i_level].attn[i_block](h)
|
451 |
+
hs.append(h)
|
452 |
+
if i_level != self.num_resolutions - 1:
|
453 |
+
hs.append(self.down[i_level].downsample(hs[-1]))
|
454 |
+
|
455 |
+
# middle
|
456 |
+
h = hs[-1]
|
457 |
+
h = self.mid.block_1(h, temb)
|
458 |
+
h = self.mid.attn_1(h)
|
459 |
+
h = self.mid.block_2(h, temb)
|
460 |
+
|
461 |
+
# upsampling
|
462 |
+
for i_level in reversed(range(self.num_resolutions)):
|
463 |
+
for i_block in range(self.num_res_blocks + 1):
|
464 |
+
h = self.up[i_level].block[i_block](
|
465 |
+
torch.cat([h, hs.pop()], dim=1), temb
|
466 |
+
)
|
467 |
+
if len(self.up[i_level].attn) > 0:
|
468 |
+
h = self.up[i_level].attn[i_block](h)
|
469 |
+
if i_level != 0:
|
470 |
+
h = self.up[i_level].upsample(h)
|
471 |
+
|
472 |
+
# end
|
473 |
+
h = self.norm_out(h)
|
474 |
+
h = nonlinearity(h)
|
475 |
+
h = self.conv_out(h)
|
476 |
+
return h
|
477 |
+
|
478 |
+
def get_last_layer(self):
|
479 |
+
return self.conv_out.weight
|
480 |
+
|
481 |
+
|
482 |
+
class Encoder(nn.Module):
|
483 |
+
def __init__(
|
484 |
+
self,
|
485 |
+
*,
|
486 |
+
ch,
|
487 |
+
out_ch,
|
488 |
+
ch_mult=(1, 2, 4, 8),
|
489 |
+
num_res_blocks,
|
490 |
+
attn_resolutions,
|
491 |
+
dropout=0.0,
|
492 |
+
resamp_with_conv=True,
|
493 |
+
in_channels,
|
494 |
+
resolution,
|
495 |
+
z_channels,
|
496 |
+
double_z=True,
|
497 |
+
use_linear_attn=False,
|
498 |
+
attn_type="vanilla",
|
499 |
+
**ignore_kwargs,
|
500 |
+
):
|
501 |
+
super().__init__()
|
502 |
+
if use_linear_attn:
|
503 |
+
attn_type = "linear"
|
504 |
+
self.ch = ch
|
505 |
+
self.temb_ch = 0
|
506 |
+
self.num_resolutions = len(ch_mult)
|
507 |
+
self.num_res_blocks = num_res_blocks
|
508 |
+
self.resolution = resolution
|
509 |
+
self.in_channels = in_channels
|
510 |
+
|
511 |
+
# downsampling
|
512 |
+
self.conv_in = torch.nn.Conv2d(
|
513 |
+
in_channels, self.ch, kernel_size=3, stride=1, padding=1
|
514 |
+
)
|
515 |
+
|
516 |
+
curr_res = resolution
|
517 |
+
in_ch_mult = (1,) + tuple(ch_mult)
|
518 |
+
self.in_ch_mult = in_ch_mult
|
519 |
+
self.down = nn.ModuleList()
|
520 |
+
for i_level in range(self.num_resolutions):
|
521 |
+
block = nn.ModuleList()
|
522 |
+
attn = nn.ModuleList()
|
523 |
+
block_in = ch * in_ch_mult[i_level]
|
524 |
+
block_out = ch * ch_mult[i_level]
|
525 |
+
for i_block in range(self.num_res_blocks):
|
526 |
+
block.append(
|
527 |
+
ResnetBlock(
|
528 |
+
in_channels=block_in,
|
529 |
+
out_channels=block_out,
|
530 |
+
temb_channels=self.temb_ch,
|
531 |
+
dropout=dropout,
|
532 |
+
)
|
533 |
+
)
|
534 |
+
block_in = block_out
|
535 |
+
if curr_res in attn_resolutions:
|
536 |
+
attn.append(make_attn(block_in, attn_type=attn_type))
|
537 |
+
down = nn.Module()
|
538 |
+
down.block = block
|
539 |
+
down.attn = attn
|
540 |
+
if i_level != self.num_resolutions - 1:
|
541 |
+
down.downsample = Downsample(block_in, resamp_with_conv)
|
542 |
+
curr_res = curr_res // 2
|
543 |
+
self.down.append(down)
|
544 |
+
|
545 |
+
# middle
|
546 |
+
self.mid = nn.Module()
|
547 |
+
self.mid.block_1 = ResnetBlock(
|
548 |
+
in_channels=block_in,
|
549 |
+
out_channels=block_in,
|
550 |
+
temb_channels=self.temb_ch,
|
551 |
+
dropout=dropout,
|
552 |
+
)
|
553 |
+
self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
|
554 |
+
self.mid.block_2 = ResnetBlock(
|
555 |
+
in_channels=block_in,
|
556 |
+
out_channels=block_in,
|
557 |
+
temb_channels=self.temb_ch,
|
558 |
+
dropout=dropout,
|
559 |
+
)
|
560 |
+
|
561 |
+
# end
|
562 |
+
self.norm_out = Normalize(block_in)
|
563 |
+
self.conv_out = torch.nn.Conv2d(
|
564 |
+
block_in,
|
565 |
+
2 * z_channels if double_z else z_channels,
|
566 |
+
kernel_size=3,
|
567 |
+
stride=1,
|
568 |
+
padding=1,
|
569 |
+
)
|
570 |
+
|
571 |
+
def forward(self, x):
|
572 |
+
# timestep embedding
|
573 |
+
temb = None
|
574 |
+
|
575 |
+
# downsampling
|
576 |
+
hs = [self.conv_in(x)]
|
577 |
+
for i_level in range(self.num_resolutions):
|
578 |
+
for i_block in range(self.num_res_blocks):
|
579 |
+
h = self.down[i_level].block[i_block](hs[-1], temb)
|
580 |
+
if len(self.down[i_level].attn) > 0:
|
581 |
+
h = self.down[i_level].attn[i_block](h)
|
582 |
+
hs.append(h)
|
583 |
+
if i_level != self.num_resolutions - 1:
|
584 |
+
hs.append(self.down[i_level].downsample(hs[-1]))
|
585 |
+
|
586 |
+
# middle
|
587 |
+
h = hs[-1]
|
588 |
+
h = self.mid.block_1(h, temb)
|
589 |
+
h = self.mid.attn_1(h)
|
590 |
+
h = self.mid.block_2(h, temb)
|
591 |
+
|
592 |
+
# end
|
593 |
+
h = self.norm_out(h)
|
594 |
+
h = nonlinearity(h)
|
595 |
+
h = self.conv_out(h)
|
596 |
+
return h
|
597 |
+
|
598 |
+
|
599 |
+
class Decoder(nn.Module):
|
600 |
+
def __init__(
|
601 |
+
self,
|
602 |
+
*,
|
603 |
+
ch,
|
604 |
+
out_ch,
|
605 |
+
ch_mult=(1, 2, 4, 8),
|
606 |
+
num_res_blocks,
|
607 |
+
attn_resolutions,
|
608 |
+
dropout=0.0,
|
609 |
+
resamp_with_conv=True,
|
610 |
+
in_channels,
|
611 |
+
resolution,
|
612 |
+
z_channels,
|
613 |
+
give_pre_end=False,
|
614 |
+
tanh_out=False,
|
615 |
+
use_linear_attn=False,
|
616 |
+
attn_type="vanilla",
|
617 |
+
**ignorekwargs,
|
618 |
+
):
|
619 |
+
super().__init__()
|
620 |
+
if use_linear_attn:
|
621 |
+
attn_type = "linear"
|
622 |
+
self.ch = ch
|
623 |
+
self.temb_ch = 0
|
624 |
+
self.num_resolutions = len(ch_mult)
|
625 |
+
self.num_res_blocks = num_res_blocks
|
626 |
+
self.resolution = resolution
|
627 |
+
self.in_channels = in_channels
|
628 |
+
self.give_pre_end = give_pre_end
|
629 |
+
self.tanh_out = tanh_out
|
630 |
+
|
631 |
+
# compute in_ch_mult, block_in and curr_res at lowest res
|
632 |
+
in_ch_mult = (1,) + tuple(ch_mult)
|
633 |
+
block_in = ch * ch_mult[self.num_resolutions - 1]
|
634 |
+
curr_res = resolution // 2 ** (self.num_resolutions - 1)
|
635 |
+
self.z_shape = (1, z_channels, curr_res, curr_res)
|
636 |
+
print(
|
637 |
+
"Working with z of shape {} = {} dimensions.".format(
|
638 |
+
self.z_shape, np.prod(self.z_shape)
|
639 |
+
)
|
640 |
+
)
|
641 |
+
|
642 |
+
make_attn_cls = self._make_attn()
|
643 |
+
make_resblock_cls = self._make_resblock()
|
644 |
+
make_conv_cls = self._make_conv()
|
645 |
+
# z to block_in
|
646 |
+
self.conv_in = torch.nn.Conv2d(
|
647 |
+
z_channels, block_in, kernel_size=3, stride=1, padding=1
|
648 |
+
)
|
649 |
+
|
650 |
+
# middle
|
651 |
+
self.mid = nn.Module()
|
652 |
+
self.mid.block_1 = make_resblock_cls(
|
653 |
+
in_channels=block_in,
|
654 |
+
out_channels=block_in,
|
655 |
+
temb_channels=self.temb_ch,
|
656 |
+
dropout=dropout,
|
657 |
+
)
|
658 |
+
self.mid.attn_1 = make_attn_cls(block_in, attn_type=attn_type)
|
659 |
+
self.mid.block_2 = make_resblock_cls(
|
660 |
+
in_channels=block_in,
|
661 |
+
out_channels=block_in,
|
662 |
+
temb_channels=self.temb_ch,
|
663 |
+
dropout=dropout,
|
664 |
+
)
|
665 |
+
|
666 |
+
# upsampling
|
667 |
+
self.up = nn.ModuleList()
|
668 |
+
for i_level in reversed(range(self.num_resolutions)):
|
669 |
+
block = nn.ModuleList()
|
670 |
+
attn = nn.ModuleList()
|
671 |
+
block_out = ch * ch_mult[i_level]
|
672 |
+
for i_block in range(self.num_res_blocks + 1):
|
673 |
+
block.append(
|
674 |
+
make_resblock_cls(
|
675 |
+
in_channels=block_in,
|
676 |
+
out_channels=block_out,
|
677 |
+
temb_channels=self.temb_ch,
|
678 |
+
dropout=dropout,
|
679 |
+
)
|
680 |
+
)
|
681 |
+
block_in = block_out
|
682 |
+
if curr_res in attn_resolutions:
|
683 |
+
attn.append(make_attn_cls(block_in, attn_type=attn_type))
|
684 |
+
up = nn.Module()
|
685 |
+
up.block = block
|
686 |
+
up.attn = attn
|
687 |
+
if i_level != 0:
|
688 |
+
up.upsample = Upsample(block_in, resamp_with_conv)
|
689 |
+
curr_res = curr_res * 2
|
690 |
+
self.up.insert(0, up) # prepend to get consistent order
|
691 |
+
|
692 |
+
# end
|
693 |
+
self.norm_out = Normalize(block_in)
|
694 |
+
self.conv_out = make_conv_cls(
|
695 |
+
block_in, out_ch, kernel_size=3, stride=1, padding=1
|
696 |
+
)
|
697 |
+
|
698 |
+
def _make_attn(self) -> Callable:
|
699 |
+
return make_attn
|
700 |
+
|
701 |
+
def _make_resblock(self) -> Callable:
|
702 |
+
return ResnetBlock
|
703 |
+
|
704 |
+
def _make_conv(self) -> Callable:
|
705 |
+
return torch.nn.Conv2d
|
706 |
+
|
707 |
+
def get_last_layer(self, **kwargs):
|
708 |
+
return self.conv_out.weight
|
709 |
+
|
710 |
+
def forward(self, z, **kwargs):
|
711 |
+
# assert z.shape[1:] == self.z_shape[1:]
|
712 |
+
self.last_z_shape = z.shape
|
713 |
+
|
714 |
+
# timestep embedding
|
715 |
+
temb = None
|
716 |
+
|
717 |
+
# z to block_in
|
718 |
+
h = self.conv_in(z)
|
719 |
+
|
720 |
+
# middle
|
721 |
+
h = self.mid.block_1(h, temb, **kwargs)
|
722 |
+
h = self.mid.attn_1(h, **kwargs)
|
723 |
+
h = self.mid.block_2(h, temb, **kwargs)
|
724 |
+
|
725 |
+
# upsampling
|
726 |
+
for i_level in reversed(range(self.num_resolutions)):
|
727 |
+
for i_block in range(self.num_res_blocks + 1):
|
728 |
+
h = self.up[i_level].block[i_block](h, temb, **kwargs)
|
729 |
+
if len(self.up[i_level].attn) > 0:
|
730 |
+
h = self.up[i_level].attn[i_block](h, **kwargs)
|
731 |
+
if i_level != 0:
|
732 |
+
h = self.up[i_level].upsample(h)
|
733 |
+
|
734 |
+
# end
|
735 |
+
if self.give_pre_end:
|
736 |
+
return h
|
737 |
+
|
738 |
+
h = self.norm_out(h)
|
739 |
+
h = nonlinearity(h)
|
740 |
+
h = self.conv_out(h, **kwargs)
|
741 |
+
if self.tanh_out:
|
742 |
+
h = torch.tanh(h)
|
743 |
+
return h
|
sgm/modules/diffusionmodules/openaimodel.py
ADDED
@@ -0,0 +1,1272 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from abc import abstractmethod
|
3 |
+
from functools import partial
|
4 |
+
from typing import Iterable
|
5 |
+
|
6 |
+
import numpy as np
|
7 |
+
import torch as th
|
8 |
+
import torch.nn as nn
|
9 |
+
import torch.nn.functional as F
|
10 |
+
# from einops._torch_specific import allow_ops_in_compiled_graph
|
11 |
+
# allow_ops_in_compiled_graph()
|
12 |
+
from einops import rearrange
|
13 |
+
|
14 |
+
from ...modules.attention import SpatialTransformer
|
15 |
+
from ...modules.diffusionmodules.util import (
|
16 |
+
avg_pool_nd,
|
17 |
+
checkpoint,
|
18 |
+
conv_nd,
|
19 |
+
linear,
|
20 |
+
normalization,
|
21 |
+
timestep_embedding,
|
22 |
+
zero_module,
|
23 |
+
)
|
24 |
+
from ...util import default, exists
|
25 |
+
|
26 |
+
|
27 |
+
# dummy replace
|
28 |
+
def convert_module_to_f16(x):
|
29 |
+
pass
|
30 |
+
|
31 |
+
|
32 |
+
def convert_module_to_f32(x):
|
33 |
+
pass
|
34 |
+
|
35 |
+
|
36 |
+
## go
|
37 |
+
class AttentionPool2d(nn.Module):
|
38 |
+
"""
|
39 |
+
Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
|
40 |
+
"""
|
41 |
+
|
42 |
+
def __init__(
|
43 |
+
self,
|
44 |
+
spacial_dim: int,
|
45 |
+
embed_dim: int,
|
46 |
+
num_heads_channels: int,
|
47 |
+
output_dim: int = None,
|
48 |
+
):
|
49 |
+
super().__init__()
|
50 |
+
self.positional_embedding = nn.Parameter(
|
51 |
+
th.randn(embed_dim, spacial_dim**2 + 1) / embed_dim**0.5
|
52 |
+
)
|
53 |
+
self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
|
54 |
+
self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
|
55 |
+
self.num_heads = embed_dim // num_heads_channels
|
56 |
+
self.attention = QKVAttention(self.num_heads)
|
57 |
+
|
58 |
+
def forward(self, x):
|
59 |
+
b, c, *_spatial = x.shape
|
60 |
+
x = x.reshape(b, c, -1) # NC(HW)
|
61 |
+
x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1)
|
62 |
+
x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1)
|
63 |
+
x = self.qkv_proj(x)
|
64 |
+
x = self.attention(x)
|
65 |
+
x = self.c_proj(x)
|
66 |
+
return x[:, :, 0]
|
67 |
+
|
68 |
+
|
69 |
+
class TimestepBlock(nn.Module):
|
70 |
+
"""
|
71 |
+
Any module where forward() takes timestep embeddings as a second argument.
|
72 |
+
"""
|
73 |
+
|
74 |
+
@abstractmethod
|
75 |
+
def forward(self, x, emb):
|
76 |
+
"""
|
77 |
+
Apply the module to `x` given `emb` timestep embeddings.
|
78 |
+
"""
|
79 |
+
|
80 |
+
|
81 |
+
class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
|
82 |
+
"""
|
83 |
+
A sequential module that passes timestep embeddings to the children that
|
84 |
+
support it as an extra input.
|
85 |
+
"""
|
86 |
+
|
87 |
+
def forward(
|
88 |
+
self,
|
89 |
+
x,
|
90 |
+
emb,
|
91 |
+
context=None,
|
92 |
+
skip_time_mix=False,
|
93 |
+
time_context=None,
|
94 |
+
num_video_frames=None,
|
95 |
+
time_context_cat=None,
|
96 |
+
use_crossframe_attention_in_spatial_layers=False,
|
97 |
+
):
|
98 |
+
for layer in self:
|
99 |
+
if isinstance(layer, TimestepBlock):
|
100 |
+
x = layer(x, emb)
|
101 |
+
elif isinstance(layer, SpatialTransformer):
|
102 |
+
x = layer(x, context)
|
103 |
+
else:
|
104 |
+
x = layer(x)
|
105 |
+
return x
|
106 |
+
|
107 |
+
|
108 |
+
class Upsample(nn.Module):
|
109 |
+
"""
|
110 |
+
An upsampling layer with an optional convolution.
|
111 |
+
:param channels: channels in the inputs and outputs.
|
112 |
+
:param use_conv: a bool determining if a convolution is applied.
|
113 |
+
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
|
114 |
+
upsampling occurs in the inner-two dimensions.
|
115 |
+
"""
|
116 |
+
|
117 |
+
def __init__(
|
118 |
+
self, channels, use_conv, dims=2, out_channels=None, padding=1, third_up=False
|
119 |
+
):
|
120 |
+
super().__init__()
|
121 |
+
self.channels = channels
|
122 |
+
self.out_channels = out_channels or channels
|
123 |
+
self.use_conv = use_conv
|
124 |
+
self.dims = dims
|
125 |
+
self.third_up = third_up
|
126 |
+
if use_conv:
|
127 |
+
self.conv = conv_nd(
|
128 |
+
dims, self.channels, self.out_channels, 3, padding=padding
|
129 |
+
)
|
130 |
+
|
131 |
+
def forward(self, x):
|
132 |
+
# support fp32 only
|
133 |
+
_dtype = x.dtype
|
134 |
+
x = x.to(th.float32)
|
135 |
+
|
136 |
+
assert x.shape[1] == self.channels
|
137 |
+
if self.dims == 3:
|
138 |
+
t_factor = 1 if not self.third_up else 2
|
139 |
+
x = F.interpolate(
|
140 |
+
x,
|
141 |
+
(t_factor * x.shape[2], x.shape[3] * 2, x.shape[4] * 2),
|
142 |
+
mode="nearest",
|
143 |
+
)
|
144 |
+
else:
|
145 |
+
x = F.interpolate(x, scale_factor=2, mode="nearest")
|
146 |
+
|
147 |
+
x = x.to(_dtype) # support fp32 only
|
148 |
+
|
149 |
+
if self.use_conv:
|
150 |
+
x = self.conv(x)
|
151 |
+
return x
|
152 |
+
|
153 |
+
|
154 |
+
class TransposedUpsample(nn.Module):
|
155 |
+
"Learned 2x upsampling without padding"
|
156 |
+
|
157 |
+
def __init__(self, channels, out_channels=None, ks=5):
|
158 |
+
super().__init__()
|
159 |
+
self.channels = channels
|
160 |
+
self.out_channels = out_channels or channels
|
161 |
+
|
162 |
+
self.up = nn.ConvTranspose2d(
|
163 |
+
self.channels, self.out_channels, kernel_size=ks, stride=2
|
164 |
+
)
|
165 |
+
|
166 |
+
def forward(self, x):
|
167 |
+
return self.up(x)
|
168 |
+
|
169 |
+
|
170 |
+
class Downsample(nn.Module):
|
171 |
+
"""
|
172 |
+
A downsampling layer with an optional convolution.
|
173 |
+
:param channels: channels in the inputs and outputs.
|
174 |
+
:param use_conv: a bool determining if a convolution is applied.
|
175 |
+
:param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
|
176 |
+
downsampling occurs in the inner-two dimensions.
|
177 |
+
"""
|
178 |
+
|
179 |
+
def __init__(
|
180 |
+
self, channels, use_conv, dims=2, out_channels=None, padding=1, third_down=False
|
181 |
+
):
|
182 |
+
super().__init__()
|
183 |
+
self.channels = channels
|
184 |
+
self.out_channels = out_channels or channels
|
185 |
+
self.use_conv = use_conv
|
186 |
+
self.dims = dims
|
187 |
+
stride = 2 if dims != 3 else ((1, 2, 2) if not third_down else (2, 2, 2))
|
188 |
+
if use_conv:
|
189 |
+
print(f"Building a Downsample layer with {dims} dims.")
|
190 |
+
print(
|
191 |
+
f" --> settings are: \n in-chn: {self.channels}, out-chn: {self.out_channels}, "
|
192 |
+
f"kernel-size: 3, stride: {stride}, padding: {padding}"
|
193 |
+
)
|
194 |
+
if dims == 3:
|
195 |
+
print(f" --> Downsampling third axis (time): {third_down}")
|
196 |
+
self.op = conv_nd(
|
197 |
+
dims,
|
198 |
+
self.channels,
|
199 |
+
self.out_channels,
|
200 |
+
3,
|
201 |
+
stride=stride,
|
202 |
+
padding=padding,
|
203 |
+
)
|
204 |
+
else:
|
205 |
+
assert self.channels == self.out_channels
|
206 |
+
self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
|
207 |
+
|
208 |
+
def forward(self, x):
|
209 |
+
assert x.shape[1] == self.channels
|
210 |
+
return self.op(x)
|
211 |
+
|
212 |
+
|
213 |
+
class ResBlock(TimestepBlock):
|
214 |
+
"""
|
215 |
+
A residual block that can optionally change the number of channels.
|
216 |
+
:param channels: the number of input channels.
|
217 |
+
:param emb_channels: the number of timestep embedding channels.
|
218 |
+
:param dropout: the rate of dropout.
|
219 |
+
:param out_channels: if specified, the number of out channels.
|
220 |
+
:param use_conv: if True and out_channels is specified, use a spatial
|
221 |
+
convolution instead of a smaller 1x1 convolution to change the
|
222 |
+
channels in the skip connection.
|
223 |
+
:param dims: determines if the signal is 1D, 2D, or 3D.
|
224 |
+
:param use_checkpoint: if True, use gradient checkpointing on this module.
|
225 |
+
:param up: if True, use this block for upsampling.
|
226 |
+
:param down: if True, use this block for downsampling.
|
227 |
+
"""
|
228 |
+
|
229 |
+
def __init__(
|
230 |
+
self,
|
231 |
+
channels,
|
232 |
+
emb_channels,
|
233 |
+
dropout,
|
234 |
+
out_channels=None,
|
235 |
+
use_conv=False,
|
236 |
+
use_scale_shift_norm=False,
|
237 |
+
dims=2,
|
238 |
+
use_checkpoint=False,
|
239 |
+
up=False,
|
240 |
+
down=False,
|
241 |
+
kernel_size=3,
|
242 |
+
exchange_temb_dims=False,
|
243 |
+
skip_t_emb=False,
|
244 |
+
):
|
245 |
+
super().__init__()
|
246 |
+
self.channels = channels
|
247 |
+
self.emb_channels = emb_channels
|
248 |
+
self.dropout = dropout
|
249 |
+
self.out_channels = out_channels or channels
|
250 |
+
self.use_conv = use_conv
|
251 |
+
self.use_checkpoint = use_checkpoint
|
252 |
+
self.use_scale_shift_norm = use_scale_shift_norm
|
253 |
+
self.exchange_temb_dims = exchange_temb_dims
|
254 |
+
|
255 |
+
if isinstance(kernel_size, Iterable):
|
256 |
+
padding = [k // 2 for k in kernel_size]
|
257 |
+
else:
|
258 |
+
padding = kernel_size // 2
|
259 |
+
|
260 |
+
self.in_layers = nn.Sequential(
|
261 |
+
normalization(channels),
|
262 |
+
nn.SiLU(),
|
263 |
+
conv_nd(dims, channels, self.out_channels, kernel_size, padding=padding),
|
264 |
+
)
|
265 |
+
|
266 |
+
self.updown = up or down
|
267 |
+
|
268 |
+
if up:
|
269 |
+
self.h_upd = Upsample(channels, False, dims)
|
270 |
+
self.x_upd = Upsample(channels, False, dims)
|
271 |
+
elif down:
|
272 |
+
self.h_upd = Downsample(channels, False, dims)
|
273 |
+
self.x_upd = Downsample(channels, False, dims)
|
274 |
+
else:
|
275 |
+
self.h_upd = self.x_upd = nn.Identity()
|
276 |
+
|
277 |
+
self.skip_t_emb = skip_t_emb
|
278 |
+
self.emb_out_channels = (
|
279 |
+
2 * self.out_channels if use_scale_shift_norm else self.out_channels
|
280 |
+
)
|
281 |
+
if self.skip_t_emb:
|
282 |
+
print(f"Skipping timestep embedding in {self.__class__.__name__}")
|
283 |
+
assert not self.use_scale_shift_norm
|
284 |
+
self.emb_layers = None
|
285 |
+
self.exchange_temb_dims = False
|
286 |
+
else:
|
287 |
+
self.emb_layers = nn.Sequential(
|
288 |
+
nn.SiLU(),
|
289 |
+
linear(
|
290 |
+
emb_channels,
|
291 |
+
self.emb_out_channels,
|
292 |
+
),
|
293 |
+
)
|
294 |
+
|
295 |
+
self.out_layers = nn.Sequential(
|
296 |
+
normalization(self.out_channels),
|
297 |
+
nn.SiLU(),
|
298 |
+
nn.Dropout(p=dropout),
|
299 |
+
zero_module(
|
300 |
+
conv_nd(
|
301 |
+
dims,
|
302 |
+
self.out_channels,
|
303 |
+
self.out_channels,
|
304 |
+
kernel_size,
|
305 |
+
padding=padding,
|
306 |
+
)
|
307 |
+
),
|
308 |
+
)
|
309 |
+
|
310 |
+
if self.out_channels == channels:
|
311 |
+
self.skip_connection = nn.Identity()
|
312 |
+
elif use_conv:
|
313 |
+
self.skip_connection = conv_nd(
|
314 |
+
dims, channels, self.out_channels, kernel_size, padding=padding
|
315 |
+
)
|
316 |
+
else:
|
317 |
+
self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
|
318 |
+
|
319 |
+
def forward(self, x, emb):
|
320 |
+
"""
|
321 |
+
Apply the block to a Tensor, conditioned on a timestep embedding.
|
322 |
+
:param x: an [N x C x ...] Tensor of features.
|
323 |
+
:param emb: an [N x emb_channels] Tensor of timestep embeddings.
|
324 |
+
:return: an [N x C x ...] Tensor of outputs.
|
325 |
+
"""
|
326 |
+
return checkpoint(
|
327 |
+
self._forward, (x, emb), self.parameters(), self.use_checkpoint
|
328 |
+
)
|
329 |
+
|
330 |
+
def _forward(self, x, emb):
|
331 |
+
if self.updown:
|
332 |
+
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
|
333 |
+
h = in_rest(x)
|
334 |
+
h = self.h_upd(h)
|
335 |
+
x = self.x_upd(x)
|
336 |
+
h = in_conv(h)
|
337 |
+
else:
|
338 |
+
h = self.in_layers(x)
|
339 |
+
|
340 |
+
if self.skip_t_emb:
|
341 |
+
emb_out = th.zeros_like(h)
|
342 |
+
else:
|
343 |
+
emb_out = self.emb_layers(emb).type(h.dtype)
|
344 |
+
while len(emb_out.shape) < len(h.shape):
|
345 |
+
emb_out = emb_out[..., None]
|
346 |
+
if self.use_scale_shift_norm:
|
347 |
+
out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
|
348 |
+
scale, shift = th.chunk(emb_out, 2, dim=1)
|
349 |
+
h = out_norm(h) * (1 + scale) + shift
|
350 |
+
h = out_rest(h)
|
351 |
+
else:
|
352 |
+
if self.exchange_temb_dims:
|
353 |
+
emb_out = rearrange(emb_out, "b t c ... -> b c t ...")
|
354 |
+
h = h + emb_out
|
355 |
+
h = self.out_layers(h)
|
356 |
+
return self.skip_connection(x) + h
|
357 |
+
|
358 |
+
|
359 |
+
class AttentionBlock(nn.Module):
|
360 |
+
"""
|
361 |
+
An attention block that allows spatial positions to attend to each other.
|
362 |
+
Originally ported from here, but adapted to the N-d case.
|
363 |
+
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
|
364 |
+
"""
|
365 |
+
|
366 |
+
def __init__(
|
367 |
+
self,
|
368 |
+
channels,
|
369 |
+
num_heads=1,
|
370 |
+
num_head_channels=-1,
|
371 |
+
use_checkpoint=False,
|
372 |
+
use_new_attention_order=False,
|
373 |
+
):
|
374 |
+
super().__init__()
|
375 |
+
self.channels = channels
|
376 |
+
if num_head_channels == -1:
|
377 |
+
self.num_heads = num_heads
|
378 |
+
else:
|
379 |
+
assert (
|
380 |
+
channels % num_head_channels == 0
|
381 |
+
), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
|
382 |
+
self.num_heads = channels // num_head_channels
|
383 |
+
self.use_checkpoint = use_checkpoint
|
384 |
+
self.norm = normalization(channels)
|
385 |
+
self.qkv = conv_nd(1, channels, channels * 3, 1)
|
386 |
+
if use_new_attention_order:
|
387 |
+
# split qkv before split heads
|
388 |
+
self.attention = QKVAttention(self.num_heads)
|
389 |
+
else:
|
390 |
+
# split heads before split qkv
|
391 |
+
self.attention = QKVAttentionLegacy(self.num_heads)
|
392 |
+
|
393 |
+
self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
|
394 |
+
|
395 |
+
def forward(self, x, **kwargs):
|
396 |
+
# TODO add crossframe attention and use mixed checkpoint
|
397 |
+
return checkpoint(
|
398 |
+
self._forward, (x,), self.parameters(), True
|
399 |
+
) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
|
400 |
+
# return pt_checkpoint(self._forward, x) # pytorch
|
401 |
+
|
402 |
+
def _forward(self, x):
|
403 |
+
b, c, *spatial = x.shape
|
404 |
+
x = x.reshape(b, c, -1)
|
405 |
+
qkv = self.qkv(self.norm(x))
|
406 |
+
h = self.attention(qkv)
|
407 |
+
h = self.proj_out(h)
|
408 |
+
return (x + h).reshape(b, c, *spatial)
|
409 |
+
|
410 |
+
|
411 |
+
def count_flops_attn(model, _x, y):
|
412 |
+
"""
|
413 |
+
A counter for the `thop` package to count the operations in an
|
414 |
+
attention operation.
|
415 |
+
Meant to be used like:
|
416 |
+
macs, params = thop.profile(
|
417 |
+
model,
|
418 |
+
inputs=(inputs, timestamps),
|
419 |
+
custom_ops={QKVAttention: QKVAttention.count_flops},
|
420 |
+
)
|
421 |
+
"""
|
422 |
+
b, c, *spatial = y[0].shape
|
423 |
+
num_spatial = int(np.prod(spatial))
|
424 |
+
# We perform two matmuls with the same number of ops.
|
425 |
+
# The first computes the weight matrix, the second computes
|
426 |
+
# the combination of the value vectors.
|
427 |
+
matmul_ops = 2 * b * (num_spatial**2) * c
|
428 |
+
model.total_ops += th.DoubleTensor([matmul_ops])
|
429 |
+
|
430 |
+
|
431 |
+
class QKVAttentionLegacy(nn.Module):
|
432 |
+
"""
|
433 |
+
A module which performs QKV attention. Matches legacy QKVAttention + input/output heads shaping
|
434 |
+
"""
|
435 |
+
|
436 |
+
def __init__(self, n_heads):
|
437 |
+
super().__init__()
|
438 |
+
self.n_heads = n_heads
|
439 |
+
|
440 |
+
def forward(self, qkv):
|
441 |
+
"""
|
442 |
+
Apply QKV attention.
|
443 |
+
:param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
|
444 |
+
:return: an [N x (H * C) x T] tensor after attention.
|
445 |
+
"""
|
446 |
+
bs, width, length = qkv.shape
|
447 |
+
assert width % (3 * self.n_heads) == 0
|
448 |
+
ch = width // (3 * self.n_heads)
|
449 |
+
q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
|
450 |
+
scale = 1 / math.sqrt(math.sqrt(ch))
|
451 |
+
weight = th.einsum(
|
452 |
+
"bct,bcs->bts", q * scale, k * scale
|
453 |
+
) # More stable with f16 than dividing afterwards
|
454 |
+
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
|
455 |
+
a = th.einsum("bts,bcs->bct", weight, v)
|
456 |
+
return a.reshape(bs, -1, length)
|
457 |
+
|
458 |
+
@staticmethod
|
459 |
+
def count_flops(model, _x, y):
|
460 |
+
return count_flops_attn(model, _x, y)
|
461 |
+
|
462 |
+
|
463 |
+
class QKVAttention(nn.Module):
|
464 |
+
"""
|
465 |
+
A module which performs QKV attention and splits in a different order.
|
466 |
+
"""
|
467 |
+
|
468 |
+
def __init__(self, n_heads):
|
469 |
+
super().__init__()
|
470 |
+
self.n_heads = n_heads
|
471 |
+
|
472 |
+
def forward(self, qkv):
|
473 |
+
"""
|
474 |
+
Apply QKV attention.
|
475 |
+
:param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
|
476 |
+
:return: an [N x (H * C) x T] tensor after attention.
|
477 |
+
"""
|
478 |
+
bs, width, length = qkv.shape
|
479 |
+
assert width % (3 * self.n_heads) == 0
|
480 |
+
ch = width // (3 * self.n_heads)
|
481 |
+
q, k, v = qkv.chunk(3, dim=1)
|
482 |
+
scale = 1 / math.sqrt(math.sqrt(ch))
|
483 |
+
weight = th.einsum(
|
484 |
+
"bct,bcs->bts",
|
485 |
+
(q * scale).view(bs * self.n_heads, ch, length),
|
486 |
+
(k * scale).view(bs * self.n_heads, ch, length),
|
487 |
+
) # More stable with f16 than dividing afterwards
|
488 |
+
weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
|
489 |
+
a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
|
490 |
+
return a.reshape(bs, -1, length)
|
491 |
+
|
492 |
+
@staticmethod
|
493 |
+
def count_flops(model, _x, y):
|
494 |
+
return count_flops_attn(model, _x, y)
|
495 |
+
|
496 |
+
|
497 |
+
class Timestep(nn.Module):
|
498 |
+
def __init__(self, dim):
|
499 |
+
super().__init__()
|
500 |
+
self.dim = dim
|
501 |
+
|
502 |
+
def forward(self, t):
|
503 |
+
return timestep_embedding(t, self.dim)
|
504 |
+
|
505 |
+
|
506 |
+
class UNetModel(nn.Module):
|
507 |
+
"""
|
508 |
+
The full UNet model with attention and timestep embedding.
|
509 |
+
:param in_channels: channels in the input Tensor.
|
510 |
+
:param model_channels: base channel count for the model.
|
511 |
+
:param out_channels: channels in the output Tensor.
|
512 |
+
:param num_res_blocks: number of residual blocks per downsample.
|
513 |
+
:param attention_resolutions: a collection of downsample rates at which
|
514 |
+
attention will take place. May be a set, list, or tuple.
|
515 |
+
For example, if this contains 4, then at 4x downsampling, attention
|
516 |
+
will be used.
|
517 |
+
:param dropout: the dropout probability.
|
518 |
+
:param channel_mult: channel multiplier for each level of the UNet.
|
519 |
+
:param conv_resample: if True, use learned convolutions for upsampling and
|
520 |
+
downsampling.
|
521 |
+
:param dims: determines if the signal is 1D, 2D, or 3D.
|
522 |
+
:param num_classes: if specified (as an int), then this model will be
|
523 |
+
class-conditional with `num_classes` classes.
|
524 |
+
:param use_checkpoint: use gradient checkpointing to reduce memory usage.
|
525 |
+
:param num_heads: the number of attention heads in each attention layer.
|
526 |
+
:param num_heads_channels: if specified, ignore num_heads and instead use
|
527 |
+
a fixed channel width per attention head.
|
528 |
+
:param num_heads_upsample: works with num_heads to set a different number
|
529 |
+
of heads for upsampling. Deprecated.
|
530 |
+
:param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
|
531 |
+
:param resblock_updown: use residual blocks for up/downsampling.
|
532 |
+
:param use_new_attention_order: use a different attention pattern for potentially
|
533 |
+
increased efficiency.
|
534 |
+
"""
|
535 |
+
|
536 |
+
def __init__(
|
537 |
+
self,
|
538 |
+
in_channels,
|
539 |
+
model_channels,
|
540 |
+
out_channels,
|
541 |
+
num_res_blocks,
|
542 |
+
attention_resolutions,
|
543 |
+
dropout=0,
|
544 |
+
channel_mult=(1, 2, 4, 8),
|
545 |
+
conv_resample=True,
|
546 |
+
dims=2,
|
547 |
+
num_classes=None,
|
548 |
+
use_checkpoint=False,
|
549 |
+
use_fp16=False,
|
550 |
+
num_heads=-1,
|
551 |
+
num_head_channels=-1,
|
552 |
+
num_heads_upsample=-1,
|
553 |
+
use_scale_shift_norm=False,
|
554 |
+
resblock_updown=False,
|
555 |
+
use_new_attention_order=False,
|
556 |
+
use_spatial_transformer=False, # custom transformer support
|
557 |
+
transformer_depth=1, # custom transformer support
|
558 |
+
context_dim=None, # custom transformer support
|
559 |
+
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
|
560 |
+
legacy=True,
|
561 |
+
disable_self_attentions=None,
|
562 |
+
num_attention_blocks=None,
|
563 |
+
disable_middle_self_attn=False,
|
564 |
+
use_linear_in_transformer=False,
|
565 |
+
spatial_transformer_attn_type="softmax",
|
566 |
+
adm_in_channels=None,
|
567 |
+
use_fairscale_checkpoint=False,
|
568 |
+
offload_to_cpu=False,
|
569 |
+
transformer_depth_middle=None,
|
570 |
+
):
|
571 |
+
super().__init__()
|
572 |
+
from omegaconf.listconfig import ListConfig
|
573 |
+
|
574 |
+
if use_spatial_transformer:
|
575 |
+
assert (
|
576 |
+
context_dim is not None
|
577 |
+
), "Fool!! You forgot to include the dimension of your cross-attention conditioning..."
|
578 |
+
|
579 |
+
if context_dim is not None:
|
580 |
+
assert (
|
581 |
+
use_spatial_transformer
|
582 |
+
), "Fool!! You forgot to use the spatial transformer for your cross-attention conditioning..."
|
583 |
+
if type(context_dim) == ListConfig:
|
584 |
+
context_dim = list(context_dim)
|
585 |
+
|
586 |
+
if num_heads_upsample == -1:
|
587 |
+
num_heads_upsample = num_heads
|
588 |
+
|
589 |
+
if num_heads == -1:
|
590 |
+
assert (
|
591 |
+
num_head_channels != -1
|
592 |
+
), "Either num_heads or num_head_channels has to be set"
|
593 |
+
|
594 |
+
if num_head_channels == -1:
|
595 |
+
assert (
|
596 |
+
num_heads != -1
|
597 |
+
), "Either num_heads or num_head_channels has to be set"
|
598 |
+
|
599 |
+
self.in_channels = in_channels
|
600 |
+
self.model_channels = model_channels
|
601 |
+
self.out_channels = out_channels
|
602 |
+
if isinstance(transformer_depth, int):
|
603 |
+
transformer_depth = len(channel_mult) * [transformer_depth]
|
604 |
+
elif isinstance(transformer_depth, ListConfig):
|
605 |
+
transformer_depth = list(transformer_depth)
|
606 |
+
transformer_depth_middle = default(
|
607 |
+
transformer_depth_middle, transformer_depth[-1]
|
608 |
+
)
|
609 |
+
|
610 |
+
if isinstance(num_res_blocks, int):
|
611 |
+
self.num_res_blocks = len(channel_mult) * [num_res_blocks]
|
612 |
+
else:
|
613 |
+
if len(num_res_blocks) != len(channel_mult):
|
614 |
+
raise ValueError(
|
615 |
+
"provide num_res_blocks either as an int (globally constant) or "
|
616 |
+
"as a list/tuple (per-level) with the same length as channel_mult"
|
617 |
+
)
|
618 |
+
self.num_res_blocks = num_res_blocks
|
619 |
+
# self.num_res_blocks = num_res_blocks
|
620 |
+
if disable_self_attentions is not None:
|
621 |
+
# should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
|
622 |
+
assert len(disable_self_attentions) == len(channel_mult)
|
623 |
+
if num_attention_blocks is not None:
|
624 |
+
assert len(num_attention_blocks) == len(self.num_res_blocks)
|
625 |
+
assert all(
|
626 |
+
map(
|
627 |
+
lambda i: self.num_res_blocks[i] >= num_attention_blocks[i],
|
628 |
+
range(len(num_attention_blocks)),
|
629 |
+
)
|
630 |
+
)
|
631 |
+
print(
|
632 |
+
f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
|
633 |
+
f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
|
634 |
+
f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
|
635 |
+
f"attention will still not be set."
|
636 |
+
) # todo: convert to warning
|
637 |
+
|
638 |
+
self.attention_resolutions = attention_resolutions
|
639 |
+
self.dropout = dropout
|
640 |
+
self.channel_mult = channel_mult
|
641 |
+
self.conv_resample = conv_resample
|
642 |
+
self.num_classes = num_classes
|
643 |
+
self.use_checkpoint = use_checkpoint
|
644 |
+
if use_fp16:
|
645 |
+
print("WARNING: use_fp16 was dropped and has no effect anymore.")
|
646 |
+
# self.dtype = th.float16 if use_fp16 else th.float32
|
647 |
+
self.num_heads = num_heads
|
648 |
+
self.num_head_channels = num_head_channels
|
649 |
+
self.num_heads_upsample = num_heads_upsample
|
650 |
+
self.predict_codebook_ids = n_embed is not None
|
651 |
+
|
652 |
+
assert use_fairscale_checkpoint != use_checkpoint or not (
|
653 |
+
use_checkpoint or use_fairscale_checkpoint
|
654 |
+
)
|
655 |
+
|
656 |
+
self.use_fairscale_checkpoint = False
|
657 |
+
checkpoint_wrapper_fn = (
|
658 |
+
partial(checkpoint_wrapper, offload_to_cpu=offload_to_cpu)
|
659 |
+
if self.use_fairscale_checkpoint
|
660 |
+
else lambda x: x
|
661 |
+
)
|
662 |
+
|
663 |
+
time_embed_dim = model_channels * 4
|
664 |
+
self.time_embed = checkpoint_wrapper_fn(
|
665 |
+
nn.Sequential(
|
666 |
+
linear(model_channels, time_embed_dim),
|
667 |
+
nn.SiLU(),
|
668 |
+
linear(time_embed_dim, time_embed_dim),
|
669 |
+
)
|
670 |
+
)
|
671 |
+
|
672 |
+
if self.num_classes is not None:
|
673 |
+
if isinstance(self.num_classes, int):
|
674 |
+
self.label_emb = nn.Embedding(num_classes, time_embed_dim)
|
675 |
+
elif self.num_classes == "continuous":
|
676 |
+
print("setting up linear c_adm embedding layer")
|
677 |
+
self.label_emb = nn.Linear(1, time_embed_dim)
|
678 |
+
elif self.num_classes == "timestep":
|
679 |
+
self.label_emb = checkpoint_wrapper_fn(
|
680 |
+
nn.Sequential(
|
681 |
+
Timestep(model_channels),
|
682 |
+
nn.Sequential(
|
683 |
+
linear(model_channels, time_embed_dim),
|
684 |
+
nn.SiLU(),
|
685 |
+
linear(time_embed_dim, time_embed_dim),
|
686 |
+
),
|
687 |
+
)
|
688 |
+
)
|
689 |
+
elif self.num_classes == "sequential":
|
690 |
+
assert adm_in_channels is not None
|
691 |
+
self.label_emb = nn.Sequential(
|
692 |
+
nn.Sequential(
|
693 |
+
linear(adm_in_channels, time_embed_dim),
|
694 |
+
nn.SiLU(),
|
695 |
+
linear(time_embed_dim, time_embed_dim),
|
696 |
+
)
|
697 |
+
)
|
698 |
+
else:
|
699 |
+
raise ValueError()
|
700 |
+
|
701 |
+
self.input_blocks = nn.ModuleList(
|
702 |
+
[
|
703 |
+
TimestepEmbedSequential(
|
704 |
+
conv_nd(dims, in_channels, model_channels, 3, padding=1)
|
705 |
+
)
|
706 |
+
]
|
707 |
+
)
|
708 |
+
self._feature_size = model_channels
|
709 |
+
input_block_chans = [model_channels]
|
710 |
+
ch = model_channels
|
711 |
+
ds = 1
|
712 |
+
for level, mult in enumerate(channel_mult):
|
713 |
+
for nr in range(self.num_res_blocks[level]):
|
714 |
+
layers = [
|
715 |
+
checkpoint_wrapper_fn(
|
716 |
+
ResBlock(
|
717 |
+
ch,
|
718 |
+
time_embed_dim,
|
719 |
+
dropout,
|
720 |
+
out_channels=mult * model_channels,
|
721 |
+
dims=dims,
|
722 |
+
use_checkpoint=use_checkpoint,
|
723 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
724 |
+
)
|
725 |
+
)
|
726 |
+
]
|
727 |
+
ch = mult * model_channels
|
728 |
+
if ds in attention_resolutions:
|
729 |
+
if num_head_channels == -1:
|
730 |
+
dim_head = ch // num_heads
|
731 |
+
else:
|
732 |
+
num_heads = ch // num_head_channels
|
733 |
+
dim_head = num_head_channels
|
734 |
+
if legacy:
|
735 |
+
# num_heads = 1
|
736 |
+
dim_head = (
|
737 |
+
ch // num_heads
|
738 |
+
if use_spatial_transformer
|
739 |
+
else num_head_channels
|
740 |
+
)
|
741 |
+
if exists(disable_self_attentions):
|
742 |
+
disabled_sa = disable_self_attentions[level]
|
743 |
+
else:
|
744 |
+
disabled_sa = False
|
745 |
+
|
746 |
+
if (
|
747 |
+
not exists(num_attention_blocks)
|
748 |
+
or nr < num_attention_blocks[level]
|
749 |
+
):
|
750 |
+
layers.append(
|
751 |
+
checkpoint_wrapper_fn(
|
752 |
+
AttentionBlock(
|
753 |
+
ch,
|
754 |
+
use_checkpoint=use_checkpoint,
|
755 |
+
num_heads=num_heads,
|
756 |
+
num_head_channels=dim_head,
|
757 |
+
use_new_attention_order=use_new_attention_order,
|
758 |
+
)
|
759 |
+
)
|
760 |
+
if not use_spatial_transformer
|
761 |
+
else checkpoint_wrapper_fn(
|
762 |
+
SpatialTransformer(
|
763 |
+
ch,
|
764 |
+
num_heads,
|
765 |
+
dim_head,
|
766 |
+
depth=transformer_depth[level],
|
767 |
+
context_dim=context_dim,
|
768 |
+
disable_self_attn=disabled_sa,
|
769 |
+
use_linear=use_linear_in_transformer,
|
770 |
+
attn_type=spatial_transformer_attn_type,
|
771 |
+
use_checkpoint=use_checkpoint,
|
772 |
+
)
|
773 |
+
)
|
774 |
+
)
|
775 |
+
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
776 |
+
self._feature_size += ch
|
777 |
+
input_block_chans.append(ch)
|
778 |
+
if level != len(channel_mult) - 1:
|
779 |
+
out_ch = ch
|
780 |
+
self.input_blocks.append(
|
781 |
+
TimestepEmbedSequential(
|
782 |
+
checkpoint_wrapper_fn(
|
783 |
+
ResBlock(
|
784 |
+
ch,
|
785 |
+
time_embed_dim,
|
786 |
+
dropout,
|
787 |
+
out_channels=out_ch,
|
788 |
+
dims=dims,
|
789 |
+
use_checkpoint=use_checkpoint,
|
790 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
791 |
+
down=True,
|
792 |
+
)
|
793 |
+
)
|
794 |
+
if resblock_updown
|
795 |
+
else Downsample(
|
796 |
+
ch, conv_resample, dims=dims, out_channels=out_ch
|
797 |
+
)
|
798 |
+
)
|
799 |
+
)
|
800 |
+
ch = out_ch
|
801 |
+
input_block_chans.append(ch)
|
802 |
+
ds *= 2
|
803 |
+
self._feature_size += ch
|
804 |
+
|
805 |
+
if num_head_channels == -1:
|
806 |
+
dim_head = ch // num_heads
|
807 |
+
else:
|
808 |
+
num_heads = ch // num_head_channels
|
809 |
+
dim_head = num_head_channels
|
810 |
+
if legacy:
|
811 |
+
# num_heads = 1
|
812 |
+
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
813 |
+
self.middle_block = TimestepEmbedSequential(
|
814 |
+
checkpoint_wrapper_fn(
|
815 |
+
ResBlock(
|
816 |
+
ch,
|
817 |
+
time_embed_dim,
|
818 |
+
dropout,
|
819 |
+
dims=dims,
|
820 |
+
use_checkpoint=use_checkpoint,
|
821 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
822 |
+
)
|
823 |
+
),
|
824 |
+
checkpoint_wrapper_fn(
|
825 |
+
AttentionBlock(
|
826 |
+
ch,
|
827 |
+
use_checkpoint=use_checkpoint,
|
828 |
+
num_heads=num_heads,
|
829 |
+
num_head_channels=dim_head,
|
830 |
+
use_new_attention_order=use_new_attention_order,
|
831 |
+
)
|
832 |
+
)
|
833 |
+
if not use_spatial_transformer
|
834 |
+
else checkpoint_wrapper_fn(
|
835 |
+
SpatialTransformer( # always uses a self-attn
|
836 |
+
ch,
|
837 |
+
num_heads,
|
838 |
+
dim_head,
|
839 |
+
depth=transformer_depth_middle,
|
840 |
+
context_dim=context_dim,
|
841 |
+
disable_self_attn=disable_middle_self_attn,
|
842 |
+
use_linear=use_linear_in_transformer,
|
843 |
+
attn_type=spatial_transformer_attn_type,
|
844 |
+
use_checkpoint=use_checkpoint,
|
845 |
+
)
|
846 |
+
),
|
847 |
+
checkpoint_wrapper_fn(
|
848 |
+
ResBlock(
|
849 |
+
ch,
|
850 |
+
time_embed_dim,
|
851 |
+
dropout,
|
852 |
+
dims=dims,
|
853 |
+
use_checkpoint=use_checkpoint,
|
854 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
855 |
+
)
|
856 |
+
),
|
857 |
+
)
|
858 |
+
self._feature_size += ch
|
859 |
+
|
860 |
+
self.output_blocks = nn.ModuleList([])
|
861 |
+
for level, mult in list(enumerate(channel_mult))[::-1]:
|
862 |
+
for i in range(self.num_res_blocks[level] + 1):
|
863 |
+
ich = input_block_chans.pop()
|
864 |
+
layers = [
|
865 |
+
checkpoint_wrapper_fn(
|
866 |
+
ResBlock(
|
867 |
+
ch + ich,
|
868 |
+
time_embed_dim,
|
869 |
+
dropout,
|
870 |
+
out_channels=model_channels * mult,
|
871 |
+
dims=dims,
|
872 |
+
use_checkpoint=use_checkpoint,
|
873 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
874 |
+
)
|
875 |
+
)
|
876 |
+
]
|
877 |
+
ch = model_channels * mult
|
878 |
+
if ds in attention_resolutions:
|
879 |
+
if num_head_channels == -1:
|
880 |
+
dim_head = ch // num_heads
|
881 |
+
else:
|
882 |
+
num_heads = ch // num_head_channels
|
883 |
+
dim_head = num_head_channels
|
884 |
+
if legacy:
|
885 |
+
# num_heads = 1
|
886 |
+
dim_head = (
|
887 |
+
ch // num_heads
|
888 |
+
if use_spatial_transformer
|
889 |
+
else num_head_channels
|
890 |
+
)
|
891 |
+
if exists(disable_self_attentions):
|
892 |
+
disabled_sa = disable_self_attentions[level]
|
893 |
+
else:
|
894 |
+
disabled_sa = False
|
895 |
+
|
896 |
+
if (
|
897 |
+
not exists(num_attention_blocks)
|
898 |
+
or i < num_attention_blocks[level]
|
899 |
+
):
|
900 |
+
layers.append(
|
901 |
+
checkpoint_wrapper_fn(
|
902 |
+
AttentionBlock(
|
903 |
+
ch,
|
904 |
+
use_checkpoint=use_checkpoint,
|
905 |
+
num_heads=num_heads_upsample,
|
906 |
+
num_head_channels=dim_head,
|
907 |
+
use_new_attention_order=use_new_attention_order,
|
908 |
+
)
|
909 |
+
)
|
910 |
+
if not use_spatial_transformer
|
911 |
+
else checkpoint_wrapper_fn(
|
912 |
+
SpatialTransformer(
|
913 |
+
ch,
|
914 |
+
num_heads,
|
915 |
+
dim_head,
|
916 |
+
depth=transformer_depth[level],
|
917 |
+
context_dim=context_dim,
|
918 |
+
disable_self_attn=disabled_sa,
|
919 |
+
use_linear=use_linear_in_transformer,
|
920 |
+
attn_type=spatial_transformer_attn_type,
|
921 |
+
use_checkpoint=use_checkpoint,
|
922 |
+
)
|
923 |
+
)
|
924 |
+
)
|
925 |
+
if level and i == self.num_res_blocks[level]:
|
926 |
+
out_ch = ch
|
927 |
+
layers.append(
|
928 |
+
checkpoint_wrapper_fn(
|
929 |
+
ResBlock(
|
930 |
+
ch,
|
931 |
+
time_embed_dim,
|
932 |
+
dropout,
|
933 |
+
out_channels=out_ch,
|
934 |
+
dims=dims,
|
935 |
+
use_checkpoint=use_checkpoint,
|
936 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
937 |
+
up=True,
|
938 |
+
)
|
939 |
+
)
|
940 |
+
if resblock_updown
|
941 |
+
else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
|
942 |
+
)
|
943 |
+
ds //= 2
|
944 |
+
self.output_blocks.append(TimestepEmbedSequential(*layers))
|
945 |
+
self._feature_size += ch
|
946 |
+
|
947 |
+
self.out = checkpoint_wrapper_fn(
|
948 |
+
nn.Sequential(
|
949 |
+
normalization(ch),
|
950 |
+
nn.SiLU(),
|
951 |
+
zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
|
952 |
+
)
|
953 |
+
)
|
954 |
+
if self.predict_codebook_ids:
|
955 |
+
self.id_predictor = checkpoint_wrapper_fn(
|
956 |
+
nn.Sequential(
|
957 |
+
normalization(ch),
|
958 |
+
conv_nd(dims, model_channels, n_embed, 1),
|
959 |
+
# nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits
|
960 |
+
)
|
961 |
+
)
|
962 |
+
|
963 |
+
def convert_to_fp16(self):
|
964 |
+
"""
|
965 |
+
Convert the torso of the model to float16.
|
966 |
+
"""
|
967 |
+
self.input_blocks.apply(convert_module_to_f16)
|
968 |
+
self.middle_block.apply(convert_module_to_f16)
|
969 |
+
self.output_blocks.apply(convert_module_to_f16)
|
970 |
+
|
971 |
+
def convert_to_fp32(self):
|
972 |
+
"""
|
973 |
+
Convert the torso of the model to float32.
|
974 |
+
"""
|
975 |
+
self.input_blocks.apply(convert_module_to_f32)
|
976 |
+
self.middle_block.apply(convert_module_to_f32)
|
977 |
+
self.output_blocks.apply(convert_module_to_f32)
|
978 |
+
|
979 |
+
def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
|
980 |
+
"""
|
981 |
+
Apply the model to an input batch.
|
982 |
+
:param x: an [N x C x ...] Tensor of inputs.
|
983 |
+
:param timesteps: a 1-D batch of timesteps.
|
984 |
+
:param context: conditioning plugged in via crossattn
|
985 |
+
:param y: an [N] Tensor of labels, if class-conditional.
|
986 |
+
:return: an [N x C x ...] Tensor of outputs.
|
987 |
+
"""
|
988 |
+
assert (y is not None) == (
|
989 |
+
self.num_classes is not None
|
990 |
+
), "must specify y if and only if the model is class-conditional"
|
991 |
+
hs = []
|
992 |
+
|
993 |
+
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
|
994 |
+
emb = self.time_embed(t_emb)
|
995 |
+
|
996 |
+
if self.num_classes is not None:
|
997 |
+
assert y.shape[0] == x.shape[0]
|
998 |
+
emb = emb + self.label_emb(y)
|
999 |
+
|
1000 |
+
# h = x.type(self.dtype)
|
1001 |
+
h = x
|
1002 |
+
for module in self.input_blocks:
|
1003 |
+
h = module(h, emb, context)
|
1004 |
+
hs.append(h)
|
1005 |
+
h = self.middle_block(h, emb, context)
|
1006 |
+
for module in self.output_blocks:
|
1007 |
+
h = th.cat([h, hs.pop()], dim=1)
|
1008 |
+
h = module(h, emb, context)
|
1009 |
+
h = h.type(x.dtype)
|
1010 |
+
if self.predict_codebook_ids:
|
1011 |
+
assert False, "not supported anymore. what the f*** are you doing?"
|
1012 |
+
else:
|
1013 |
+
return self.out(h)
|
1014 |
+
|
1015 |
+
|
1016 |
+
class NoTimeUNetModel(UNetModel):
|
1017 |
+
def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
|
1018 |
+
timesteps = th.zeros_like(timesteps)
|
1019 |
+
return super().forward(x, timesteps, context, y, **kwargs)
|
1020 |
+
|
1021 |
+
|
1022 |
+
class EncoderUNetModel(nn.Module):
|
1023 |
+
"""
|
1024 |
+
The half UNet model with attention and timestep embedding.
|
1025 |
+
For usage, see UNet.
|
1026 |
+
"""
|
1027 |
+
|
1028 |
+
def __init__(
|
1029 |
+
self,
|
1030 |
+
image_size,
|
1031 |
+
in_channels,
|
1032 |
+
model_channels,
|
1033 |
+
out_channels,
|
1034 |
+
num_res_blocks,
|
1035 |
+
attention_resolutions,
|
1036 |
+
dropout=0,
|
1037 |
+
channel_mult=(1, 2, 4, 8),
|
1038 |
+
conv_resample=True,
|
1039 |
+
dims=2,
|
1040 |
+
use_checkpoint=False,
|
1041 |
+
use_fp16=False,
|
1042 |
+
num_heads=1,
|
1043 |
+
num_head_channels=-1,
|
1044 |
+
num_heads_upsample=-1,
|
1045 |
+
use_scale_shift_norm=False,
|
1046 |
+
resblock_updown=False,
|
1047 |
+
use_new_attention_order=False,
|
1048 |
+
pool="adaptive",
|
1049 |
+
*args,
|
1050 |
+
**kwargs,
|
1051 |
+
):
|
1052 |
+
super().__init__()
|
1053 |
+
|
1054 |
+
if num_heads_upsample == -1:
|
1055 |
+
num_heads_upsample = num_heads
|
1056 |
+
|
1057 |
+
self.in_channels = in_channels
|
1058 |
+
self.model_channels = model_channels
|
1059 |
+
self.out_channels = out_channels
|
1060 |
+
self.num_res_blocks = num_res_blocks
|
1061 |
+
self.attention_resolutions = attention_resolutions
|
1062 |
+
self.dropout = dropout
|
1063 |
+
self.channel_mult = channel_mult
|
1064 |
+
self.conv_resample = conv_resample
|
1065 |
+
self.use_checkpoint = use_checkpoint
|
1066 |
+
self.dtype = th.float16 if use_fp16 else th.float32
|
1067 |
+
self.num_heads = num_heads
|
1068 |
+
self.num_head_channels = num_head_channels
|
1069 |
+
self.num_heads_upsample = num_heads_upsample
|
1070 |
+
|
1071 |
+
time_embed_dim = model_channels * 4
|
1072 |
+
self.time_embed = nn.Sequential(
|
1073 |
+
linear(model_channels, time_embed_dim),
|
1074 |
+
nn.SiLU(),
|
1075 |
+
linear(time_embed_dim, time_embed_dim),
|
1076 |
+
)
|
1077 |
+
|
1078 |
+
self.input_blocks = nn.ModuleList(
|
1079 |
+
[
|
1080 |
+
TimestepEmbedSequential(
|
1081 |
+
conv_nd(dims, in_channels, model_channels, 3, padding=1)
|
1082 |
+
)
|
1083 |
+
]
|
1084 |
+
)
|
1085 |
+
self._feature_size = model_channels
|
1086 |
+
input_block_chans = [model_channels]
|
1087 |
+
ch = model_channels
|
1088 |
+
ds = 1
|
1089 |
+
for level, mult in enumerate(channel_mult):
|
1090 |
+
for _ in range(num_res_blocks):
|
1091 |
+
layers = [
|
1092 |
+
ResBlock(
|
1093 |
+
ch,
|
1094 |
+
time_embed_dim,
|
1095 |
+
dropout,
|
1096 |
+
out_channels=mult * model_channels,
|
1097 |
+
dims=dims,
|
1098 |
+
use_checkpoint=use_checkpoint,
|
1099 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
1100 |
+
)
|
1101 |
+
]
|
1102 |
+
ch = mult * model_channels
|
1103 |
+
if ds in attention_resolutions:
|
1104 |
+
layers.append(
|
1105 |
+
AttentionBlock(
|
1106 |
+
ch,
|
1107 |
+
use_checkpoint=use_checkpoint,
|
1108 |
+
num_heads=num_heads,
|
1109 |
+
num_head_channels=num_head_channels,
|
1110 |
+
use_new_attention_order=use_new_attention_order,
|
1111 |
+
)
|
1112 |
+
)
|
1113 |
+
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
1114 |
+
self._feature_size += ch
|
1115 |
+
input_block_chans.append(ch)
|
1116 |
+
if level != len(channel_mult) - 1:
|
1117 |
+
out_ch = ch
|
1118 |
+
self.input_blocks.append(
|
1119 |
+
TimestepEmbedSequential(
|
1120 |
+
ResBlock(
|
1121 |
+
ch,
|
1122 |
+
time_embed_dim,
|
1123 |
+
dropout,
|
1124 |
+
out_channels=out_ch,
|
1125 |
+
dims=dims,
|
1126 |
+
use_checkpoint=use_checkpoint,
|
1127 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
1128 |
+
down=True,
|
1129 |
+
)
|
1130 |
+
if resblock_updown
|
1131 |
+
else Downsample(
|
1132 |
+
ch, conv_resample, dims=dims, out_channels=out_ch
|
1133 |
+
)
|
1134 |
+
)
|
1135 |
+
)
|
1136 |
+
ch = out_ch
|
1137 |
+
input_block_chans.append(ch)
|
1138 |
+
ds *= 2
|
1139 |
+
self._feature_size += ch
|
1140 |
+
|
1141 |
+
self.middle_block = TimestepEmbedSequential(
|
1142 |
+
ResBlock(
|
1143 |
+
ch,
|
1144 |
+
time_embed_dim,
|
1145 |
+
dropout,
|
1146 |
+
dims=dims,
|
1147 |
+
use_checkpoint=use_checkpoint,
|
1148 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
1149 |
+
),
|
1150 |
+
AttentionBlock(
|
1151 |
+
ch,
|
1152 |
+
use_checkpoint=use_checkpoint,
|
1153 |
+
num_heads=num_heads,
|
1154 |
+
num_head_channels=num_head_channels,
|
1155 |
+
use_new_attention_order=use_new_attention_order,
|
1156 |
+
),
|
1157 |
+
ResBlock(
|
1158 |
+
ch,
|
1159 |
+
time_embed_dim,
|
1160 |
+
dropout,
|
1161 |
+
dims=dims,
|
1162 |
+
use_checkpoint=use_checkpoint,
|
1163 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
1164 |
+
),
|
1165 |
+
)
|
1166 |
+
self._feature_size += ch
|
1167 |
+
self.pool = pool
|
1168 |
+
if pool == "adaptive":
|
1169 |
+
self.out = nn.Sequential(
|
1170 |
+
normalization(ch),
|
1171 |
+
nn.SiLU(),
|
1172 |
+
nn.AdaptiveAvgPool2d((1, 1)),
|
1173 |
+
zero_module(conv_nd(dims, ch, out_channels, 1)),
|
1174 |
+
nn.Flatten(),
|
1175 |
+
)
|
1176 |
+
elif pool == "attention":
|
1177 |
+
assert num_head_channels != -1
|
1178 |
+
self.out = nn.Sequential(
|
1179 |
+
normalization(ch),
|
1180 |
+
nn.SiLU(),
|
1181 |
+
AttentionPool2d(
|
1182 |
+
(image_size // ds), ch, num_head_channels, out_channels
|
1183 |
+
),
|
1184 |
+
)
|
1185 |
+
elif pool == "spatial":
|
1186 |
+
self.out = nn.Sequential(
|
1187 |
+
nn.Linear(self._feature_size, 2048),
|
1188 |
+
nn.ReLU(),
|
1189 |
+
nn.Linear(2048, self.out_channels),
|
1190 |
+
)
|
1191 |
+
elif pool == "spatial_v2":
|
1192 |
+
self.out = nn.Sequential(
|
1193 |
+
nn.Linear(self._feature_size, 2048),
|
1194 |
+
normalization(2048),
|
1195 |
+
nn.SiLU(),
|
1196 |
+
nn.Linear(2048, self.out_channels),
|
1197 |
+
)
|
1198 |
+
else:
|
1199 |
+
raise NotImplementedError(f"Unexpected {pool} pooling")
|
1200 |
+
|
1201 |
+
def convert_to_fp16(self):
|
1202 |
+
"""
|
1203 |
+
Convert the torso of the model to float16.
|
1204 |
+
"""
|
1205 |
+
self.input_blocks.apply(convert_module_to_f16)
|
1206 |
+
self.middle_block.apply(convert_module_to_f16)
|
1207 |
+
|
1208 |
+
def convert_to_fp32(self):
|
1209 |
+
"""
|
1210 |
+
Convert the torso of the model to float32.
|
1211 |
+
"""
|
1212 |
+
self.input_blocks.apply(convert_module_to_f32)
|
1213 |
+
self.middle_block.apply(convert_module_to_f32)
|
1214 |
+
|
1215 |
+
def forward(self, x, timesteps):
|
1216 |
+
"""
|
1217 |
+
Apply the model to an input batch.
|
1218 |
+
:param x: an [N x C x ...] Tensor of inputs.
|
1219 |
+
:param timesteps: a 1-D batch of timesteps.
|
1220 |
+
:return: an [N x K] Tensor of outputs.
|
1221 |
+
"""
|
1222 |
+
emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
|
1223 |
+
|
1224 |
+
results = []
|
1225 |
+
# h = x.type(self.dtype)
|
1226 |
+
h = x
|
1227 |
+
for module in self.input_blocks:
|
1228 |
+
h = module(h, emb)
|
1229 |
+
if self.pool.startswith("spatial"):
|
1230 |
+
results.append(h.type(x.dtype).mean(dim=(2, 3)))
|
1231 |
+
h = self.middle_block(h, emb)
|
1232 |
+
if self.pool.startswith("spatial"):
|
1233 |
+
results.append(h.type(x.dtype).mean(dim=(2, 3)))
|
1234 |
+
h = th.cat(results, axis=-1)
|
1235 |
+
return self.out(h)
|
1236 |
+
else:
|
1237 |
+
h = h.type(x.dtype)
|
1238 |
+
return self.out(h)
|
1239 |
+
|
1240 |
+
|
1241 |
+
if __name__ == "__main__":
|
1242 |
+
|
1243 |
+
class Dummy(nn.Module):
|
1244 |
+
def __init__(self, in_channels=3, model_channels=64):
|
1245 |
+
super().__init__()
|
1246 |
+
self.input_blocks = nn.ModuleList(
|
1247 |
+
[
|
1248 |
+
TimestepEmbedSequential(
|
1249 |
+
conv_nd(2, in_channels, model_channels, 3, padding=1)
|
1250 |
+
)
|
1251 |
+
]
|
1252 |
+
)
|
1253 |
+
|
1254 |
+
model = UNetModel(
|
1255 |
+
use_checkpoint=True,
|
1256 |
+
image_size=64,
|
1257 |
+
in_channels=4,
|
1258 |
+
out_channels=4,
|
1259 |
+
model_channels=128,
|
1260 |
+
attention_resolutions=[4, 2],
|
1261 |
+
num_res_blocks=2,
|
1262 |
+
channel_mult=[1, 2, 4],
|
1263 |
+
num_head_channels=64,
|
1264 |
+
use_spatial_transformer=False,
|
1265 |
+
use_linear_in_transformer=True,
|
1266 |
+
transformer_depth=1,
|
1267 |
+
legacy=False,
|
1268 |
+
).cuda()
|
1269 |
+
x = th.randn(11, 4, 64, 64).cuda()
|
1270 |
+
t = th.randint(low=0, high=10, size=(11,), device="cuda")
|
1271 |
+
o = model(x, t)
|
1272 |
+
print("done.")
|
sgm/modules/diffusionmodules/sampling.py
ADDED
@@ -0,0 +1,766 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Partially ported from https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py
|
3 |
+
"""
|
4 |
+
|
5 |
+
|
6 |
+
from typing import Dict, Union
|
7 |
+
|
8 |
+
import torch
|
9 |
+
from omegaconf import ListConfig, OmegaConf
|
10 |
+
from tqdm import tqdm
|
11 |
+
|
12 |
+
from ...modules.diffusionmodules.sampling_utils import (
|
13 |
+
get_ancestral_step,
|
14 |
+
linear_multistep_coeff,
|
15 |
+
to_d,
|
16 |
+
to_neg_log_sigma,
|
17 |
+
to_sigma,
|
18 |
+
)
|
19 |
+
from ...util import append_dims, default, instantiate_from_config
|
20 |
+
from k_diffusion.sampling import get_sigmas_karras, BrownianTreeNoiseSampler
|
21 |
+
|
22 |
+
DEFAULT_GUIDER = {"target": "sgm.modules.diffusionmodules.guiders.IdentityGuider"}
|
23 |
+
|
24 |
+
|
25 |
+
class BaseDiffusionSampler:
|
26 |
+
def __init__(
|
27 |
+
self,
|
28 |
+
discretization_config: Union[Dict, ListConfig, OmegaConf],
|
29 |
+
num_steps: Union[int, None] = None,
|
30 |
+
guider_config: Union[Dict, ListConfig, OmegaConf, None] = None,
|
31 |
+
verbose: bool = False,
|
32 |
+
device: str = "cuda",
|
33 |
+
):
|
34 |
+
self.num_steps = num_steps
|
35 |
+
self.discretization = instantiate_from_config(discretization_config)
|
36 |
+
self.guider = instantiate_from_config(
|
37 |
+
default(
|
38 |
+
guider_config,
|
39 |
+
DEFAULT_GUIDER,
|
40 |
+
)
|
41 |
+
)
|
42 |
+
self.verbose = verbose
|
43 |
+
self.device = device
|
44 |
+
|
45 |
+
def prepare_sampling_loop(self, x, cond, uc=None, num_steps=None):
|
46 |
+
sigmas = self.discretization(
|
47 |
+
self.num_steps if num_steps is None else num_steps, device=self.device
|
48 |
+
)
|
49 |
+
uc = default(uc, cond)
|
50 |
+
|
51 |
+
x *= torch.sqrt(1.0 + sigmas[0] ** 2.0)
|
52 |
+
num_sigmas = len(sigmas)
|
53 |
+
|
54 |
+
s_in = x.new_ones([x.shape[0]])
|
55 |
+
|
56 |
+
return x, s_in, sigmas, num_sigmas, cond, uc
|
57 |
+
|
58 |
+
def denoise(self, x, denoiser, sigma, cond, uc):
|
59 |
+
denoised = denoiser(*self.guider.prepare_inputs(x, sigma, cond, uc))
|
60 |
+
denoised = self.guider(denoised, sigma)
|
61 |
+
return denoised
|
62 |
+
|
63 |
+
def get_sigma_gen(self, num_sigmas):
|
64 |
+
sigma_generator = range(num_sigmas - 1)
|
65 |
+
if self.verbose:
|
66 |
+
print("#" * 30, " Sampling setting ", "#" * 30)
|
67 |
+
print(f"Sampler: {self.__class__.__name__}")
|
68 |
+
print(f"Discretization: {self.discretization.__class__.__name__}")
|
69 |
+
print(f"Guider: {self.guider.__class__.__name__}")
|
70 |
+
sigma_generator = tqdm(
|
71 |
+
sigma_generator,
|
72 |
+
total=num_sigmas,
|
73 |
+
desc=f"Sampling with {self.__class__.__name__} for {num_sigmas} steps",
|
74 |
+
)
|
75 |
+
return sigma_generator
|
76 |
+
|
77 |
+
|
78 |
+
class SingleStepDiffusionSampler(BaseDiffusionSampler):
|
79 |
+
def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc, *args, **kwargs):
|
80 |
+
raise NotImplementedError
|
81 |
+
|
82 |
+
def euler_step(self, x, d, dt):
|
83 |
+
return x + dt * d
|
84 |
+
|
85 |
+
|
86 |
+
class EDMSampler(SingleStepDiffusionSampler):
|
87 |
+
def __init__(
|
88 |
+
self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, *args, **kwargs
|
89 |
+
):
|
90 |
+
super().__init__(*args, **kwargs)
|
91 |
+
|
92 |
+
self.s_churn = s_churn
|
93 |
+
self.s_tmin = s_tmin
|
94 |
+
self.s_tmax = s_tmax
|
95 |
+
self.s_noise = s_noise
|
96 |
+
|
97 |
+
def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, gamma=0.0):
|
98 |
+
sigma_hat = sigma * (gamma + 1.0)
|
99 |
+
if gamma > 0:
|
100 |
+
eps = torch.randn_like(x) * self.s_noise
|
101 |
+
x = x + eps * append_dims(sigma_hat**2 - sigma**2, x.ndim) ** 0.5
|
102 |
+
|
103 |
+
denoised = self.denoise(x, denoiser, sigma_hat, cond, uc)
|
104 |
+
# print('denoised', denoised.mean(axis=[0, 2, 3]))
|
105 |
+
d = to_d(x, sigma_hat, denoised)
|
106 |
+
dt = append_dims(next_sigma - sigma_hat, x.ndim)
|
107 |
+
|
108 |
+
euler_step = self.euler_step(x, d, dt)
|
109 |
+
x = self.possible_correction_step(
|
110 |
+
euler_step, x, d, dt, next_sigma, denoiser, cond, uc
|
111 |
+
)
|
112 |
+
return x
|
113 |
+
|
114 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None):
|
115 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
116 |
+
x, cond, uc, num_steps
|
117 |
+
)
|
118 |
+
|
119 |
+
for i in self.get_sigma_gen(num_sigmas):
|
120 |
+
gamma = (
|
121 |
+
min(self.s_churn / (num_sigmas - 1), 2**0.5 - 1)
|
122 |
+
if self.s_tmin <= sigmas[i] <= self.s_tmax
|
123 |
+
else 0.0
|
124 |
+
)
|
125 |
+
x = self.sampler_step(
|
126 |
+
s_in * sigmas[i],
|
127 |
+
s_in * sigmas[i + 1],
|
128 |
+
denoiser,
|
129 |
+
x,
|
130 |
+
cond,
|
131 |
+
uc,
|
132 |
+
gamma,
|
133 |
+
)
|
134 |
+
|
135 |
+
return x
|
136 |
+
|
137 |
+
|
138 |
+
class AncestralSampler(SingleStepDiffusionSampler):
|
139 |
+
def __init__(self, eta=1.0, s_noise=1.0, *args, **kwargs):
|
140 |
+
super().__init__(*args, **kwargs)
|
141 |
+
|
142 |
+
self.eta = eta
|
143 |
+
self.s_noise = s_noise
|
144 |
+
self.noise_sampler = lambda x: torch.randn_like(x)
|
145 |
+
|
146 |
+
def ancestral_euler_step(self, x, denoised, sigma, sigma_down):
|
147 |
+
d = to_d(x, sigma, denoised)
|
148 |
+
dt = append_dims(sigma_down - sigma, x.ndim)
|
149 |
+
|
150 |
+
return self.euler_step(x, d, dt)
|
151 |
+
|
152 |
+
def ancestral_step(self, x, sigma, next_sigma, sigma_up):
|
153 |
+
x = torch.where(
|
154 |
+
append_dims(next_sigma, x.ndim) > 0.0,
|
155 |
+
x + self.noise_sampler(x) * self.s_noise * append_dims(sigma_up, x.ndim),
|
156 |
+
x,
|
157 |
+
)
|
158 |
+
return x
|
159 |
+
|
160 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None):
|
161 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
162 |
+
x, cond, uc, num_steps
|
163 |
+
)
|
164 |
+
|
165 |
+
for i in self.get_sigma_gen(num_sigmas):
|
166 |
+
x = self.sampler_step(
|
167 |
+
s_in * sigmas[i],
|
168 |
+
s_in * sigmas[i + 1],
|
169 |
+
denoiser,
|
170 |
+
x,
|
171 |
+
cond,
|
172 |
+
uc,
|
173 |
+
)
|
174 |
+
|
175 |
+
return x
|
176 |
+
|
177 |
+
|
178 |
+
class LinearMultistepSampler(BaseDiffusionSampler):
|
179 |
+
def __init__(
|
180 |
+
self,
|
181 |
+
order=4,
|
182 |
+
*args,
|
183 |
+
**kwargs,
|
184 |
+
):
|
185 |
+
super().__init__(*args, **kwargs)
|
186 |
+
|
187 |
+
self.order = order
|
188 |
+
|
189 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None, **kwargs):
|
190 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
191 |
+
x, cond, uc, num_steps
|
192 |
+
)
|
193 |
+
|
194 |
+
ds = []
|
195 |
+
sigmas_cpu = sigmas.detach().cpu().numpy()
|
196 |
+
for i in self.get_sigma_gen(num_sigmas):
|
197 |
+
sigma = s_in * sigmas[i]
|
198 |
+
denoised = denoiser(
|
199 |
+
*self.guider.prepare_inputs(x, sigma, cond, uc), **kwargs
|
200 |
+
)
|
201 |
+
denoised = self.guider(denoised, sigma)
|
202 |
+
d = to_d(x, sigma, denoised)
|
203 |
+
ds.append(d)
|
204 |
+
if len(ds) > self.order:
|
205 |
+
ds.pop(0)
|
206 |
+
cur_order = min(i + 1, self.order)
|
207 |
+
coeffs = [
|
208 |
+
linear_multistep_coeff(cur_order, sigmas_cpu, i, j)
|
209 |
+
for j in range(cur_order)
|
210 |
+
]
|
211 |
+
x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
|
212 |
+
|
213 |
+
return x
|
214 |
+
|
215 |
+
|
216 |
+
class EulerEDMSampler(EDMSampler):
|
217 |
+
def possible_correction_step(
|
218 |
+
self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc
|
219 |
+
):
|
220 |
+
# print("euler_step: ", euler_step.mean(axis=[0, 2, 3]))
|
221 |
+
return euler_step
|
222 |
+
|
223 |
+
|
224 |
+
class HeunEDMSampler(EDMSampler):
|
225 |
+
def possible_correction_step(
|
226 |
+
self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc
|
227 |
+
):
|
228 |
+
if torch.sum(next_sigma) < 1e-14:
|
229 |
+
# Save a network evaluation if all noise levels are 0
|
230 |
+
return euler_step
|
231 |
+
else:
|
232 |
+
denoised = self.denoise(euler_step, denoiser, next_sigma, cond, uc)
|
233 |
+
d_new = to_d(euler_step, next_sigma, denoised)
|
234 |
+
d_prime = (d + d_new) / 2.0
|
235 |
+
|
236 |
+
# apply correction if noise level is not 0
|
237 |
+
x = torch.where(
|
238 |
+
append_dims(next_sigma, x.ndim) > 0.0, x + d_prime * dt, euler_step
|
239 |
+
)
|
240 |
+
return x
|
241 |
+
|
242 |
+
|
243 |
+
class EulerAncestralSampler(AncestralSampler):
|
244 |
+
def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc):
|
245 |
+
sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta)
|
246 |
+
denoised = self.denoise(x, denoiser, sigma, cond, uc)
|
247 |
+
x = self.ancestral_euler_step(x, denoised, sigma, sigma_down)
|
248 |
+
x = self.ancestral_step(x, sigma, next_sigma, sigma_up)
|
249 |
+
|
250 |
+
return x
|
251 |
+
|
252 |
+
|
253 |
+
class DPMPP2SAncestralSampler(AncestralSampler):
|
254 |
+
def get_variables(self, sigma, sigma_down):
|
255 |
+
t, t_next = [to_neg_log_sigma(s) for s in (sigma, sigma_down)]
|
256 |
+
h = t_next - t
|
257 |
+
s = t + 0.5 * h
|
258 |
+
return h, s, t, t_next
|
259 |
+
|
260 |
+
def get_mult(self, h, s, t, t_next):
|
261 |
+
mult1 = to_sigma(s) / to_sigma(t)
|
262 |
+
mult2 = (-0.5 * h).expm1()
|
263 |
+
mult3 = to_sigma(t_next) / to_sigma(t)
|
264 |
+
mult4 = (-h).expm1()
|
265 |
+
|
266 |
+
return mult1, mult2, mult3, mult4
|
267 |
+
|
268 |
+
def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, **kwargs):
|
269 |
+
sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta)
|
270 |
+
denoised = self.denoise(x, denoiser, sigma, cond, uc)
|
271 |
+
x_euler = self.ancestral_euler_step(x, denoised, sigma, sigma_down)
|
272 |
+
|
273 |
+
if torch.sum(sigma_down) < 1e-14:
|
274 |
+
# Save a network evaluation if all noise levels are 0
|
275 |
+
x = x_euler
|
276 |
+
else:
|
277 |
+
h, s, t, t_next = self.get_variables(sigma, sigma_down)
|
278 |
+
mult = [
|
279 |
+
append_dims(mult, x.ndim) for mult in self.get_mult(h, s, t, t_next)
|
280 |
+
]
|
281 |
+
|
282 |
+
x2 = mult[0] * x - mult[1] * denoised
|
283 |
+
denoised2 = self.denoise(x2, denoiser, to_sigma(s), cond, uc)
|
284 |
+
x_dpmpp2s = mult[2] * x - mult[3] * denoised2
|
285 |
+
|
286 |
+
# apply correction if noise level is not 0
|
287 |
+
x = torch.where(append_dims(sigma_down, x.ndim) > 0.0, x_dpmpp2s, x_euler)
|
288 |
+
|
289 |
+
x = self.ancestral_step(x, sigma, next_sigma, sigma_up)
|
290 |
+
return x
|
291 |
+
|
292 |
+
|
293 |
+
class DPMPP2MSampler(BaseDiffusionSampler):
|
294 |
+
def get_variables(self, sigma, next_sigma, previous_sigma=None):
|
295 |
+
t, t_next = [to_neg_log_sigma(s) for s in (sigma, next_sigma)]
|
296 |
+
h = t_next - t
|
297 |
+
|
298 |
+
if previous_sigma is not None:
|
299 |
+
h_last = t - to_neg_log_sigma(previous_sigma)
|
300 |
+
r = h_last / h
|
301 |
+
return h, r, t, t_next
|
302 |
+
else:
|
303 |
+
return h, None, t, t_next
|
304 |
+
|
305 |
+
def get_mult(self, h, r, t, t_next, previous_sigma):
|
306 |
+
mult1 = to_sigma(t_next) / to_sigma(t)
|
307 |
+
mult2 = (-h).expm1()
|
308 |
+
|
309 |
+
if previous_sigma is not None:
|
310 |
+
mult3 = 1 + 1 / (2 * r)
|
311 |
+
mult4 = 1 / (2 * r)
|
312 |
+
return mult1, mult2, mult3, mult4
|
313 |
+
else:
|
314 |
+
return mult1, mult2
|
315 |
+
|
316 |
+
def sampler_step(
|
317 |
+
self,
|
318 |
+
old_denoised,
|
319 |
+
previous_sigma,
|
320 |
+
sigma,
|
321 |
+
next_sigma,
|
322 |
+
denoiser,
|
323 |
+
x,
|
324 |
+
cond,
|
325 |
+
uc=None,
|
326 |
+
):
|
327 |
+
denoised = self.denoise(x, denoiser, sigma, cond, uc)
|
328 |
+
|
329 |
+
h, r, t, t_next = self.get_variables(sigma, next_sigma, previous_sigma)
|
330 |
+
mult = [
|
331 |
+
append_dims(mult, x.ndim)
|
332 |
+
for mult in self.get_mult(h, r, t, t_next, previous_sigma)
|
333 |
+
]
|
334 |
+
|
335 |
+
x_standard = mult[0] * x - mult[1] * denoised
|
336 |
+
if old_denoised is None or torch.sum(next_sigma) < 1e-14:
|
337 |
+
# Save a network evaluation if all noise levels are 0 or on the first step
|
338 |
+
return x_standard, denoised
|
339 |
+
else:
|
340 |
+
denoised_d = mult[2] * denoised - mult[3] * old_denoised
|
341 |
+
x_advanced = mult[0] * x - mult[1] * denoised_d
|
342 |
+
|
343 |
+
# apply correction if noise level is not 0 and not first step
|
344 |
+
x = torch.where(
|
345 |
+
append_dims(next_sigma, x.ndim) > 0.0, x_advanced, x_standard
|
346 |
+
)
|
347 |
+
|
348 |
+
return x, denoised
|
349 |
+
|
350 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None, **kwargs):
|
351 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
352 |
+
x, cond, uc, num_steps
|
353 |
+
)
|
354 |
+
|
355 |
+
old_denoised = None
|
356 |
+
for i in self.get_sigma_gen(num_sigmas):
|
357 |
+
x, old_denoised = self.sampler_step(
|
358 |
+
old_denoised,
|
359 |
+
None if i == 0 else s_in * sigmas[i - 1],
|
360 |
+
s_in * sigmas[i],
|
361 |
+
s_in * sigmas[i + 1],
|
362 |
+
denoiser,
|
363 |
+
x,
|
364 |
+
cond,
|
365 |
+
uc=uc,
|
366 |
+
)
|
367 |
+
|
368 |
+
return x
|
369 |
+
|
370 |
+
|
371 |
+
class SubstepSampler(EulerAncestralSampler):
|
372 |
+
def __init__(self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, restore_cfg=4.0,
|
373 |
+
restore_cfg_s_tmin=0.05, eta=1., n_sample_steps=4, *args, **kwargs):
|
374 |
+
super().__init__(*args, **kwargs)
|
375 |
+
self.n_sample_steps = n_sample_steps
|
376 |
+
self.steps_subset = [0, 100, 200, 300, 1000]
|
377 |
+
|
378 |
+
def prepare_sampling_loop(self, x, cond, uc=None, num_steps=None):
|
379 |
+
sigmas = self.discretization(1000, device=self.device)
|
380 |
+
sigmas = sigmas[
|
381 |
+
self.steps_subset[: self.num_steps] + self.steps_subset[-1:]
|
382 |
+
]
|
383 |
+
print(sigmas)
|
384 |
+
# uc = cond
|
385 |
+
x *= torch.sqrt(1.0 + sigmas[0] ** 2.0)
|
386 |
+
num_sigmas = len(sigmas)
|
387 |
+
s_in = x.new_ones([x.shape[0]])
|
388 |
+
return x, s_in, sigmas, num_sigmas, cond, uc
|
389 |
+
|
390 |
+
def denoise(self, x, denoiser, sigma, cond, uc, control_scale=1.0):
|
391 |
+
denoised = denoiser(*self.guider.prepare_inputs(x, sigma, cond, uc), control_scale)
|
392 |
+
denoised = self.guider(denoised, sigma)
|
393 |
+
return denoised
|
394 |
+
|
395 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None, control_scale=1.0, *args, **kwargs):
|
396 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
397 |
+
x, cond, uc, num_steps
|
398 |
+
)
|
399 |
+
|
400 |
+
for i in self.get_sigma_gen(num_sigmas):
|
401 |
+
x = self.sampler_step(
|
402 |
+
s_in * sigmas[i],
|
403 |
+
s_in * sigmas[i + 1],
|
404 |
+
denoiser,
|
405 |
+
x,
|
406 |
+
cond,
|
407 |
+
uc,
|
408 |
+
control_scale=control_scale,
|
409 |
+
)
|
410 |
+
|
411 |
+
return x
|
412 |
+
|
413 |
+
def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc, control_scale=1.0):
|
414 |
+
sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta)
|
415 |
+
denoised = self.denoise(x, denoiser, sigma, cond, uc, control_scale=control_scale)
|
416 |
+
x = self.ancestral_euler_step(x, denoised, sigma, sigma_down)
|
417 |
+
x = self.ancestral_step(x, sigma, next_sigma, sigma_up)
|
418 |
+
|
419 |
+
return x
|
420 |
+
|
421 |
+
|
422 |
+
class RestoreDPMPP2MSampler(DPMPP2MSampler):
|
423 |
+
def __init__(self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, restore_cfg=4.0,
|
424 |
+
restore_cfg_s_tmin=0.05, eta=1., *args, **kwargs):
|
425 |
+
self.s_noise = s_noise
|
426 |
+
self.eta = eta
|
427 |
+
super().__init__(*args, **kwargs)
|
428 |
+
|
429 |
+
def denoise(self, x, denoiser, sigma, cond, uc, control_scale=1.0):
|
430 |
+
denoised = denoiser(*self.guider.prepare_inputs(x, sigma, cond, uc), control_scale)
|
431 |
+
denoised = self.guider(denoised, sigma)
|
432 |
+
return denoised
|
433 |
+
|
434 |
+
def get_mult(self, h, r, t, t_next, previous_sigma):
|
435 |
+
eta_h = self.eta * h
|
436 |
+
mult1 = to_sigma(t_next) / to_sigma(t) * (-eta_h).exp()
|
437 |
+
mult2 = (-h -eta_h).expm1()
|
438 |
+
|
439 |
+
if previous_sigma is not None:
|
440 |
+
mult3 = 1 + 1 / (2 * r)
|
441 |
+
mult4 = 1 / (2 * r)
|
442 |
+
return mult1, mult2, mult3, mult4
|
443 |
+
else:
|
444 |
+
return mult1, mult2
|
445 |
+
|
446 |
+
|
447 |
+
def sampler_step(
|
448 |
+
self,
|
449 |
+
old_denoised,
|
450 |
+
previous_sigma,
|
451 |
+
sigma,
|
452 |
+
next_sigma,
|
453 |
+
denoiser,
|
454 |
+
x,
|
455 |
+
cond,
|
456 |
+
uc=None,
|
457 |
+
eps_noise=None,
|
458 |
+
control_scale=1.0,
|
459 |
+
):
|
460 |
+
denoised = self.denoise(x, denoiser, sigma, cond, uc, control_scale=control_scale)
|
461 |
+
|
462 |
+
h, r, t, t_next = self.get_variables(sigma, next_sigma, previous_sigma)
|
463 |
+
eta_h = self.eta * h
|
464 |
+
mult = [
|
465 |
+
append_dims(mult, x.ndim)
|
466 |
+
for mult in self.get_mult(h, r, t, t_next, previous_sigma)
|
467 |
+
]
|
468 |
+
|
469 |
+
x_standard = mult[0] * x - mult[1] * denoised
|
470 |
+
if old_denoised is None or torch.sum(next_sigma) < 1e-14:
|
471 |
+
# Save a network evaluation if all noise levels are 0 or on the first step
|
472 |
+
return x_standard, denoised
|
473 |
+
else:
|
474 |
+
denoised_d = mult[2] * denoised - mult[3] * old_denoised
|
475 |
+
x_advanced = mult[0] * x - mult[1] * denoised_d
|
476 |
+
|
477 |
+
# apply correction if noise level is not 0 and not first step
|
478 |
+
x = torch.where(
|
479 |
+
append_dims(next_sigma, x.ndim) > 0.0, x_advanced, x_standard
|
480 |
+
)
|
481 |
+
if self.eta:
|
482 |
+
x = x + eps_noise * next_sigma * (-2 * eta_h).expm1().neg().sqrt() * self.s_noise
|
483 |
+
|
484 |
+
return x, denoised
|
485 |
+
|
486 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None, control_scale=1.0, **kwargs):
|
487 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
488 |
+
x, cond, uc, num_steps
|
489 |
+
)
|
490 |
+
sigmas_min, sigmas_max = sigmas[-2].cpu(), sigmas[0].cpu()
|
491 |
+
sigmas_new = get_sigmas_karras(self.num_steps, sigmas_min, sigmas_max, device=x.device)
|
492 |
+
sigmas = sigmas_new
|
493 |
+
|
494 |
+
noise_sampler = BrownianTreeNoiseSampler(x, sigmas_min, sigmas_max)
|
495 |
+
|
496 |
+
old_denoised = None
|
497 |
+
for i in self.get_sigma_gen(num_sigmas):
|
498 |
+
if i > 0 and torch.sum(s_in * sigmas[i + 1]) > 1e-14:
|
499 |
+
eps_noise = noise_sampler(s_in * sigmas[i], s_in * sigmas[i + 1])
|
500 |
+
else:
|
501 |
+
eps_noise = None
|
502 |
+
x, old_denoised = self.sampler_step(
|
503 |
+
old_denoised,
|
504 |
+
None if i == 0 else s_in * sigmas[i - 1],
|
505 |
+
s_in * sigmas[i],
|
506 |
+
s_in * sigmas[i + 1],
|
507 |
+
denoiser,
|
508 |
+
x,
|
509 |
+
cond,
|
510 |
+
uc=uc,
|
511 |
+
eps_noise=eps_noise,
|
512 |
+
control_scale=control_scale,
|
513 |
+
)
|
514 |
+
|
515 |
+
return x
|
516 |
+
|
517 |
+
|
518 |
+
def to_d_center(denoised, x_center, x):
|
519 |
+
b = denoised.shape[0]
|
520 |
+
v_center = (denoised - x_center).view(b, -1)
|
521 |
+
v_denoise = (x - denoised).view(b, -1)
|
522 |
+
d_center = v_center - v_denoise * (v_center * v_denoise).sum(dim=1).view(b, 1) / \
|
523 |
+
(v_denoise * v_denoise).sum(dim=1).view(b, 1)
|
524 |
+
d_center = d_center / d_center.view(x.shape[0], -1).norm(dim=1).view(-1, 1)
|
525 |
+
return d_center.view(denoised.shape)
|
526 |
+
|
527 |
+
|
528 |
+
class RestoreEDMSampler(SingleStepDiffusionSampler):
|
529 |
+
def __init__(
|
530 |
+
self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, restore_cfg=4.0,
|
531 |
+
restore_cfg_s_tmin=0.05, *args, **kwargs
|
532 |
+
):
|
533 |
+
super().__init__(*args, **kwargs)
|
534 |
+
|
535 |
+
self.s_churn = s_churn
|
536 |
+
self.s_tmin = s_tmin
|
537 |
+
self.s_tmax = s_tmax
|
538 |
+
self.s_noise = s_noise
|
539 |
+
self.restore_cfg = restore_cfg
|
540 |
+
self.restore_cfg_s_tmin = restore_cfg_s_tmin
|
541 |
+
self.sigma_max = 14.6146
|
542 |
+
|
543 |
+
def denoise(self, x, denoiser, sigma, cond, uc, control_scale=1.0):
|
544 |
+
denoised = denoiser(*self.guider.prepare_inputs(x, sigma, cond, uc), control_scale)
|
545 |
+
denoised = self.guider(denoised, sigma)
|
546 |
+
return denoised
|
547 |
+
|
548 |
+
def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, gamma=0.0, x_center=None, eps_noise=None,
|
549 |
+
control_scale=1.0, use_linear_control_scale=False, control_scale_start=0.0):
|
550 |
+
sigma_hat = sigma * (gamma + 1.0)
|
551 |
+
if gamma > 0:
|
552 |
+
if eps_noise is not None:
|
553 |
+
eps = eps_noise * self.s_noise
|
554 |
+
else:
|
555 |
+
eps = torch.randn_like(x) * self.s_noise
|
556 |
+
x = x + eps * append_dims(sigma_hat**2 - sigma**2, x.ndim) ** 0.5
|
557 |
+
|
558 |
+
if use_linear_control_scale:
|
559 |
+
control_scale = (sigma[0].item() / self.sigma_max) * (control_scale_start - control_scale) + control_scale
|
560 |
+
|
561 |
+
denoised = self.denoise(x, denoiser, sigma_hat, cond, uc, control_scale=control_scale)
|
562 |
+
|
563 |
+
if (next_sigma[0] > self.restore_cfg_s_tmin) and (self.restore_cfg > 0):
|
564 |
+
d_center = (denoised - x_center)
|
565 |
+
denoised = denoised - d_center * ((sigma.view(-1, 1, 1, 1) / self.sigma_max) ** self.restore_cfg)
|
566 |
+
|
567 |
+
d = to_d(x, sigma_hat, denoised)
|
568 |
+
dt = append_dims(next_sigma - sigma_hat, x.ndim)
|
569 |
+
x = self.euler_step(x, d, dt)
|
570 |
+
return x
|
571 |
+
|
572 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None, x_center=None, control_scale=1.0,
|
573 |
+
use_linear_control_scale=False, control_scale_start=0.0):
|
574 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
575 |
+
x, cond, uc, num_steps
|
576 |
+
)
|
577 |
+
|
578 |
+
for _idx, i in enumerate(self.get_sigma_gen(num_sigmas)):
|
579 |
+
gamma = (
|
580 |
+
min(self.s_churn / (num_sigmas - 1), 2**0.5 - 1)
|
581 |
+
if self.s_tmin <= sigmas[i] <= self.s_tmax
|
582 |
+
else 0.0
|
583 |
+
)
|
584 |
+
x = self.sampler_step(
|
585 |
+
s_in * sigmas[i],
|
586 |
+
s_in * sigmas[i + 1],
|
587 |
+
denoiser,
|
588 |
+
x,
|
589 |
+
cond,
|
590 |
+
uc,
|
591 |
+
gamma,
|
592 |
+
x_center,
|
593 |
+
control_scale=control_scale,
|
594 |
+
use_linear_control_scale=use_linear_control_scale,
|
595 |
+
control_scale_start=control_scale_start,
|
596 |
+
)
|
597 |
+
return x
|
598 |
+
|
599 |
+
|
600 |
+
class TiledRestoreEDMSampler(RestoreEDMSampler):
|
601 |
+
def __init__(self, tile_size=128, tile_stride=64, *args, **kwargs):
|
602 |
+
super().__init__(*args, **kwargs)
|
603 |
+
self.tile_size = tile_size
|
604 |
+
self.tile_stride = tile_stride
|
605 |
+
self.tile_weights = gaussian_weights(self.tile_size, self.tile_size, 1)
|
606 |
+
|
607 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None, x_center=None, control_scale=1.0,
|
608 |
+
use_linear_control_scale=False, control_scale_start=0.0):
|
609 |
+
use_local_prompt = isinstance(cond, list)
|
610 |
+
b, _, h, w = x.shape
|
611 |
+
latent_tiles_iterator = _sliding_windows(h, w, self.tile_size, self.tile_stride)
|
612 |
+
tile_weights = self.tile_weights.repeat(b, 1, 1, 1)
|
613 |
+
if not use_local_prompt:
|
614 |
+
LQ_latent = cond['control']
|
615 |
+
else:
|
616 |
+
assert len(cond) == len(latent_tiles_iterator), "Number of local prompts should be equal to number of tiles"
|
617 |
+
LQ_latent = cond[0]['control']
|
618 |
+
clean_LQ_latent = x_center
|
619 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
620 |
+
x, cond, uc, num_steps
|
621 |
+
)
|
622 |
+
|
623 |
+
for _idx, i in enumerate(self.get_sigma_gen(num_sigmas)):
|
624 |
+
gamma = (
|
625 |
+
min(self.s_churn / (num_sigmas - 1), 2**0.5 - 1)
|
626 |
+
if self.s_tmin <= sigmas[i] <= self.s_tmax
|
627 |
+
else 0.0
|
628 |
+
)
|
629 |
+
x_next = torch.zeros_like(x)
|
630 |
+
count = torch.zeros_like(x)
|
631 |
+
eps_noise = torch.randn_like(x)
|
632 |
+
for j, (hi, hi_end, wi, wi_end) in enumerate(latent_tiles_iterator):
|
633 |
+
x_tile = x[:, :, hi:hi_end, wi:wi_end]
|
634 |
+
_eps_noise = eps_noise[:, :, hi:hi_end, wi:wi_end]
|
635 |
+
x_center_tile = clean_LQ_latent[:, :, hi:hi_end, wi:wi_end]
|
636 |
+
if use_local_prompt:
|
637 |
+
_cond = cond[j]
|
638 |
+
else:
|
639 |
+
_cond = cond
|
640 |
+
_cond['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
|
641 |
+
uc['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
|
642 |
+
_x = self.sampler_step(
|
643 |
+
s_in * sigmas[i],
|
644 |
+
s_in * sigmas[i + 1],
|
645 |
+
denoiser,
|
646 |
+
x_tile,
|
647 |
+
_cond,
|
648 |
+
uc,
|
649 |
+
gamma,
|
650 |
+
x_center_tile,
|
651 |
+
eps_noise=_eps_noise,
|
652 |
+
control_scale=control_scale,
|
653 |
+
use_linear_control_scale=use_linear_control_scale,
|
654 |
+
control_scale_start=control_scale_start,
|
655 |
+
)
|
656 |
+
x_next[:, :, hi:hi_end, wi:wi_end] += _x * tile_weights
|
657 |
+
count[:, :, hi:hi_end, wi:wi_end] += tile_weights
|
658 |
+
x_next /= count
|
659 |
+
x = x_next
|
660 |
+
return x
|
661 |
+
|
662 |
+
|
663 |
+
class TiledRestoreDPMPP2MSampler(RestoreDPMPP2MSampler):
|
664 |
+
def __init__(self, tile_size=128, tile_stride=64, *args, **kwargs):
|
665 |
+
super().__init__(*args, **kwargs)
|
666 |
+
self.tile_size = tile_size
|
667 |
+
self.tile_stride = tile_stride
|
668 |
+
self.tile_weights = gaussian_weights(self.tile_size, self.tile_size, 1)
|
669 |
+
|
670 |
+
def __call__(self, denoiser, x, cond, uc=None, num_steps=None, control_scale=1.0, **kwargs):
|
671 |
+
use_local_prompt = isinstance(cond, list)
|
672 |
+
b, _, h, w = x.shape
|
673 |
+
latent_tiles_iterator = _sliding_windows(h, w, self.tile_size, self.tile_stride)
|
674 |
+
tile_weights = self.tile_weights.repeat(b, 1, 1, 1)
|
675 |
+
if not use_local_prompt:
|
676 |
+
LQ_latent = cond['control']
|
677 |
+
else:
|
678 |
+
assert len(cond) == len(latent_tiles_iterator), "Number of local prompts should be equal to number of tiles"
|
679 |
+
LQ_latent = cond[0]['control']
|
680 |
+
x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
|
681 |
+
x, cond, uc, num_steps
|
682 |
+
)
|
683 |
+
sigmas_min, sigmas_max = sigmas[-2].cpu(), sigmas[0].cpu()
|
684 |
+
sigmas_new = get_sigmas_karras(self.num_steps, sigmas_min, sigmas_max, device=x.device)
|
685 |
+
sigmas = sigmas_new
|
686 |
+
|
687 |
+
noise_sampler = BrownianTreeNoiseSampler(x, sigmas_min, sigmas_max)
|
688 |
+
|
689 |
+
old_denoised = None
|
690 |
+
for _idx, i in enumerate(self.get_sigma_gen(num_sigmas)):
|
691 |
+
if i > 0 and torch.sum(s_in * sigmas[i + 1]) > 1e-14:
|
692 |
+
eps_noise = noise_sampler(s_in * sigmas[i], s_in * sigmas[i + 1])
|
693 |
+
else:
|
694 |
+
eps_noise = torch.zeros_like(x)
|
695 |
+
x_next = torch.zeros_like(x)
|
696 |
+
old_denoised_next = torch.zeros_like(x)
|
697 |
+
count = torch.zeros_like(x)
|
698 |
+
for j, (hi, hi_end, wi, wi_end) in enumerate(latent_tiles_iterator):
|
699 |
+
x_tile = x[:, :, hi:hi_end, wi:wi_end]
|
700 |
+
_eps_noise = eps_noise[:, :, hi:hi_end, wi:wi_end]
|
701 |
+
if old_denoised is not None:
|
702 |
+
old_denoised_tile = old_denoised[:, :, hi:hi_end, wi:wi_end]
|
703 |
+
else:
|
704 |
+
old_denoised_tile = None
|
705 |
+
if use_local_prompt:
|
706 |
+
_cond = cond[j]
|
707 |
+
else:
|
708 |
+
_cond = cond
|
709 |
+
_cond['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
|
710 |
+
uc['control'] = LQ_latent[:, :, hi:hi_end, wi:wi_end]
|
711 |
+
_x, _old_denoised = self.sampler_step(
|
712 |
+
old_denoised_tile,
|
713 |
+
None if i == 0 else s_in * sigmas[i - 1],
|
714 |
+
s_in * sigmas[i],
|
715 |
+
s_in * sigmas[i + 1],
|
716 |
+
denoiser,
|
717 |
+
x_tile,
|
718 |
+
_cond,
|
719 |
+
uc=uc,
|
720 |
+
eps_noise=_eps_noise,
|
721 |
+
control_scale=control_scale,
|
722 |
+
)
|
723 |
+
x_next[:, :, hi:hi_end, wi:wi_end] += _x * tile_weights
|
724 |
+
old_denoised_next[:, :, hi:hi_end, wi:wi_end] += _old_denoised * tile_weights
|
725 |
+
count[:, :, hi:hi_end, wi:wi_end] += tile_weights
|
726 |
+
old_denoised_next /= count
|
727 |
+
x_next /= count
|
728 |
+
x = x_next
|
729 |
+
old_denoised = old_denoised_next
|
730 |
+
return x
|
731 |
+
|
732 |
+
|
733 |
+
def gaussian_weights(tile_width, tile_height, nbatches):
|
734 |
+
"""Generates a gaussian mask of weights for tile contributions"""
|
735 |
+
from numpy import pi, exp, sqrt
|
736 |
+
import numpy as np
|
737 |
+
|
738 |
+
latent_width = tile_width
|
739 |
+
latent_height = tile_height
|
740 |
+
|
741 |
+
var = 0.01
|
742 |
+
midpoint = (latent_width - 1) / 2 # -1 because index goes from 0 to latent_width - 1
|
743 |
+
x_probs = [exp(-(x - midpoint) * (x - midpoint) / (latent_width * latent_width) / (2 * var)) / sqrt(2 * pi * var)
|
744 |
+
for x in range(latent_width)]
|
745 |
+
midpoint = latent_height / 2
|
746 |
+
y_probs = [exp(-(y - midpoint) * (y - midpoint) / (latent_height * latent_height) / (2 * var)) / sqrt(2 * pi * var)
|
747 |
+
for y in range(latent_height)]
|
748 |
+
|
749 |
+
weights = np.outer(y_probs, x_probs)
|
750 |
+
return torch.tile(torch.tensor(weights, device='cuda'), (nbatches, 4, 1, 1))
|
751 |
+
|
752 |
+
|
753 |
+
def _sliding_windows(h: int, w: int, tile_size: int, tile_stride: int):
|
754 |
+
hi_list = list(range(0, h - tile_size + 1, tile_stride))
|
755 |
+
if (h - tile_size) % tile_stride != 0:
|
756 |
+
hi_list.append(h - tile_size)
|
757 |
+
|
758 |
+
wi_list = list(range(0, w - tile_size + 1, tile_stride))
|
759 |
+
if (w - tile_size) % tile_stride != 0:
|
760 |
+
wi_list.append(w - tile_size)
|
761 |
+
|
762 |
+
coords = []
|
763 |
+
for hi in hi_list:
|
764 |
+
for wi in wi_list:
|
765 |
+
coords.append((hi, hi + tile_size, wi, wi + tile_size))
|
766 |
+
return coords
|
sgm/modules/diffusionmodules/sampling_utils.py
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from scipy import integrate
|
3 |
+
|
4 |
+
from ...util import append_dims
|
5 |
+
|
6 |
+
|
7 |
+
class NoDynamicThresholding:
|
8 |
+
def __call__(self, uncond, cond, scale):
|
9 |
+
return uncond + scale.view(-1, 1, 1, 1) * (cond - uncond)
|
10 |
+
|
11 |
+
|
12 |
+
def linear_multistep_coeff(order, t, i, j, epsrel=1e-4):
|
13 |
+
if order - 1 > i:
|
14 |
+
raise ValueError(f"Order {order} too high for step {i}")
|
15 |
+
|
16 |
+
def fn(tau):
|
17 |
+
prod = 1.0
|
18 |
+
for k in range(order):
|
19 |
+
if j == k:
|
20 |
+
continue
|
21 |
+
prod *= (tau - t[i - k]) / (t[i - j] - t[i - k])
|
22 |
+
return prod
|
23 |
+
|
24 |
+
return integrate.quad(fn, t[i], t[i + 1], epsrel=epsrel)[0]
|
25 |
+
|
26 |
+
|
27 |
+
def get_ancestral_step(sigma_from, sigma_to, eta=1.0):
|
28 |
+
if not eta:
|
29 |
+
return sigma_to, 0.0
|
30 |
+
sigma_up = torch.minimum(
|
31 |
+
sigma_to,
|
32 |
+
eta
|
33 |
+
* (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5,
|
34 |
+
)
|
35 |
+
sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5
|
36 |
+
return sigma_down, sigma_up
|
37 |
+
|
38 |
+
|
39 |
+
def to_d(x, sigma, denoised):
|
40 |
+
return (x - denoised) / append_dims(sigma, x.ndim)
|
41 |
+
|
42 |
+
|
43 |
+
def to_neg_log_sigma(sigma):
|
44 |
+
return sigma.log().neg()
|
45 |
+
|
46 |
+
|
47 |
+
def to_sigma(neg_log_sigma):
|
48 |
+
return neg_log_sigma.neg().exp()
|
sgm/modules/diffusionmodules/sigma_sampling.py
ADDED
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
from ...util import default, instantiate_from_config
|
4 |
+
|
5 |
+
|
6 |
+
class EDMSampling:
|
7 |
+
def __init__(self, p_mean=-1.2, p_std=1.2):
|
8 |
+
self.p_mean = p_mean
|
9 |
+
self.p_std = p_std
|
10 |
+
|
11 |
+
def __call__(self, n_samples, rand=None):
|
12 |
+
log_sigma = self.p_mean + self.p_std * default(rand, torch.randn((n_samples,)))
|
13 |
+
return log_sigma.exp()
|
14 |
+
|
15 |
+
|
16 |
+
class DiscreteSampling:
|
17 |
+
def __init__(self, discretization_config, num_idx, do_append_zero=False, flip=True, idx_range=None):
|
18 |
+
self.num_idx = num_idx
|
19 |
+
self.sigmas = instantiate_from_config(discretization_config)(
|
20 |
+
num_idx, do_append_zero=do_append_zero, flip=flip
|
21 |
+
)
|
22 |
+
self.idx_range = idx_range
|
23 |
+
|
24 |
+
def idx_to_sigma(self, idx):
|
25 |
+
# print(self.sigmas[idx])
|
26 |
+
return self.sigmas[idx]
|
27 |
+
|
28 |
+
def __call__(self, n_samples, rand=None):
|
29 |
+
if self.idx_range is None:
|
30 |
+
idx = default(
|
31 |
+
rand,
|
32 |
+
torch.randint(0, self.num_idx, (n_samples,)),
|
33 |
+
)
|
34 |
+
else:
|
35 |
+
idx = default(
|
36 |
+
rand,
|
37 |
+
torch.randint(self.idx_range[0], self.idx_range[1], (n_samples,)),
|
38 |
+
)
|
39 |
+
return self.idx_to_sigma(idx)
|
40 |
+
|
sgm/modules/diffusionmodules/util.py
ADDED
@@ -0,0 +1,309 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
adopted from
|
3 |
+
https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
|
4 |
+
and
|
5 |
+
https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
|
6 |
+
and
|
7 |
+
https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
|
8 |
+
|
9 |
+
thanks!
|
10 |
+
"""
|
11 |
+
|
12 |
+
import math
|
13 |
+
|
14 |
+
import torch
|
15 |
+
import torch.nn as nn
|
16 |
+
from einops import repeat
|
17 |
+
|
18 |
+
|
19 |
+
def make_beta_schedule(
|
20 |
+
schedule,
|
21 |
+
n_timestep,
|
22 |
+
linear_start=1e-4,
|
23 |
+
linear_end=2e-2,
|
24 |
+
):
|
25 |
+
if schedule == "linear":
|
26 |
+
betas = (
|
27 |
+
torch.linspace(
|
28 |
+
linear_start**0.5, linear_end**0.5, n_timestep, dtype=torch.float64
|
29 |
+
)
|
30 |
+
** 2
|
31 |
+
)
|
32 |
+
return betas.numpy()
|
33 |
+
|
34 |
+
|
35 |
+
def extract_into_tensor(a, t, x_shape):
|
36 |
+
b, *_ = t.shape
|
37 |
+
out = a.gather(-1, t)
|
38 |
+
return out.reshape(b, *((1,) * (len(x_shape) - 1)))
|
39 |
+
|
40 |
+
|
41 |
+
def mixed_checkpoint(func, inputs: dict, params, flag):
|
42 |
+
"""
|
43 |
+
Evaluate a function without caching intermediate activations, allowing for
|
44 |
+
reduced memory at the expense of extra compute in the backward pass. This differs from the original checkpoint function
|
45 |
+
borrowed from https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py in that
|
46 |
+
it also works with non-tensor inputs
|
47 |
+
:param func: the function to evaluate.
|
48 |
+
:param inputs: the argument dictionary to pass to `func`.
|
49 |
+
:param params: a sequence of parameters `func` depends on but does not
|
50 |
+
explicitly take as arguments.
|
51 |
+
:param flag: if False, disable gradient checkpointing.
|
52 |
+
"""
|
53 |
+
if flag:
|
54 |
+
tensor_keys = [key for key in inputs if isinstance(inputs[key], torch.Tensor)]
|
55 |
+
tensor_inputs = [
|
56 |
+
inputs[key] for key in inputs if isinstance(inputs[key], torch.Tensor)
|
57 |
+
]
|
58 |
+
non_tensor_keys = [
|
59 |
+
key for key in inputs if not isinstance(inputs[key], torch.Tensor)
|
60 |
+
]
|
61 |
+
non_tensor_inputs = [
|
62 |
+
inputs[key] for key in inputs if not isinstance(inputs[key], torch.Tensor)
|
63 |
+
]
|
64 |
+
args = tuple(tensor_inputs) + tuple(non_tensor_inputs) + tuple(params)
|
65 |
+
return MixedCheckpointFunction.apply(
|
66 |
+
func,
|
67 |
+
len(tensor_inputs),
|
68 |
+
len(non_tensor_inputs),
|
69 |
+
tensor_keys,
|
70 |
+
non_tensor_keys,
|
71 |
+
*args,
|
72 |
+
)
|
73 |
+
else:
|
74 |
+
return func(**inputs)
|
75 |
+
|
76 |
+
|
77 |
+
class MixedCheckpointFunction(torch.autograd.Function):
|
78 |
+
@staticmethod
|
79 |
+
def forward(
|
80 |
+
ctx,
|
81 |
+
run_function,
|
82 |
+
length_tensors,
|
83 |
+
length_non_tensors,
|
84 |
+
tensor_keys,
|
85 |
+
non_tensor_keys,
|
86 |
+
*args,
|
87 |
+
):
|
88 |
+
ctx.end_tensors = length_tensors
|
89 |
+
ctx.end_non_tensors = length_tensors + length_non_tensors
|
90 |
+
ctx.gpu_autocast_kwargs = {
|
91 |
+
"enabled": torch.is_autocast_enabled(),
|
92 |
+
"dtype": torch.get_autocast_gpu_dtype(),
|
93 |
+
"cache_enabled": torch.is_autocast_cache_enabled(),
|
94 |
+
}
|
95 |
+
assert (
|
96 |
+
len(tensor_keys) == length_tensors
|
97 |
+
and len(non_tensor_keys) == length_non_tensors
|
98 |
+
)
|
99 |
+
|
100 |
+
ctx.input_tensors = {
|
101 |
+
key: val for (key, val) in zip(tensor_keys, list(args[: ctx.end_tensors]))
|
102 |
+
}
|
103 |
+
ctx.input_non_tensors = {
|
104 |
+
key: val
|
105 |
+
for (key, val) in zip(
|
106 |
+
non_tensor_keys, list(args[ctx.end_tensors : ctx.end_non_tensors])
|
107 |
+
)
|
108 |
+
}
|
109 |
+
ctx.run_function = run_function
|
110 |
+
ctx.input_params = list(args[ctx.end_non_tensors :])
|
111 |
+
|
112 |
+
with torch.no_grad():
|
113 |
+
output_tensors = ctx.run_function(
|
114 |
+
**ctx.input_tensors, **ctx.input_non_tensors
|
115 |
+
)
|
116 |
+
return output_tensors
|
117 |
+
|
118 |
+
@staticmethod
|
119 |
+
def backward(ctx, *output_grads):
|
120 |
+
# additional_args = {key: ctx.input_tensors[key] for key in ctx.input_tensors if not isinstance(ctx.input_tensors[key],torch.Tensor)}
|
121 |
+
ctx.input_tensors = {
|
122 |
+
key: ctx.input_tensors[key].detach().requires_grad_(True)
|
123 |
+
for key in ctx.input_tensors
|
124 |
+
}
|
125 |
+
|
126 |
+
with torch.enable_grad(), torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs):
|
127 |
+
# Fixes a bug where the first op in run_function modifies the
|
128 |
+
# Tensor storage in place, which is not allowed for detach()'d
|
129 |
+
# Tensors.
|
130 |
+
shallow_copies = {
|
131 |
+
key: ctx.input_tensors[key].view_as(ctx.input_tensors[key])
|
132 |
+
for key in ctx.input_tensors
|
133 |
+
}
|
134 |
+
# shallow_copies.update(additional_args)
|
135 |
+
output_tensors = ctx.run_function(**shallow_copies, **ctx.input_non_tensors)
|
136 |
+
input_grads = torch.autograd.grad(
|
137 |
+
output_tensors,
|
138 |
+
list(ctx.input_tensors.values()) + ctx.input_params,
|
139 |
+
output_grads,
|
140 |
+
allow_unused=True,
|
141 |
+
)
|
142 |
+
del ctx.input_tensors
|
143 |
+
del ctx.input_params
|
144 |
+
del output_tensors
|
145 |
+
return (
|
146 |
+
(None, None, None, None, None)
|
147 |
+
+ input_grads[: ctx.end_tensors]
|
148 |
+
+ (None,) * (ctx.end_non_tensors - ctx.end_tensors)
|
149 |
+
+ input_grads[ctx.end_tensors :]
|
150 |
+
)
|
151 |
+
|
152 |
+
|
153 |
+
def checkpoint(func, inputs, params, flag):
|
154 |
+
"""
|
155 |
+
Evaluate a function without caching intermediate activations, allowing for
|
156 |
+
reduced memory at the expense of extra compute in the backward pass.
|
157 |
+
:param func: the function to evaluate.
|
158 |
+
:param inputs: the argument sequence to pass to `func`.
|
159 |
+
:param params: a sequence of parameters `func` depends on but does not
|
160 |
+
explicitly take as arguments.
|
161 |
+
:param flag: if False, disable gradient checkpointing.
|
162 |
+
"""
|
163 |
+
if flag:
|
164 |
+
args = tuple(inputs) + tuple(params)
|
165 |
+
return CheckpointFunction.apply(func, len(inputs), *args)
|
166 |
+
else:
|
167 |
+
return func(*inputs)
|
168 |
+
|
169 |
+
|
170 |
+
class CheckpointFunction(torch.autograd.Function):
|
171 |
+
@staticmethod
|
172 |
+
def forward(ctx, run_function, length, *args):
|
173 |
+
ctx.run_function = run_function
|
174 |
+
ctx.input_tensors = list(args[:length])
|
175 |
+
ctx.input_params = list(args[length:])
|
176 |
+
ctx.gpu_autocast_kwargs = {
|
177 |
+
"enabled": torch.is_autocast_enabled(),
|
178 |
+
"dtype": torch.get_autocast_gpu_dtype(),
|
179 |
+
"cache_enabled": torch.is_autocast_cache_enabled(),
|
180 |
+
}
|
181 |
+
with torch.no_grad():
|
182 |
+
output_tensors = ctx.run_function(*ctx.input_tensors)
|
183 |
+
return output_tensors
|
184 |
+
|
185 |
+
@staticmethod
|
186 |
+
def backward(ctx, *output_grads):
|
187 |
+
ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
|
188 |
+
with torch.enable_grad(), torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs):
|
189 |
+
# Fixes a bug where the first op in run_function modifies the
|
190 |
+
# Tensor storage in place, which is not allowed for detach()'d
|
191 |
+
# Tensors.
|
192 |
+
shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
|
193 |
+
output_tensors = ctx.run_function(*shallow_copies)
|
194 |
+
input_grads = torch.autograd.grad(
|
195 |
+
output_tensors,
|
196 |
+
ctx.input_tensors + ctx.input_params,
|
197 |
+
output_grads,
|
198 |
+
allow_unused=True,
|
199 |
+
)
|
200 |
+
del ctx.input_tensors
|
201 |
+
del ctx.input_params
|
202 |
+
del output_tensors
|
203 |
+
return (None, None) + input_grads
|
204 |
+
|
205 |
+
|
206 |
+
def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
|
207 |
+
"""
|
208 |
+
Create sinusoidal timestep embeddings.
|
209 |
+
:param timesteps: a 1-D Tensor of N indices, one per batch element.
|
210 |
+
These may be fractional.
|
211 |
+
:param dim: the dimension of the output.
|
212 |
+
:param max_period: controls the minimum frequency of the embeddings.
|
213 |
+
:return: an [N x dim] Tensor of positional embeddings.
|
214 |
+
"""
|
215 |
+
if not repeat_only:
|
216 |
+
half = dim // 2
|
217 |
+
freqs = torch.exp(
|
218 |
+
-math.log(max_period)
|
219 |
+
* torch.arange(start=0, end=half, dtype=torch.float32)
|
220 |
+
/ half
|
221 |
+
).to(device=timesteps.device)
|
222 |
+
args = timesteps[:, None].float() * freqs[None]
|
223 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
224 |
+
if dim % 2:
|
225 |
+
embedding = torch.cat(
|
226 |
+
[embedding, torch.zeros_like(embedding[:, :1])], dim=-1
|
227 |
+
)
|
228 |
+
else:
|
229 |
+
embedding = repeat(timesteps, "b -> b d", d=dim)
|
230 |
+
return embedding
|
231 |
+
|
232 |
+
|
233 |
+
def zero_module(module):
|
234 |
+
"""
|
235 |
+
Zero out the parameters of a module and return it.
|
236 |
+
"""
|
237 |
+
for p in module.parameters():
|
238 |
+
p.detach().zero_()
|
239 |
+
return module
|
240 |
+
|
241 |
+
|
242 |
+
def scale_module(module, scale):
|
243 |
+
"""
|
244 |
+
Scale the parameters of a module and return it.
|
245 |
+
"""
|
246 |
+
for p in module.parameters():
|
247 |
+
p.detach().mul_(scale)
|
248 |
+
return module
|
249 |
+
|
250 |
+
|
251 |
+
def mean_flat(tensor):
|
252 |
+
"""
|
253 |
+
Take the mean over all non-batch dimensions.
|
254 |
+
"""
|
255 |
+
return tensor.mean(dim=list(range(1, len(tensor.shape))))
|
256 |
+
|
257 |
+
|
258 |
+
def normalization(channels):
|
259 |
+
"""
|
260 |
+
Make a standard normalization layer.
|
261 |
+
:param channels: number of input channels.
|
262 |
+
:return: an nn.Module for normalization.
|
263 |
+
"""
|
264 |
+
return GroupNorm32(32, channels)
|
265 |
+
|
266 |
+
|
267 |
+
# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
|
268 |
+
class SiLU(nn.Module):
|
269 |
+
def forward(self, x):
|
270 |
+
return x * torch.sigmoid(x)
|
271 |
+
|
272 |
+
|
273 |
+
class GroupNorm32(nn.GroupNorm):
|
274 |
+
def forward(self, x):
|
275 |
+
# return super().forward(x.float()).type(x.dtype)
|
276 |
+
return super().forward(x)
|
277 |
+
|
278 |
+
|
279 |
+
def conv_nd(dims, *args, **kwargs):
|
280 |
+
"""
|
281 |
+
Create a 1D, 2D, or 3D convolution module.
|
282 |
+
"""
|
283 |
+
if dims == 1:
|
284 |
+
return nn.Conv1d(*args, **kwargs)
|
285 |
+
elif dims == 2:
|
286 |
+
return nn.Conv2d(*args, **kwargs)
|
287 |
+
elif dims == 3:
|
288 |
+
return nn.Conv3d(*args, **kwargs)
|
289 |
+
raise ValueError(f"unsupported dimensions: {dims}")
|
290 |
+
|
291 |
+
|
292 |
+
def linear(*args, **kwargs):
|
293 |
+
"""
|
294 |
+
Create a linear module.
|
295 |
+
"""
|
296 |
+
return nn.Linear(*args, **kwargs)
|
297 |
+
|
298 |
+
|
299 |
+
def avg_pool_nd(dims, *args, **kwargs):
|
300 |
+
"""
|
301 |
+
Create a 1D, 2D, or 3D average pooling module.
|
302 |
+
"""
|
303 |
+
if dims == 1:
|
304 |
+
return nn.AvgPool1d(*args, **kwargs)
|
305 |
+
elif dims == 2:
|
306 |
+
return nn.AvgPool2d(*args, **kwargs)
|
307 |
+
elif dims == 3:
|
308 |
+
return nn.AvgPool3d(*args, **kwargs)
|
309 |
+
raise ValueError(f"unsupported dimensions: {dims}")
|
sgm/modules/diffusionmodules/wrappers.py
ADDED
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from packaging import version
|
4 |
+
# import torch._dynamo
|
5 |
+
# torch._dynamo.config.suppress_errors = True
|
6 |
+
# torch._dynamo.config.cache_size_limit = 512
|
7 |
+
|
8 |
+
OPENAIUNETWRAPPER = "sgm.modules.diffusionmodules.wrappers.OpenAIWrapper"
|
9 |
+
|
10 |
+
|
11 |
+
class IdentityWrapper(nn.Module):
|
12 |
+
def __init__(self, diffusion_model, compile_model: bool = False):
|
13 |
+
super().__init__()
|
14 |
+
compile = (
|
15 |
+
torch.compile
|
16 |
+
if (version.parse(torch.__version__) >= version.parse("2.0.0"))
|
17 |
+
and compile_model
|
18 |
+
else lambda x: x
|
19 |
+
)
|
20 |
+
self.diffusion_model = compile(diffusion_model)
|
21 |
+
|
22 |
+
def forward(self, *args, **kwargs):
|
23 |
+
return self.diffusion_model(*args, **kwargs)
|
24 |
+
|
25 |
+
|
26 |
+
class OpenAIWrapper(IdentityWrapper):
|
27 |
+
def forward(
|
28 |
+
self, x: torch.Tensor, t: torch.Tensor, c: dict, **kwargs
|
29 |
+
) -> torch.Tensor:
|
30 |
+
x = torch.cat((x, c.get("concat", torch.Tensor([]).type_as(x))), dim=1)
|
31 |
+
return self.diffusion_model(
|
32 |
+
x,
|
33 |
+
timesteps=t,
|
34 |
+
context=c.get("crossattn", None),
|
35 |
+
y=c.get("vector", None),
|
36 |
+
**kwargs,
|
37 |
+
)
|
38 |
+
|
39 |
+
|
40 |
+
class OpenAIHalfWrapper(IdentityWrapper):
|
41 |
+
def __init__(self, *args, **kwargs):
|
42 |
+
super().__init__(*args, **kwargs)
|
43 |
+
self.diffusion_model = self.diffusion_model.half()
|
44 |
+
|
45 |
+
def forward(
|
46 |
+
self, x: torch.Tensor, t: torch.Tensor, c: dict, **kwargs
|
47 |
+
) -> torch.Tensor:
|
48 |
+
x = torch.cat((x, c.get("concat", torch.Tensor([]).type_as(x))), dim=1)
|
49 |
+
_context = c.get("crossattn", None)
|
50 |
+
_y = c.get("vector", None)
|
51 |
+
if _context is not None:
|
52 |
+
_context = _context.half()
|
53 |
+
if _y is not None:
|
54 |
+
_y = _y.half()
|
55 |
+
x = x.half()
|
56 |
+
t = t.half()
|
57 |
+
|
58 |
+
out = self.diffusion_model(
|
59 |
+
x,
|
60 |
+
timesteps=t,
|
61 |
+
context=_context,
|
62 |
+
y=_y,
|
63 |
+
**kwargs,
|
64 |
+
)
|
65 |
+
return out.float()
|
66 |
+
|
67 |
+
|
68 |
+
class ControlWrapper(nn.Module):
|
69 |
+
def __init__(self, diffusion_model, compile_model: bool = False, dtype=torch.float32):
|
70 |
+
super().__init__()
|
71 |
+
self.compile = (
|
72 |
+
torch.compile
|
73 |
+
if (version.parse(torch.__version__) >= version.parse("2.0.0"))
|
74 |
+
and compile_model
|
75 |
+
else lambda x: x
|
76 |
+
)
|
77 |
+
self.diffusion_model = self.compile(diffusion_model)
|
78 |
+
self.control_model = None
|
79 |
+
self.dtype = dtype
|
80 |
+
|
81 |
+
def load_control_model(self, control_model):
|
82 |
+
self.control_model = self.compile(control_model)
|
83 |
+
|
84 |
+
def forward(
|
85 |
+
self, x: torch.Tensor, t: torch.Tensor, c: dict, control_scale=1, **kwargs
|
86 |
+
) -> torch.Tensor:
|
87 |
+
with torch.autocast("cuda", dtype=self.dtype):
|
88 |
+
control = self.control_model(x=c.get("control", None), timesteps=t, xt=x,
|
89 |
+
control_vector=c.get("control_vector", None),
|
90 |
+
mask_x=c.get("mask_x", None),
|
91 |
+
context=c.get("crossattn", None),
|
92 |
+
y=c.get("vector", None))
|
93 |
+
out = self.diffusion_model(
|
94 |
+
x,
|
95 |
+
timesteps=t,
|
96 |
+
context=c.get("crossattn", None),
|
97 |
+
y=c.get("vector", None),
|
98 |
+
control=control,
|
99 |
+
control_scale=control_scale,
|
100 |
+
**kwargs,
|
101 |
+
)
|
102 |
+
return out.float()
|
103 |
+
|