Spaces:
Runtime error
Runtime error
anonymous
commited on
Commit
•
4fcfd85
1
Parent(s):
fd216ce
add controlnet
Browse files- ControlNet/ldm/data/__init__.py +0 -0
- ControlNet/ldm/data/util.py +24 -0
- ControlNet/ldm/models/autoencoder.py +219 -0
- ControlNet/ldm/models/diffusion/__init__.py +0 -0
- ControlNet/ldm/models/diffusion/ddim.py +336 -0
- ControlNet/ldm/models/diffusion/ddpm.py +1797 -0
- ControlNet/ldm/models/diffusion/dpm_solver/__init__.py +1 -0
- ControlNet/ldm/models/diffusion/dpm_solver/dpm_solver.py +1154 -0
- ControlNet/ldm/models/diffusion/dpm_solver/sampler.py +87 -0
- ControlNet/ldm/models/diffusion/plms.py +244 -0
- ControlNet/ldm/models/diffusion/sampling_util.py +22 -0
ControlNet/ldm/data/__init__.py
ADDED
File without changes
|
ControlNet/ldm/data/util.py
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
from ldm.modules.midas.api import load_midas_transform
|
4 |
+
|
5 |
+
|
6 |
+
class AddMiDaS(object):
|
7 |
+
def __init__(self, model_type):
|
8 |
+
super().__init__()
|
9 |
+
self.transform = load_midas_transform(model_type)
|
10 |
+
|
11 |
+
def pt2np(self, x):
|
12 |
+
x = ((x + 1.0) * .5).detach().cpu().numpy()
|
13 |
+
return x
|
14 |
+
|
15 |
+
def np2pt(self, x):
|
16 |
+
x = torch.from_numpy(x) * 2 - 1.
|
17 |
+
return x
|
18 |
+
|
19 |
+
def __call__(self, sample):
|
20 |
+
# sample['jpg'] is tensor hwc in [-1, 1] at this point
|
21 |
+
x = self.pt2np(sample['jpg'])
|
22 |
+
x = self.transform({"image": x})["image"]
|
23 |
+
sample['midas_in'] = x
|
24 |
+
return sample
|
ControlNet/ldm/models/autoencoder.py
ADDED
@@ -0,0 +1,219 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import pytorch_lightning as pl
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from contextlib import contextmanager
|
5 |
+
|
6 |
+
from ldm.modules.diffusionmodules.model import Encoder, Decoder
|
7 |
+
from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
|
8 |
+
|
9 |
+
from ldm.util import instantiate_from_config
|
10 |
+
from ldm.modules.ema import LitEma
|
11 |
+
|
12 |
+
|
13 |
+
class AutoencoderKL(pl.LightningModule):
|
14 |
+
def __init__(self,
|
15 |
+
ddconfig,
|
16 |
+
lossconfig,
|
17 |
+
embed_dim,
|
18 |
+
ckpt_path=None,
|
19 |
+
ignore_keys=[],
|
20 |
+
image_key="image",
|
21 |
+
colorize_nlabels=None,
|
22 |
+
monitor=None,
|
23 |
+
ema_decay=None,
|
24 |
+
learn_logvar=False
|
25 |
+
):
|
26 |
+
super().__init__()
|
27 |
+
self.learn_logvar = learn_logvar
|
28 |
+
self.image_key = image_key
|
29 |
+
self.encoder = Encoder(**ddconfig)
|
30 |
+
self.decoder = Decoder(**ddconfig)
|
31 |
+
self.loss = instantiate_from_config(lossconfig)
|
32 |
+
assert ddconfig["double_z"]
|
33 |
+
self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
|
34 |
+
self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
|
35 |
+
self.embed_dim = embed_dim
|
36 |
+
if colorize_nlabels is not None:
|
37 |
+
assert type(colorize_nlabels)==int
|
38 |
+
self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
|
39 |
+
if monitor is not None:
|
40 |
+
self.monitor = monitor
|
41 |
+
|
42 |
+
self.use_ema = ema_decay is not None
|
43 |
+
if self.use_ema:
|
44 |
+
self.ema_decay = ema_decay
|
45 |
+
assert 0. < ema_decay < 1.
|
46 |
+
self.model_ema = LitEma(self, decay=ema_decay)
|
47 |
+
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
|
48 |
+
|
49 |
+
if ckpt_path is not None:
|
50 |
+
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
|
51 |
+
|
52 |
+
def init_from_ckpt(self, path, ignore_keys=list()):
|
53 |
+
sd = torch.load(path, map_location="cpu")["state_dict"]
|
54 |
+
keys = list(sd.keys())
|
55 |
+
for k in keys:
|
56 |
+
for ik in ignore_keys:
|
57 |
+
if k.startswith(ik):
|
58 |
+
print("Deleting key {} from state_dict.".format(k))
|
59 |
+
del sd[k]
|
60 |
+
self.load_state_dict(sd, strict=False)
|
61 |
+
print(f"Restored from {path}")
|
62 |
+
|
63 |
+
@contextmanager
|
64 |
+
def ema_scope(self, context=None):
|
65 |
+
if self.use_ema:
|
66 |
+
self.model_ema.store(self.parameters())
|
67 |
+
self.model_ema.copy_to(self)
|
68 |
+
if context is not None:
|
69 |
+
print(f"{context}: Switched to EMA weights")
|
70 |
+
try:
|
71 |
+
yield None
|
72 |
+
finally:
|
73 |
+
if self.use_ema:
|
74 |
+
self.model_ema.restore(self.parameters())
|
75 |
+
if context is not None:
|
76 |
+
print(f"{context}: Restored training weights")
|
77 |
+
|
78 |
+
def on_train_batch_end(self, *args, **kwargs):
|
79 |
+
if self.use_ema:
|
80 |
+
self.model_ema(self)
|
81 |
+
|
82 |
+
def encode(self, x):
|
83 |
+
h = self.encoder(x)
|
84 |
+
moments = self.quant_conv(h)
|
85 |
+
posterior = DiagonalGaussianDistribution(moments)
|
86 |
+
return posterior
|
87 |
+
|
88 |
+
def decode(self, z):
|
89 |
+
z = self.post_quant_conv(z)
|
90 |
+
dec = self.decoder(z)
|
91 |
+
return dec
|
92 |
+
|
93 |
+
def forward(self, input, sample_posterior=True):
|
94 |
+
posterior = self.encode(input)
|
95 |
+
if sample_posterior:
|
96 |
+
z = posterior.sample()
|
97 |
+
else:
|
98 |
+
z = posterior.mode()
|
99 |
+
dec = self.decode(z)
|
100 |
+
return dec, posterior
|
101 |
+
|
102 |
+
def get_input(self, batch, k):
|
103 |
+
x = batch[k]
|
104 |
+
if len(x.shape) == 3:
|
105 |
+
x = x[..., None]
|
106 |
+
x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
|
107 |
+
return x
|
108 |
+
|
109 |
+
def training_step(self, batch, batch_idx, optimizer_idx):
|
110 |
+
inputs = self.get_input(batch, self.image_key)
|
111 |
+
reconstructions, posterior = self(inputs)
|
112 |
+
|
113 |
+
if optimizer_idx == 0:
|
114 |
+
# train encoder+decoder+logvar
|
115 |
+
aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
|
116 |
+
last_layer=self.get_last_layer(), split="train")
|
117 |
+
self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
|
118 |
+
self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
|
119 |
+
return aeloss
|
120 |
+
|
121 |
+
if optimizer_idx == 1:
|
122 |
+
# train the discriminator
|
123 |
+
discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
|
124 |
+
last_layer=self.get_last_layer(), split="train")
|
125 |
+
|
126 |
+
self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
|
127 |
+
self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
|
128 |
+
return discloss
|
129 |
+
|
130 |
+
def validation_step(self, batch, batch_idx):
|
131 |
+
log_dict = self._validation_step(batch, batch_idx)
|
132 |
+
with self.ema_scope():
|
133 |
+
log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema")
|
134 |
+
return log_dict
|
135 |
+
|
136 |
+
def _validation_step(self, batch, batch_idx, postfix=""):
|
137 |
+
inputs = self.get_input(batch, self.image_key)
|
138 |
+
reconstructions, posterior = self(inputs)
|
139 |
+
aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
|
140 |
+
last_layer=self.get_last_layer(), split="val"+postfix)
|
141 |
+
|
142 |
+
discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
|
143 |
+
last_layer=self.get_last_layer(), split="val"+postfix)
|
144 |
+
|
145 |
+
self.log(f"val{postfix}/rec_loss", log_dict_ae[f"val{postfix}/rec_loss"])
|
146 |
+
self.log_dict(log_dict_ae)
|
147 |
+
self.log_dict(log_dict_disc)
|
148 |
+
return self.log_dict
|
149 |
+
|
150 |
+
def configure_optimizers(self):
|
151 |
+
lr = self.learning_rate
|
152 |
+
ae_params_list = list(self.encoder.parameters()) + list(self.decoder.parameters()) + list(
|
153 |
+
self.quant_conv.parameters()) + list(self.post_quant_conv.parameters())
|
154 |
+
if self.learn_logvar:
|
155 |
+
print(f"{self.__class__.__name__}: Learning logvar")
|
156 |
+
ae_params_list.append(self.loss.logvar)
|
157 |
+
opt_ae = torch.optim.Adam(ae_params_list,
|
158 |
+
lr=lr, betas=(0.5, 0.9))
|
159 |
+
opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
|
160 |
+
lr=lr, betas=(0.5, 0.9))
|
161 |
+
return [opt_ae, opt_disc], []
|
162 |
+
|
163 |
+
def get_last_layer(self):
|
164 |
+
return self.decoder.conv_out.weight
|
165 |
+
|
166 |
+
@torch.no_grad()
|
167 |
+
def log_images(self, batch, only_inputs=False, log_ema=False, **kwargs):
|
168 |
+
log = dict()
|
169 |
+
x = self.get_input(batch, self.image_key)
|
170 |
+
x = x.to(self.device)
|
171 |
+
if not only_inputs:
|
172 |
+
xrec, posterior = self(x)
|
173 |
+
if x.shape[1] > 3:
|
174 |
+
# colorize with random projection
|
175 |
+
assert xrec.shape[1] > 3
|
176 |
+
x = self.to_rgb(x)
|
177 |
+
xrec = self.to_rgb(xrec)
|
178 |
+
log["samples"] = self.decode(torch.randn_like(posterior.sample()))
|
179 |
+
log["reconstructions"] = xrec
|
180 |
+
if log_ema or self.use_ema:
|
181 |
+
with self.ema_scope():
|
182 |
+
xrec_ema, posterior_ema = self(x)
|
183 |
+
if x.shape[1] > 3:
|
184 |
+
# colorize with random projection
|
185 |
+
assert xrec_ema.shape[1] > 3
|
186 |
+
xrec_ema = self.to_rgb(xrec_ema)
|
187 |
+
log["samples_ema"] = self.decode(torch.randn_like(posterior_ema.sample()))
|
188 |
+
log["reconstructions_ema"] = xrec_ema
|
189 |
+
log["inputs"] = x
|
190 |
+
return log
|
191 |
+
|
192 |
+
def to_rgb(self, x):
|
193 |
+
assert self.image_key == "segmentation"
|
194 |
+
if not hasattr(self, "colorize"):
|
195 |
+
self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
|
196 |
+
x = F.conv2d(x, weight=self.colorize)
|
197 |
+
x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
|
198 |
+
return x
|
199 |
+
|
200 |
+
|
201 |
+
class IdentityFirstStage(torch.nn.Module):
|
202 |
+
def __init__(self, *args, vq_interface=False, **kwargs):
|
203 |
+
self.vq_interface = vq_interface
|
204 |
+
super().__init__()
|
205 |
+
|
206 |
+
def encode(self, x, *args, **kwargs):
|
207 |
+
return x
|
208 |
+
|
209 |
+
def decode(self, x, *args, **kwargs):
|
210 |
+
return x
|
211 |
+
|
212 |
+
def quantize(self, x, *args, **kwargs):
|
213 |
+
if self.vq_interface:
|
214 |
+
return x, None, [None, None, None]
|
215 |
+
return x
|
216 |
+
|
217 |
+
def forward(self, x, *args, **kwargs):
|
218 |
+
return x
|
219 |
+
|
ControlNet/ldm/models/diffusion/__init__.py
ADDED
File without changes
|
ControlNet/ldm/models/diffusion/ddim.py
ADDED
@@ -0,0 +1,336 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""SAMPLING ONLY."""
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import numpy as np
|
5 |
+
from tqdm import tqdm
|
6 |
+
|
7 |
+
from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor
|
8 |
+
|
9 |
+
|
10 |
+
class DDIMSampler(object):
|
11 |
+
def __init__(self, model, schedule="linear", **kwargs):
|
12 |
+
super().__init__()
|
13 |
+
self.model = model
|
14 |
+
self.ddpm_num_timesteps = model.num_timesteps
|
15 |
+
self.schedule = schedule
|
16 |
+
|
17 |
+
def register_buffer(self, name, attr):
|
18 |
+
if type(attr) == torch.Tensor:
|
19 |
+
if attr.device != torch.device("cuda"):
|
20 |
+
attr = attr.to(torch.device("cuda"))
|
21 |
+
setattr(self, name, attr)
|
22 |
+
|
23 |
+
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
|
24 |
+
self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
|
25 |
+
num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
|
26 |
+
alphas_cumprod = self.model.alphas_cumprod
|
27 |
+
assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
|
28 |
+
to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
|
29 |
+
|
30 |
+
self.register_buffer('betas', to_torch(self.model.betas))
|
31 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
32 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
|
33 |
+
|
34 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
35 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
|
36 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
|
37 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
|
38 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
|
39 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
|
40 |
+
|
41 |
+
# ddim sampling parameters
|
42 |
+
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
|
43 |
+
ddim_timesteps=self.ddim_timesteps,
|
44 |
+
eta=ddim_eta,verbose=verbose)
|
45 |
+
self.register_buffer('ddim_sigmas', ddim_sigmas)
|
46 |
+
self.register_buffer('ddim_alphas', ddim_alphas)
|
47 |
+
self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
|
48 |
+
self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
|
49 |
+
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
|
50 |
+
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
|
51 |
+
1 - self.alphas_cumprod / self.alphas_cumprod_prev))
|
52 |
+
self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
|
53 |
+
|
54 |
+
@torch.no_grad()
|
55 |
+
def sample(self,
|
56 |
+
S,
|
57 |
+
batch_size,
|
58 |
+
shape,
|
59 |
+
conditioning=None,
|
60 |
+
callback=None,
|
61 |
+
normals_sequence=None,
|
62 |
+
img_callback=None,
|
63 |
+
quantize_x0=False,
|
64 |
+
eta=0.,
|
65 |
+
mask=None,
|
66 |
+
x0=None,
|
67 |
+
temperature=1.,
|
68 |
+
noise_dropout=0.,
|
69 |
+
score_corrector=None,
|
70 |
+
corrector_kwargs=None,
|
71 |
+
verbose=True,
|
72 |
+
x_T=None,
|
73 |
+
log_every_t=100,
|
74 |
+
unconditional_guidance_scale=1.,
|
75 |
+
unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
|
76 |
+
dynamic_threshold=None,
|
77 |
+
ucg_schedule=None,
|
78 |
+
**kwargs
|
79 |
+
):
|
80 |
+
if conditioning is not None:
|
81 |
+
if isinstance(conditioning, dict):
|
82 |
+
ctmp = conditioning[list(conditioning.keys())[0]]
|
83 |
+
while isinstance(ctmp, list): ctmp = ctmp[0]
|
84 |
+
cbs = ctmp.shape[0]
|
85 |
+
if cbs != batch_size:
|
86 |
+
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
|
87 |
+
|
88 |
+
elif isinstance(conditioning, list):
|
89 |
+
for ctmp in conditioning:
|
90 |
+
if ctmp.shape[0] != batch_size:
|
91 |
+
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
|
92 |
+
|
93 |
+
else:
|
94 |
+
if conditioning.shape[0] != batch_size:
|
95 |
+
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
|
96 |
+
|
97 |
+
self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
|
98 |
+
# sampling
|
99 |
+
C, H, W = shape
|
100 |
+
size = (batch_size, C, H, W)
|
101 |
+
print(f'Data shape for DDIM sampling is {size}, eta {eta}')
|
102 |
+
|
103 |
+
samples, intermediates = self.ddim_sampling(conditioning, size,
|
104 |
+
callback=callback,
|
105 |
+
img_callback=img_callback,
|
106 |
+
quantize_denoised=quantize_x0,
|
107 |
+
mask=mask, x0=x0,
|
108 |
+
ddim_use_original_steps=False,
|
109 |
+
noise_dropout=noise_dropout,
|
110 |
+
temperature=temperature,
|
111 |
+
score_corrector=score_corrector,
|
112 |
+
corrector_kwargs=corrector_kwargs,
|
113 |
+
x_T=x_T,
|
114 |
+
log_every_t=log_every_t,
|
115 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
116 |
+
unconditional_conditioning=unconditional_conditioning,
|
117 |
+
dynamic_threshold=dynamic_threshold,
|
118 |
+
ucg_schedule=ucg_schedule
|
119 |
+
)
|
120 |
+
return samples, intermediates
|
121 |
+
|
122 |
+
@torch.no_grad()
|
123 |
+
def ddim_sampling(self, cond, shape,
|
124 |
+
x_T=None, ddim_use_original_steps=False,
|
125 |
+
callback=None, timesteps=None, quantize_denoised=False,
|
126 |
+
mask=None, x0=None, img_callback=None, log_every_t=100,
|
127 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
|
128 |
+
unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None,
|
129 |
+
ucg_schedule=None):
|
130 |
+
device = self.model.betas.device
|
131 |
+
b = shape[0]
|
132 |
+
if x_T is None:
|
133 |
+
img = torch.randn(shape, device=device)
|
134 |
+
else:
|
135 |
+
img = x_T
|
136 |
+
|
137 |
+
if timesteps is None:
|
138 |
+
timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
|
139 |
+
elif timesteps is not None and not ddim_use_original_steps:
|
140 |
+
subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
|
141 |
+
timesteps = self.ddim_timesteps[:subset_end]
|
142 |
+
|
143 |
+
intermediates = {'x_inter': [img], 'pred_x0': [img]}
|
144 |
+
time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
|
145 |
+
total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
|
146 |
+
print(f"Running DDIM Sampling with {total_steps} timesteps")
|
147 |
+
|
148 |
+
iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
|
149 |
+
|
150 |
+
for i, step in enumerate(iterator):
|
151 |
+
index = total_steps - i - 1
|
152 |
+
ts = torch.full((b,), step, device=device, dtype=torch.long)
|
153 |
+
|
154 |
+
if mask is not None:
|
155 |
+
assert x0 is not None
|
156 |
+
img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
|
157 |
+
img = img_orig * mask + (1. - mask) * img
|
158 |
+
|
159 |
+
if ucg_schedule is not None:
|
160 |
+
assert len(ucg_schedule) == len(time_range)
|
161 |
+
unconditional_guidance_scale = ucg_schedule[i]
|
162 |
+
|
163 |
+
outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
|
164 |
+
quantize_denoised=quantize_denoised, temperature=temperature,
|
165 |
+
noise_dropout=noise_dropout, score_corrector=score_corrector,
|
166 |
+
corrector_kwargs=corrector_kwargs,
|
167 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
168 |
+
unconditional_conditioning=unconditional_conditioning,
|
169 |
+
dynamic_threshold=dynamic_threshold)
|
170 |
+
img, pred_x0 = outs
|
171 |
+
if callback: callback(i)
|
172 |
+
if img_callback: img_callback(pred_x0, i)
|
173 |
+
|
174 |
+
if index % log_every_t == 0 or index == total_steps - 1:
|
175 |
+
intermediates['x_inter'].append(img)
|
176 |
+
intermediates['pred_x0'].append(pred_x0)
|
177 |
+
|
178 |
+
return img, intermediates
|
179 |
+
|
180 |
+
@torch.no_grad()
|
181 |
+
def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
|
182 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
|
183 |
+
unconditional_guidance_scale=1., unconditional_conditioning=None,
|
184 |
+
dynamic_threshold=None):
|
185 |
+
b, *_, device = *x.shape, x.device
|
186 |
+
|
187 |
+
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
|
188 |
+
model_output = self.model.apply_model(x, t, c)
|
189 |
+
else:
|
190 |
+
x_in = torch.cat([x] * 2)
|
191 |
+
t_in = torch.cat([t] * 2)
|
192 |
+
if isinstance(c, dict):
|
193 |
+
assert isinstance(unconditional_conditioning, dict)
|
194 |
+
c_in = dict()
|
195 |
+
for k in c:
|
196 |
+
if isinstance(c[k], list):
|
197 |
+
c_in[k] = [torch.cat([
|
198 |
+
unconditional_conditioning[k][i],
|
199 |
+
c[k][i]]) for i in range(len(c[k]))]
|
200 |
+
else:
|
201 |
+
c_in[k] = torch.cat([
|
202 |
+
unconditional_conditioning[k],
|
203 |
+
c[k]])
|
204 |
+
elif isinstance(c, list):
|
205 |
+
c_in = list()
|
206 |
+
assert isinstance(unconditional_conditioning, list)
|
207 |
+
for i in range(len(c)):
|
208 |
+
c_in.append(torch.cat([unconditional_conditioning[i], c[i]]))
|
209 |
+
else:
|
210 |
+
c_in = torch.cat([unconditional_conditioning, c])
|
211 |
+
model_uncond, model_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
|
212 |
+
model_output = model_uncond + unconditional_guidance_scale * (model_t - model_uncond)
|
213 |
+
|
214 |
+
if self.model.parameterization == "v":
|
215 |
+
e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)
|
216 |
+
else:
|
217 |
+
e_t = model_output
|
218 |
+
|
219 |
+
if score_corrector is not None:
|
220 |
+
assert self.model.parameterization == "eps", 'not implemented'
|
221 |
+
e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
|
222 |
+
|
223 |
+
alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
|
224 |
+
alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
|
225 |
+
sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
|
226 |
+
sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
|
227 |
+
# select parameters corresponding to the currently considered timestep
|
228 |
+
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
|
229 |
+
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
|
230 |
+
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
|
231 |
+
sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
|
232 |
+
|
233 |
+
# current prediction for x_0
|
234 |
+
if self.model.parameterization != "v":
|
235 |
+
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
|
236 |
+
else:
|
237 |
+
pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
|
238 |
+
|
239 |
+
if quantize_denoised:
|
240 |
+
pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
|
241 |
+
|
242 |
+
if dynamic_threshold is not None:
|
243 |
+
raise NotImplementedError()
|
244 |
+
|
245 |
+
# direction pointing to x_t
|
246 |
+
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
|
247 |
+
noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
|
248 |
+
if noise_dropout > 0.:
|
249 |
+
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
250 |
+
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
|
251 |
+
return x_prev, pred_x0
|
252 |
+
|
253 |
+
@torch.no_grad()
|
254 |
+
def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None,
|
255 |
+
unconditional_guidance_scale=1.0, unconditional_conditioning=None, callback=None):
|
256 |
+
num_reference_steps = self.ddpm_num_timesteps if use_original_steps else self.ddim_timesteps.shape[0]
|
257 |
+
|
258 |
+
assert t_enc <= num_reference_steps
|
259 |
+
num_steps = t_enc
|
260 |
+
|
261 |
+
if use_original_steps:
|
262 |
+
alphas_next = self.alphas_cumprod[:num_steps]
|
263 |
+
alphas = self.alphas_cumprod_prev[:num_steps]
|
264 |
+
else:
|
265 |
+
alphas_next = self.ddim_alphas[:num_steps]
|
266 |
+
alphas = torch.tensor(self.ddim_alphas_prev[:num_steps])
|
267 |
+
|
268 |
+
x_next = x0
|
269 |
+
intermediates = []
|
270 |
+
inter_steps = []
|
271 |
+
for i in tqdm(range(num_steps), desc='Encoding Image'):
|
272 |
+
t = torch.full((x0.shape[0],), i, device=self.model.device, dtype=torch.long)
|
273 |
+
if unconditional_guidance_scale == 1.:
|
274 |
+
noise_pred = self.model.apply_model(x_next, t, c)
|
275 |
+
else:
|
276 |
+
assert unconditional_conditioning is not None
|
277 |
+
e_t_uncond, noise_pred = torch.chunk(
|
278 |
+
self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)),
|
279 |
+
torch.cat((unconditional_conditioning, c))), 2)
|
280 |
+
noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond)
|
281 |
+
|
282 |
+
xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next
|
283 |
+
weighted_noise_pred = alphas_next[i].sqrt() * (
|
284 |
+
(1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred
|
285 |
+
x_next = xt_weighted + weighted_noise_pred
|
286 |
+
if return_intermediates and i % (
|
287 |
+
num_steps // return_intermediates) == 0 and i < num_steps - 1:
|
288 |
+
intermediates.append(x_next)
|
289 |
+
inter_steps.append(i)
|
290 |
+
elif return_intermediates and i >= num_steps - 2:
|
291 |
+
intermediates.append(x_next)
|
292 |
+
inter_steps.append(i)
|
293 |
+
if callback: callback(i)
|
294 |
+
|
295 |
+
out = {'x_encoded': x_next, 'intermediate_steps': inter_steps}
|
296 |
+
if return_intermediates:
|
297 |
+
out.update({'intermediates': intermediates})
|
298 |
+
return x_next, out
|
299 |
+
|
300 |
+
@torch.no_grad()
|
301 |
+
def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
|
302 |
+
# fast, but does not allow for exact reconstruction
|
303 |
+
# t serves as an index to gather the correct alphas
|
304 |
+
if use_original_steps:
|
305 |
+
sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
|
306 |
+
sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
|
307 |
+
else:
|
308 |
+
sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
|
309 |
+
sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
|
310 |
+
|
311 |
+
if noise is None:
|
312 |
+
noise = torch.randn_like(x0)
|
313 |
+
return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
|
314 |
+
extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
|
315 |
+
|
316 |
+
@torch.no_grad()
|
317 |
+
def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
|
318 |
+
use_original_steps=False, callback=None):
|
319 |
+
|
320 |
+
timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
|
321 |
+
timesteps = timesteps[:t_start]
|
322 |
+
|
323 |
+
time_range = np.flip(timesteps)
|
324 |
+
total_steps = timesteps.shape[0]
|
325 |
+
print(f"Running DDIM Sampling with {total_steps} timesteps")
|
326 |
+
|
327 |
+
iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
|
328 |
+
x_dec = x_latent
|
329 |
+
for i, step in enumerate(iterator):
|
330 |
+
index = total_steps - i - 1
|
331 |
+
ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
|
332 |
+
x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
|
333 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
334 |
+
unconditional_conditioning=unconditional_conditioning)
|
335 |
+
if callback: callback(i)
|
336 |
+
return x_dec
|
ControlNet/ldm/models/diffusion/ddpm.py
ADDED
@@ -0,0 +1,1797 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
wild mixture of
|
3 |
+
https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
|
4 |
+
https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
|
5 |
+
https://github.com/CompVis/taming-transformers
|
6 |
+
-- merci
|
7 |
+
"""
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
import numpy as np
|
12 |
+
import pytorch_lightning as pl
|
13 |
+
from torch.optim.lr_scheduler import LambdaLR
|
14 |
+
from einops import rearrange, repeat
|
15 |
+
from contextlib import contextmanager, nullcontext
|
16 |
+
from functools import partial
|
17 |
+
import itertools
|
18 |
+
from tqdm import tqdm
|
19 |
+
from torchvision.utils import make_grid
|
20 |
+
from pytorch_lightning.utilities.distributed import rank_zero_only
|
21 |
+
from omegaconf import ListConfig
|
22 |
+
|
23 |
+
from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
|
24 |
+
from ldm.modules.ema import LitEma
|
25 |
+
from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
|
26 |
+
from ldm.models.autoencoder import IdentityFirstStage, AutoencoderKL
|
27 |
+
from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
|
28 |
+
from ldm.models.diffusion.ddim import DDIMSampler
|
29 |
+
|
30 |
+
|
31 |
+
__conditioning_keys__ = {'concat': 'c_concat',
|
32 |
+
'crossattn': 'c_crossattn',
|
33 |
+
'adm': 'y'}
|
34 |
+
|
35 |
+
|
36 |
+
def disabled_train(self, mode=True):
|
37 |
+
"""Overwrite model.train with this function to make sure train/eval mode
|
38 |
+
does not change anymore."""
|
39 |
+
return self
|
40 |
+
|
41 |
+
|
42 |
+
def uniform_on_device(r1, r2, shape, device):
|
43 |
+
return (r1 - r2) * torch.rand(*shape, device=device) + r2
|
44 |
+
|
45 |
+
|
46 |
+
class DDPM(pl.LightningModule):
|
47 |
+
# classic DDPM with Gaussian diffusion, in image space
|
48 |
+
def __init__(self,
|
49 |
+
unet_config,
|
50 |
+
timesteps=1000,
|
51 |
+
beta_schedule="linear",
|
52 |
+
loss_type="l2",
|
53 |
+
ckpt_path=None,
|
54 |
+
ignore_keys=[],
|
55 |
+
load_only_unet=False,
|
56 |
+
monitor="val/loss",
|
57 |
+
use_ema=True,
|
58 |
+
first_stage_key="image",
|
59 |
+
image_size=256,
|
60 |
+
channels=3,
|
61 |
+
log_every_t=100,
|
62 |
+
clip_denoised=True,
|
63 |
+
linear_start=1e-4,
|
64 |
+
linear_end=2e-2,
|
65 |
+
cosine_s=8e-3,
|
66 |
+
given_betas=None,
|
67 |
+
original_elbo_weight=0.,
|
68 |
+
v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
|
69 |
+
l_simple_weight=1.,
|
70 |
+
conditioning_key=None,
|
71 |
+
parameterization="eps", # all assuming fixed variance schedules
|
72 |
+
scheduler_config=None,
|
73 |
+
use_positional_encodings=False,
|
74 |
+
learn_logvar=False,
|
75 |
+
logvar_init=0.,
|
76 |
+
make_it_fit=False,
|
77 |
+
ucg_training=None,
|
78 |
+
reset_ema=False,
|
79 |
+
reset_num_ema_updates=False,
|
80 |
+
):
|
81 |
+
super().__init__()
|
82 |
+
assert parameterization in ["eps", "x0", "v"], 'currently only supporting "eps" and "x0" and "v"'
|
83 |
+
self.parameterization = parameterization
|
84 |
+
print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
|
85 |
+
self.cond_stage_model = None
|
86 |
+
self.clip_denoised = clip_denoised
|
87 |
+
self.log_every_t = log_every_t
|
88 |
+
self.first_stage_key = first_stage_key
|
89 |
+
self.image_size = image_size # try conv?
|
90 |
+
self.channels = channels
|
91 |
+
self.use_positional_encodings = use_positional_encodings
|
92 |
+
self.model = DiffusionWrapper(unet_config, conditioning_key)
|
93 |
+
count_params(self.model, verbose=True)
|
94 |
+
self.use_ema = use_ema
|
95 |
+
if self.use_ema:
|
96 |
+
self.model_ema = LitEma(self.model)
|
97 |
+
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
|
98 |
+
|
99 |
+
self.use_scheduler = scheduler_config is not None
|
100 |
+
if self.use_scheduler:
|
101 |
+
self.scheduler_config = scheduler_config
|
102 |
+
|
103 |
+
self.v_posterior = v_posterior
|
104 |
+
self.original_elbo_weight = original_elbo_weight
|
105 |
+
self.l_simple_weight = l_simple_weight
|
106 |
+
|
107 |
+
if monitor is not None:
|
108 |
+
self.monitor = monitor
|
109 |
+
self.make_it_fit = make_it_fit
|
110 |
+
if reset_ema: assert exists(ckpt_path)
|
111 |
+
if ckpt_path is not None:
|
112 |
+
self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
|
113 |
+
if reset_ema:
|
114 |
+
assert self.use_ema
|
115 |
+
print(f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.")
|
116 |
+
self.model_ema = LitEma(self.model)
|
117 |
+
if reset_num_ema_updates:
|
118 |
+
print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ")
|
119 |
+
assert self.use_ema
|
120 |
+
self.model_ema.reset_num_updates()
|
121 |
+
|
122 |
+
self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
|
123 |
+
linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
|
124 |
+
|
125 |
+
self.loss_type = loss_type
|
126 |
+
|
127 |
+
self.learn_logvar = learn_logvar
|
128 |
+
logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
|
129 |
+
if self.learn_logvar:
|
130 |
+
self.logvar = nn.Parameter(self.logvar, requires_grad=True)
|
131 |
+
else:
|
132 |
+
self.register_buffer('logvar', logvar)
|
133 |
+
|
134 |
+
self.ucg_training = ucg_training or dict()
|
135 |
+
if self.ucg_training:
|
136 |
+
self.ucg_prng = np.random.RandomState()
|
137 |
+
|
138 |
+
def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
|
139 |
+
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
140 |
+
if exists(given_betas):
|
141 |
+
betas = given_betas
|
142 |
+
else:
|
143 |
+
betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
|
144 |
+
cosine_s=cosine_s)
|
145 |
+
alphas = 1. - betas
|
146 |
+
alphas_cumprod = np.cumprod(alphas, axis=0)
|
147 |
+
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
|
148 |
+
|
149 |
+
timesteps, = betas.shape
|
150 |
+
self.num_timesteps = int(timesteps)
|
151 |
+
self.linear_start = linear_start
|
152 |
+
self.linear_end = linear_end
|
153 |
+
assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
|
154 |
+
|
155 |
+
to_torch = partial(torch.tensor, dtype=torch.float32)
|
156 |
+
|
157 |
+
self.register_buffer('betas', to_torch(betas))
|
158 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
159 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
|
160 |
+
|
161 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
162 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
|
163 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
|
164 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
|
165 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
|
166 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
|
167 |
+
|
168 |
+
# calculations for posterior q(x_{t-1} | x_t, x_0)
|
169 |
+
posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
|
170 |
+
1. - alphas_cumprod) + self.v_posterior * betas
|
171 |
+
# above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
|
172 |
+
self.register_buffer('posterior_variance', to_torch(posterior_variance))
|
173 |
+
# below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
|
174 |
+
self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
|
175 |
+
self.register_buffer('posterior_mean_coef1', to_torch(
|
176 |
+
betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
|
177 |
+
self.register_buffer('posterior_mean_coef2', to_torch(
|
178 |
+
(1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
|
179 |
+
|
180 |
+
if self.parameterization == "eps":
|
181 |
+
lvlb_weights = self.betas ** 2 / (
|
182 |
+
2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
|
183 |
+
elif self.parameterization == "x0":
|
184 |
+
lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
|
185 |
+
elif self.parameterization == "v":
|
186 |
+
lvlb_weights = torch.ones_like(self.betas ** 2 / (
|
187 |
+
2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)))
|
188 |
+
else:
|
189 |
+
raise NotImplementedError("mu not supported")
|
190 |
+
lvlb_weights[0] = lvlb_weights[1]
|
191 |
+
self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
|
192 |
+
assert not torch.isnan(self.lvlb_weights).all()
|
193 |
+
|
194 |
+
@contextmanager
|
195 |
+
def ema_scope(self, context=None):
|
196 |
+
if self.use_ema:
|
197 |
+
self.model_ema.store(self.model.parameters())
|
198 |
+
self.model_ema.copy_to(self.model)
|
199 |
+
if context is not None:
|
200 |
+
print(f"{context}: Switched to EMA weights")
|
201 |
+
try:
|
202 |
+
yield None
|
203 |
+
finally:
|
204 |
+
if self.use_ema:
|
205 |
+
self.model_ema.restore(self.model.parameters())
|
206 |
+
if context is not None:
|
207 |
+
print(f"{context}: Restored training weights")
|
208 |
+
|
209 |
+
@torch.no_grad()
|
210 |
+
def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
|
211 |
+
sd = torch.load(path, map_location="cpu")
|
212 |
+
if "state_dict" in list(sd.keys()):
|
213 |
+
sd = sd["state_dict"]
|
214 |
+
keys = list(sd.keys())
|
215 |
+
for k in keys:
|
216 |
+
for ik in ignore_keys:
|
217 |
+
if k.startswith(ik):
|
218 |
+
print("Deleting key {} from state_dict.".format(k))
|
219 |
+
del sd[k]
|
220 |
+
if self.make_it_fit:
|
221 |
+
n_params = len([name for name, _ in
|
222 |
+
itertools.chain(self.named_parameters(),
|
223 |
+
self.named_buffers())])
|
224 |
+
for name, param in tqdm(
|
225 |
+
itertools.chain(self.named_parameters(),
|
226 |
+
self.named_buffers()),
|
227 |
+
desc="Fitting old weights to new weights",
|
228 |
+
total=n_params
|
229 |
+
):
|
230 |
+
if not name in sd:
|
231 |
+
continue
|
232 |
+
old_shape = sd[name].shape
|
233 |
+
new_shape = param.shape
|
234 |
+
assert len(old_shape) == len(new_shape)
|
235 |
+
if len(new_shape) > 2:
|
236 |
+
# we only modify first two axes
|
237 |
+
assert new_shape[2:] == old_shape[2:]
|
238 |
+
# assumes first axis corresponds to output dim
|
239 |
+
if not new_shape == old_shape:
|
240 |
+
new_param = param.clone()
|
241 |
+
old_param = sd[name]
|
242 |
+
if len(new_shape) == 1:
|
243 |
+
for i in range(new_param.shape[0]):
|
244 |
+
new_param[i] = old_param[i % old_shape[0]]
|
245 |
+
elif len(new_shape) >= 2:
|
246 |
+
for i in range(new_param.shape[0]):
|
247 |
+
for j in range(new_param.shape[1]):
|
248 |
+
new_param[i, j] = old_param[i % old_shape[0], j % old_shape[1]]
|
249 |
+
|
250 |
+
n_used_old = torch.ones(old_shape[1])
|
251 |
+
for j in range(new_param.shape[1]):
|
252 |
+
n_used_old[j % old_shape[1]] += 1
|
253 |
+
n_used_new = torch.zeros(new_shape[1])
|
254 |
+
for j in range(new_param.shape[1]):
|
255 |
+
n_used_new[j] = n_used_old[j % old_shape[1]]
|
256 |
+
|
257 |
+
n_used_new = n_used_new[None, :]
|
258 |
+
while len(n_used_new.shape) < len(new_shape):
|
259 |
+
n_used_new = n_used_new.unsqueeze(-1)
|
260 |
+
new_param /= n_used_new
|
261 |
+
|
262 |
+
sd[name] = new_param
|
263 |
+
|
264 |
+
missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
|
265 |
+
sd, strict=False)
|
266 |
+
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
267 |
+
if len(missing) > 0:
|
268 |
+
print(f"Missing Keys:\n {missing}")
|
269 |
+
if len(unexpected) > 0:
|
270 |
+
print(f"\nUnexpected Keys:\n {unexpected}")
|
271 |
+
|
272 |
+
def q_mean_variance(self, x_start, t):
|
273 |
+
"""
|
274 |
+
Get the distribution q(x_t | x_0).
|
275 |
+
:param x_start: the [N x C x ...] tensor of noiseless inputs.
|
276 |
+
:param t: the number of diffusion steps (minus 1). Here, 0 means one step.
|
277 |
+
:return: A tuple (mean, variance, log_variance), all of x_start's shape.
|
278 |
+
"""
|
279 |
+
mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
|
280 |
+
variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
|
281 |
+
log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
|
282 |
+
return mean, variance, log_variance
|
283 |
+
|
284 |
+
def predict_start_from_noise(self, x_t, t, noise):
|
285 |
+
return (
|
286 |
+
extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
|
287 |
+
extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
|
288 |
+
)
|
289 |
+
|
290 |
+
def predict_start_from_z_and_v(self, x_t, t, v):
|
291 |
+
# self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
|
292 |
+
# self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
|
293 |
+
return (
|
294 |
+
extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t -
|
295 |
+
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v
|
296 |
+
)
|
297 |
+
|
298 |
+
def predict_eps_from_z_and_v(self, x_t, t, v):
|
299 |
+
return (
|
300 |
+
extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * v +
|
301 |
+
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * x_t
|
302 |
+
)
|
303 |
+
|
304 |
+
def q_posterior(self, x_start, x_t, t):
|
305 |
+
posterior_mean = (
|
306 |
+
extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
|
307 |
+
extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
|
308 |
+
)
|
309 |
+
posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
|
310 |
+
posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
|
311 |
+
return posterior_mean, posterior_variance, posterior_log_variance_clipped
|
312 |
+
|
313 |
+
def p_mean_variance(self, x, t, clip_denoised: bool):
|
314 |
+
model_out = self.model(x, t)
|
315 |
+
if self.parameterization == "eps":
|
316 |
+
x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
|
317 |
+
elif self.parameterization == "x0":
|
318 |
+
x_recon = model_out
|
319 |
+
if clip_denoised:
|
320 |
+
x_recon.clamp_(-1., 1.)
|
321 |
+
|
322 |
+
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
|
323 |
+
return model_mean, posterior_variance, posterior_log_variance
|
324 |
+
|
325 |
+
@torch.no_grad()
|
326 |
+
def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
|
327 |
+
b, *_, device = *x.shape, x.device
|
328 |
+
model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
|
329 |
+
noise = noise_like(x.shape, device, repeat_noise)
|
330 |
+
# no noise when t == 0
|
331 |
+
nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
|
332 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
|
333 |
+
|
334 |
+
@torch.no_grad()
|
335 |
+
def p_sample_loop(self, shape, return_intermediates=False):
|
336 |
+
device = self.betas.device
|
337 |
+
b = shape[0]
|
338 |
+
img = torch.randn(shape, device=device)
|
339 |
+
intermediates = [img]
|
340 |
+
for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
|
341 |
+
img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
|
342 |
+
clip_denoised=self.clip_denoised)
|
343 |
+
if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
|
344 |
+
intermediates.append(img)
|
345 |
+
if return_intermediates:
|
346 |
+
return img, intermediates
|
347 |
+
return img
|
348 |
+
|
349 |
+
@torch.no_grad()
|
350 |
+
def sample(self, batch_size=16, return_intermediates=False):
|
351 |
+
image_size = self.image_size
|
352 |
+
channels = self.channels
|
353 |
+
return self.p_sample_loop((batch_size, channels, image_size, image_size),
|
354 |
+
return_intermediates=return_intermediates)
|
355 |
+
|
356 |
+
def q_sample(self, x_start, t, noise=None):
|
357 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
358 |
+
return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
|
359 |
+
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
|
360 |
+
|
361 |
+
def get_v(self, x, noise, t):
|
362 |
+
return (
|
363 |
+
extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise -
|
364 |
+
extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x
|
365 |
+
)
|
366 |
+
|
367 |
+
def get_loss(self, pred, target, mean=True):
|
368 |
+
if self.loss_type == 'l1':
|
369 |
+
loss = (target - pred).abs()
|
370 |
+
if mean:
|
371 |
+
loss = loss.mean()
|
372 |
+
elif self.loss_type == 'l2':
|
373 |
+
if mean:
|
374 |
+
loss = torch.nn.functional.mse_loss(target, pred)
|
375 |
+
else:
|
376 |
+
loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
|
377 |
+
else:
|
378 |
+
raise NotImplementedError("unknown loss type '{loss_type}'")
|
379 |
+
|
380 |
+
return loss
|
381 |
+
|
382 |
+
def p_losses(self, x_start, t, noise=None):
|
383 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
384 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
385 |
+
model_out = self.model(x_noisy, t)
|
386 |
+
|
387 |
+
loss_dict = {}
|
388 |
+
if self.parameterization == "eps":
|
389 |
+
target = noise
|
390 |
+
elif self.parameterization == "x0":
|
391 |
+
target = x_start
|
392 |
+
elif self.parameterization == "v":
|
393 |
+
target = self.get_v(x_start, noise, t)
|
394 |
+
else:
|
395 |
+
raise NotImplementedError(f"Parameterization {self.parameterization} not yet supported")
|
396 |
+
|
397 |
+
loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
|
398 |
+
|
399 |
+
log_prefix = 'train' if self.training else 'val'
|
400 |
+
|
401 |
+
loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
|
402 |
+
loss_simple = loss.mean() * self.l_simple_weight
|
403 |
+
|
404 |
+
loss_vlb = (self.lvlb_weights[t] * loss).mean()
|
405 |
+
loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
|
406 |
+
|
407 |
+
loss = loss_simple + self.original_elbo_weight * loss_vlb
|
408 |
+
|
409 |
+
loss_dict.update({f'{log_prefix}/loss': loss})
|
410 |
+
|
411 |
+
return loss, loss_dict
|
412 |
+
|
413 |
+
def forward(self, x, *args, **kwargs):
|
414 |
+
# b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
|
415 |
+
# assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
|
416 |
+
t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
|
417 |
+
return self.p_losses(x, t, *args, **kwargs)
|
418 |
+
|
419 |
+
def get_input(self, batch, k):
|
420 |
+
x = batch[k]
|
421 |
+
if len(x.shape) == 3:
|
422 |
+
x = x[..., None]
|
423 |
+
x = rearrange(x, 'b h w c -> b c h w')
|
424 |
+
x = x.to(memory_format=torch.contiguous_format).float()
|
425 |
+
return x
|
426 |
+
|
427 |
+
def shared_step(self, batch):
|
428 |
+
x = self.get_input(batch, self.first_stage_key)
|
429 |
+
loss, loss_dict = self(x)
|
430 |
+
return loss, loss_dict
|
431 |
+
|
432 |
+
def training_step(self, batch, batch_idx):
|
433 |
+
for k in self.ucg_training:
|
434 |
+
p = self.ucg_training[k]["p"]
|
435 |
+
val = self.ucg_training[k]["val"]
|
436 |
+
if val is None:
|
437 |
+
val = ""
|
438 |
+
for i in range(len(batch[k])):
|
439 |
+
if self.ucg_prng.choice(2, p=[1 - p, p]):
|
440 |
+
batch[k][i] = val
|
441 |
+
|
442 |
+
loss, loss_dict = self.shared_step(batch)
|
443 |
+
|
444 |
+
self.log_dict(loss_dict, prog_bar=True,
|
445 |
+
logger=True, on_step=True, on_epoch=True)
|
446 |
+
|
447 |
+
self.log("global_step", self.global_step,
|
448 |
+
prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
449 |
+
|
450 |
+
if self.use_scheduler:
|
451 |
+
lr = self.optimizers().param_groups[0]['lr']
|
452 |
+
self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
453 |
+
|
454 |
+
return loss
|
455 |
+
|
456 |
+
@torch.no_grad()
|
457 |
+
def validation_step(self, batch, batch_idx):
|
458 |
+
_, loss_dict_no_ema = self.shared_step(batch)
|
459 |
+
with self.ema_scope():
|
460 |
+
_, loss_dict_ema = self.shared_step(batch)
|
461 |
+
loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
|
462 |
+
self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
|
463 |
+
self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
|
464 |
+
|
465 |
+
def on_train_batch_end(self, *args, **kwargs):
|
466 |
+
if self.use_ema:
|
467 |
+
self.model_ema(self.model)
|
468 |
+
|
469 |
+
def _get_rows_from_list(self, samples):
|
470 |
+
n_imgs_per_row = len(samples)
|
471 |
+
denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
|
472 |
+
denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
|
473 |
+
denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
|
474 |
+
return denoise_grid
|
475 |
+
|
476 |
+
@torch.no_grad()
|
477 |
+
def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
|
478 |
+
log = dict()
|
479 |
+
x = self.get_input(batch, self.first_stage_key)
|
480 |
+
N = min(x.shape[0], N)
|
481 |
+
n_row = min(x.shape[0], n_row)
|
482 |
+
x = x.to(self.device)[:N]
|
483 |
+
log["inputs"] = x
|
484 |
+
|
485 |
+
# get diffusion row
|
486 |
+
diffusion_row = list()
|
487 |
+
x_start = x[:n_row]
|
488 |
+
|
489 |
+
for t in range(self.num_timesteps):
|
490 |
+
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
491 |
+
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
492 |
+
t = t.to(self.device).long()
|
493 |
+
noise = torch.randn_like(x_start)
|
494 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
495 |
+
diffusion_row.append(x_noisy)
|
496 |
+
|
497 |
+
log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
|
498 |
+
|
499 |
+
if sample:
|
500 |
+
# get denoise row
|
501 |
+
with self.ema_scope("Plotting"):
|
502 |
+
samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
|
503 |
+
|
504 |
+
log["samples"] = samples
|
505 |
+
log["denoise_row"] = self._get_rows_from_list(denoise_row)
|
506 |
+
|
507 |
+
if return_keys:
|
508 |
+
if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
|
509 |
+
return log
|
510 |
+
else:
|
511 |
+
return {key: log[key] for key in return_keys}
|
512 |
+
return log
|
513 |
+
|
514 |
+
def configure_optimizers(self):
|
515 |
+
lr = self.learning_rate
|
516 |
+
params = list(self.model.parameters())
|
517 |
+
if self.learn_logvar:
|
518 |
+
params = params + [self.logvar]
|
519 |
+
opt = torch.optim.AdamW(params, lr=lr)
|
520 |
+
return opt
|
521 |
+
|
522 |
+
|
523 |
+
class LatentDiffusion(DDPM):
|
524 |
+
"""main class"""
|
525 |
+
|
526 |
+
def __init__(self,
|
527 |
+
first_stage_config,
|
528 |
+
cond_stage_config,
|
529 |
+
num_timesteps_cond=None,
|
530 |
+
cond_stage_key="image",
|
531 |
+
cond_stage_trainable=False,
|
532 |
+
concat_mode=True,
|
533 |
+
cond_stage_forward=None,
|
534 |
+
conditioning_key=None,
|
535 |
+
scale_factor=1.0,
|
536 |
+
scale_by_std=False,
|
537 |
+
force_null_conditioning=False,
|
538 |
+
*args, **kwargs):
|
539 |
+
self.force_null_conditioning = force_null_conditioning
|
540 |
+
self.num_timesteps_cond = default(num_timesteps_cond, 1)
|
541 |
+
self.scale_by_std = scale_by_std
|
542 |
+
assert self.num_timesteps_cond <= kwargs['timesteps']
|
543 |
+
# for backwards compatibility after implementation of DiffusionWrapper
|
544 |
+
if conditioning_key is None:
|
545 |
+
conditioning_key = 'concat' if concat_mode else 'crossattn'
|
546 |
+
if cond_stage_config == '__is_unconditional__' and not self.force_null_conditioning:
|
547 |
+
conditioning_key = None
|
548 |
+
ckpt_path = kwargs.pop("ckpt_path", None)
|
549 |
+
reset_ema = kwargs.pop("reset_ema", False)
|
550 |
+
reset_num_ema_updates = kwargs.pop("reset_num_ema_updates", False)
|
551 |
+
ignore_keys = kwargs.pop("ignore_keys", [])
|
552 |
+
super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
|
553 |
+
self.concat_mode = concat_mode
|
554 |
+
self.cond_stage_trainable = cond_stage_trainable
|
555 |
+
self.cond_stage_key = cond_stage_key
|
556 |
+
try:
|
557 |
+
self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
|
558 |
+
except:
|
559 |
+
self.num_downs = 0
|
560 |
+
if not scale_by_std:
|
561 |
+
self.scale_factor = scale_factor
|
562 |
+
else:
|
563 |
+
self.register_buffer('scale_factor', torch.tensor(scale_factor))
|
564 |
+
self.instantiate_first_stage(first_stage_config)
|
565 |
+
self.instantiate_cond_stage(cond_stage_config)
|
566 |
+
self.cond_stage_forward = cond_stage_forward
|
567 |
+
self.clip_denoised = False
|
568 |
+
self.bbox_tokenizer = None
|
569 |
+
|
570 |
+
self.restarted_from_ckpt = False
|
571 |
+
if ckpt_path is not None:
|
572 |
+
self.init_from_ckpt(ckpt_path, ignore_keys)
|
573 |
+
self.restarted_from_ckpt = True
|
574 |
+
if reset_ema:
|
575 |
+
assert self.use_ema
|
576 |
+
print(
|
577 |
+
f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.")
|
578 |
+
self.model_ema = LitEma(self.model)
|
579 |
+
if reset_num_ema_updates:
|
580 |
+
print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ")
|
581 |
+
assert self.use_ema
|
582 |
+
self.model_ema.reset_num_updates()
|
583 |
+
|
584 |
+
def make_cond_schedule(self, ):
|
585 |
+
self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
|
586 |
+
ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
|
587 |
+
self.cond_ids[:self.num_timesteps_cond] = ids
|
588 |
+
|
589 |
+
@rank_zero_only
|
590 |
+
@torch.no_grad()
|
591 |
+
def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
|
592 |
+
# only for very first batch
|
593 |
+
if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
|
594 |
+
assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
|
595 |
+
# set rescale weight to 1./std of encodings
|
596 |
+
print("### USING STD-RESCALING ###")
|
597 |
+
x = super().get_input(batch, self.first_stage_key)
|
598 |
+
x = x.to(self.device)
|
599 |
+
encoder_posterior = self.encode_first_stage(x)
|
600 |
+
z = self.get_first_stage_encoding(encoder_posterior).detach()
|
601 |
+
del self.scale_factor
|
602 |
+
self.register_buffer('scale_factor', 1. / z.flatten().std())
|
603 |
+
print(f"setting self.scale_factor to {self.scale_factor}")
|
604 |
+
print("### USING STD-RESCALING ###")
|
605 |
+
|
606 |
+
def register_schedule(self,
|
607 |
+
given_betas=None, beta_schedule="linear", timesteps=1000,
|
608 |
+
linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
609 |
+
super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
|
610 |
+
|
611 |
+
self.shorten_cond_schedule = self.num_timesteps_cond > 1
|
612 |
+
if self.shorten_cond_schedule:
|
613 |
+
self.make_cond_schedule()
|
614 |
+
|
615 |
+
def instantiate_first_stage(self, config):
|
616 |
+
model = instantiate_from_config(config)
|
617 |
+
self.first_stage_model = model.eval()
|
618 |
+
self.first_stage_model.train = disabled_train
|
619 |
+
for param in self.first_stage_model.parameters():
|
620 |
+
param.requires_grad = False
|
621 |
+
|
622 |
+
def instantiate_cond_stage(self, config):
|
623 |
+
if not self.cond_stage_trainable:
|
624 |
+
if config == "__is_first_stage__":
|
625 |
+
print("Using first stage also as cond stage.")
|
626 |
+
self.cond_stage_model = self.first_stage_model
|
627 |
+
elif config == "__is_unconditional__":
|
628 |
+
print(f"Training {self.__class__.__name__} as an unconditional model.")
|
629 |
+
self.cond_stage_model = None
|
630 |
+
# self.be_unconditional = True
|
631 |
+
else:
|
632 |
+
model = instantiate_from_config(config)
|
633 |
+
self.cond_stage_model = model.eval()
|
634 |
+
self.cond_stage_model.train = disabled_train
|
635 |
+
for param in self.cond_stage_model.parameters():
|
636 |
+
param.requires_grad = False
|
637 |
+
else:
|
638 |
+
assert config != '__is_first_stage__'
|
639 |
+
assert config != '__is_unconditional__'
|
640 |
+
model = instantiate_from_config(config)
|
641 |
+
self.cond_stage_model = model
|
642 |
+
|
643 |
+
def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
|
644 |
+
denoise_row = []
|
645 |
+
for zd in tqdm(samples, desc=desc):
|
646 |
+
denoise_row.append(self.decode_first_stage(zd.to(self.device),
|
647 |
+
force_not_quantize=force_no_decoder_quantization))
|
648 |
+
n_imgs_per_row = len(denoise_row)
|
649 |
+
denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W
|
650 |
+
denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
|
651 |
+
denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
|
652 |
+
denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
|
653 |
+
return denoise_grid
|
654 |
+
|
655 |
+
def get_first_stage_encoding(self, encoder_posterior):
|
656 |
+
if isinstance(encoder_posterior, DiagonalGaussianDistribution):
|
657 |
+
z = encoder_posterior.sample()
|
658 |
+
elif isinstance(encoder_posterior, torch.Tensor):
|
659 |
+
z = encoder_posterior
|
660 |
+
else:
|
661 |
+
raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
|
662 |
+
return self.scale_factor * z
|
663 |
+
|
664 |
+
def get_learned_conditioning(self, c):
|
665 |
+
if self.cond_stage_forward is None:
|
666 |
+
if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
|
667 |
+
c = self.cond_stage_model.encode(c)
|
668 |
+
if isinstance(c, DiagonalGaussianDistribution):
|
669 |
+
c = c.mode()
|
670 |
+
else:
|
671 |
+
c = self.cond_stage_model(c)
|
672 |
+
else:
|
673 |
+
assert hasattr(self.cond_stage_model, self.cond_stage_forward)
|
674 |
+
c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
|
675 |
+
return c
|
676 |
+
|
677 |
+
def meshgrid(self, h, w):
|
678 |
+
y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
|
679 |
+
x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
|
680 |
+
|
681 |
+
arr = torch.cat([y, x], dim=-1)
|
682 |
+
return arr
|
683 |
+
|
684 |
+
def delta_border(self, h, w):
|
685 |
+
"""
|
686 |
+
:param h: height
|
687 |
+
:param w: width
|
688 |
+
:return: normalized distance to image border,
|
689 |
+
wtith min distance = 0 at border and max dist = 0.5 at image center
|
690 |
+
"""
|
691 |
+
lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
|
692 |
+
arr = self.meshgrid(h, w) / lower_right_corner
|
693 |
+
dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
|
694 |
+
dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
|
695 |
+
edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
|
696 |
+
return edge_dist
|
697 |
+
|
698 |
+
def get_weighting(self, h, w, Ly, Lx, device):
|
699 |
+
weighting = self.delta_border(h, w)
|
700 |
+
weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
|
701 |
+
self.split_input_params["clip_max_weight"], )
|
702 |
+
weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
|
703 |
+
|
704 |
+
if self.split_input_params["tie_braker"]:
|
705 |
+
L_weighting = self.delta_border(Ly, Lx)
|
706 |
+
L_weighting = torch.clip(L_weighting,
|
707 |
+
self.split_input_params["clip_min_tie_weight"],
|
708 |
+
self.split_input_params["clip_max_tie_weight"])
|
709 |
+
|
710 |
+
L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
|
711 |
+
weighting = weighting * L_weighting
|
712 |
+
return weighting
|
713 |
+
|
714 |
+
def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code
|
715 |
+
"""
|
716 |
+
:param x: img of size (bs, c, h, w)
|
717 |
+
:return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
|
718 |
+
"""
|
719 |
+
bs, nc, h, w = x.shape
|
720 |
+
|
721 |
+
# number of crops in image
|
722 |
+
Ly = (h - kernel_size[0]) // stride[0] + 1
|
723 |
+
Lx = (w - kernel_size[1]) // stride[1] + 1
|
724 |
+
|
725 |
+
if uf == 1 and df == 1:
|
726 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
727 |
+
unfold = torch.nn.Unfold(**fold_params)
|
728 |
+
|
729 |
+
fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
|
730 |
+
|
731 |
+
weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
|
732 |
+
normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap
|
733 |
+
weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
|
734 |
+
|
735 |
+
elif uf > 1 and df == 1:
|
736 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
737 |
+
unfold = torch.nn.Unfold(**fold_params)
|
738 |
+
|
739 |
+
fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
|
740 |
+
dilation=1, padding=0,
|
741 |
+
stride=(stride[0] * uf, stride[1] * uf))
|
742 |
+
fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
|
743 |
+
|
744 |
+
weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
|
745 |
+
normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap
|
746 |
+
weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
|
747 |
+
|
748 |
+
elif df > 1 and uf == 1:
|
749 |
+
fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
|
750 |
+
unfold = torch.nn.Unfold(**fold_params)
|
751 |
+
|
752 |
+
fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
|
753 |
+
dilation=1, padding=0,
|
754 |
+
stride=(stride[0] // df, stride[1] // df))
|
755 |
+
fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
|
756 |
+
|
757 |
+
weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
|
758 |
+
normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap
|
759 |
+
weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
|
760 |
+
|
761 |
+
else:
|
762 |
+
raise NotImplementedError
|
763 |
+
|
764 |
+
return fold, unfold, normalization, weighting
|
765 |
+
|
766 |
+
@torch.no_grad()
|
767 |
+
def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
|
768 |
+
cond_key=None, return_original_cond=False, bs=None, return_x=False):
|
769 |
+
x = super().get_input(batch, k)
|
770 |
+
if bs is not None:
|
771 |
+
x = x[:bs]
|
772 |
+
x = x.to(self.device)
|
773 |
+
encoder_posterior = self.encode_first_stage(x)
|
774 |
+
z = self.get_first_stage_encoding(encoder_posterior).detach()
|
775 |
+
|
776 |
+
if self.model.conditioning_key is not None and not self.force_null_conditioning:
|
777 |
+
if cond_key is None:
|
778 |
+
cond_key = self.cond_stage_key
|
779 |
+
if cond_key != self.first_stage_key:
|
780 |
+
if cond_key in ['caption', 'coordinates_bbox', "txt"]:
|
781 |
+
xc = batch[cond_key]
|
782 |
+
elif cond_key in ['class_label', 'cls']:
|
783 |
+
xc = batch
|
784 |
+
else:
|
785 |
+
xc = super().get_input(batch, cond_key).to(self.device)
|
786 |
+
else:
|
787 |
+
xc = x
|
788 |
+
if not self.cond_stage_trainable or force_c_encode:
|
789 |
+
if isinstance(xc, dict) or isinstance(xc, list):
|
790 |
+
c = self.get_learned_conditioning(xc)
|
791 |
+
else:
|
792 |
+
c = self.get_learned_conditioning(xc.to(self.device))
|
793 |
+
else:
|
794 |
+
c = xc
|
795 |
+
if bs is not None:
|
796 |
+
c = c[:bs]
|
797 |
+
|
798 |
+
if self.use_positional_encodings:
|
799 |
+
pos_x, pos_y = self.compute_latent_shifts(batch)
|
800 |
+
ckey = __conditioning_keys__[self.model.conditioning_key]
|
801 |
+
c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
|
802 |
+
|
803 |
+
else:
|
804 |
+
c = None
|
805 |
+
xc = None
|
806 |
+
if self.use_positional_encodings:
|
807 |
+
pos_x, pos_y = self.compute_latent_shifts(batch)
|
808 |
+
c = {'pos_x': pos_x, 'pos_y': pos_y}
|
809 |
+
out = [z, c]
|
810 |
+
if return_first_stage_outputs:
|
811 |
+
xrec = self.decode_first_stage(z)
|
812 |
+
out.extend([x, xrec])
|
813 |
+
if return_x:
|
814 |
+
out.extend([x])
|
815 |
+
if return_original_cond:
|
816 |
+
out.append(xc)
|
817 |
+
return out
|
818 |
+
|
819 |
+
@torch.no_grad()
|
820 |
+
def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
|
821 |
+
if predict_cids:
|
822 |
+
if z.dim() == 4:
|
823 |
+
z = torch.argmax(z.exp(), dim=1).long()
|
824 |
+
z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
|
825 |
+
z = rearrange(z, 'b h w c -> b c h w').contiguous()
|
826 |
+
|
827 |
+
z = 1. / self.scale_factor * z
|
828 |
+
return self.first_stage_model.decode(z)
|
829 |
+
|
830 |
+
@torch.no_grad()
|
831 |
+
def encode_first_stage(self, x):
|
832 |
+
return self.first_stage_model.encode(x)
|
833 |
+
|
834 |
+
def shared_step(self, batch, **kwargs):
|
835 |
+
x, c = self.get_input(batch, self.first_stage_key)
|
836 |
+
loss = self(x, c)
|
837 |
+
return loss
|
838 |
+
|
839 |
+
def forward(self, x, c, *args, **kwargs):
|
840 |
+
t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
|
841 |
+
if self.model.conditioning_key is not None:
|
842 |
+
assert c is not None
|
843 |
+
if self.cond_stage_trainable:
|
844 |
+
c = self.get_learned_conditioning(c)
|
845 |
+
if self.shorten_cond_schedule: # TODO: drop this option
|
846 |
+
tc = self.cond_ids[t].to(self.device)
|
847 |
+
c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
|
848 |
+
return self.p_losses(x, c, t, *args, **kwargs)
|
849 |
+
|
850 |
+
def apply_model(self, x_noisy, t, cond, return_ids=False):
|
851 |
+
if isinstance(cond, dict):
|
852 |
+
# hybrid case, cond is expected to be a dict
|
853 |
+
pass
|
854 |
+
else:
|
855 |
+
if not isinstance(cond, list):
|
856 |
+
cond = [cond]
|
857 |
+
key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
|
858 |
+
cond = {key: cond}
|
859 |
+
|
860 |
+
x_recon = self.model(x_noisy, t, **cond)
|
861 |
+
|
862 |
+
if isinstance(x_recon, tuple) and not return_ids:
|
863 |
+
return x_recon[0]
|
864 |
+
else:
|
865 |
+
return x_recon
|
866 |
+
|
867 |
+
def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
|
868 |
+
return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
|
869 |
+
extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
|
870 |
+
|
871 |
+
def _prior_bpd(self, x_start):
|
872 |
+
"""
|
873 |
+
Get the prior KL term for the variational lower-bound, measured in
|
874 |
+
bits-per-dim.
|
875 |
+
This term can't be optimized, as it only depends on the encoder.
|
876 |
+
:param x_start: the [N x C x ...] tensor of inputs.
|
877 |
+
:return: a batch of [N] KL values (in bits), one per batch element.
|
878 |
+
"""
|
879 |
+
batch_size = x_start.shape[0]
|
880 |
+
t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
|
881 |
+
qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
|
882 |
+
kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
|
883 |
+
return mean_flat(kl_prior) / np.log(2.0)
|
884 |
+
|
885 |
+
def p_losses(self, x_start, cond, t, noise=None):
|
886 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
887 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
888 |
+
model_output = self.apply_model(x_noisy, t, cond)
|
889 |
+
|
890 |
+
loss_dict = {}
|
891 |
+
prefix = 'train' if self.training else 'val'
|
892 |
+
|
893 |
+
if self.parameterization == "x0":
|
894 |
+
target = x_start
|
895 |
+
elif self.parameterization == "eps":
|
896 |
+
target = noise
|
897 |
+
elif self.parameterization == "v":
|
898 |
+
target = self.get_v(x_start, noise, t)
|
899 |
+
else:
|
900 |
+
raise NotImplementedError()
|
901 |
+
|
902 |
+
loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
|
903 |
+
loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
|
904 |
+
|
905 |
+
logvar_t = self.logvar[t].to(self.device)
|
906 |
+
loss = loss_simple / torch.exp(logvar_t) + logvar_t
|
907 |
+
# loss = loss_simple / torch.exp(self.logvar) + self.logvar
|
908 |
+
if self.learn_logvar:
|
909 |
+
loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
|
910 |
+
loss_dict.update({'logvar': self.logvar.data.mean()})
|
911 |
+
|
912 |
+
loss = self.l_simple_weight * loss.mean()
|
913 |
+
|
914 |
+
loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
|
915 |
+
loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
|
916 |
+
loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
|
917 |
+
loss += (self.original_elbo_weight * loss_vlb)
|
918 |
+
loss_dict.update({f'{prefix}/loss': loss})
|
919 |
+
|
920 |
+
return loss, loss_dict
|
921 |
+
|
922 |
+
def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
|
923 |
+
return_x0=False, score_corrector=None, corrector_kwargs=None):
|
924 |
+
t_in = t
|
925 |
+
model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
|
926 |
+
|
927 |
+
if score_corrector is not None:
|
928 |
+
assert self.parameterization == "eps"
|
929 |
+
model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
|
930 |
+
|
931 |
+
if return_codebook_ids:
|
932 |
+
model_out, logits = model_out
|
933 |
+
|
934 |
+
if self.parameterization == "eps":
|
935 |
+
x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
|
936 |
+
elif self.parameterization == "x0":
|
937 |
+
x_recon = model_out
|
938 |
+
else:
|
939 |
+
raise NotImplementedError()
|
940 |
+
|
941 |
+
if clip_denoised:
|
942 |
+
x_recon.clamp_(-1., 1.)
|
943 |
+
if quantize_denoised:
|
944 |
+
x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
|
945 |
+
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
|
946 |
+
if return_codebook_ids:
|
947 |
+
return model_mean, posterior_variance, posterior_log_variance, logits
|
948 |
+
elif return_x0:
|
949 |
+
return model_mean, posterior_variance, posterior_log_variance, x_recon
|
950 |
+
else:
|
951 |
+
return model_mean, posterior_variance, posterior_log_variance
|
952 |
+
|
953 |
+
@torch.no_grad()
|
954 |
+
def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
|
955 |
+
return_codebook_ids=False, quantize_denoised=False, return_x0=False,
|
956 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
|
957 |
+
b, *_, device = *x.shape, x.device
|
958 |
+
outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
|
959 |
+
return_codebook_ids=return_codebook_ids,
|
960 |
+
quantize_denoised=quantize_denoised,
|
961 |
+
return_x0=return_x0,
|
962 |
+
score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
|
963 |
+
if return_codebook_ids:
|
964 |
+
raise DeprecationWarning("Support dropped.")
|
965 |
+
model_mean, _, model_log_variance, logits = outputs
|
966 |
+
elif return_x0:
|
967 |
+
model_mean, _, model_log_variance, x0 = outputs
|
968 |
+
else:
|
969 |
+
model_mean, _, model_log_variance = outputs
|
970 |
+
|
971 |
+
noise = noise_like(x.shape, device, repeat_noise) * temperature
|
972 |
+
if noise_dropout > 0.:
|
973 |
+
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
974 |
+
# no noise when t == 0
|
975 |
+
nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
|
976 |
+
|
977 |
+
if return_codebook_ids:
|
978 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
|
979 |
+
if return_x0:
|
980 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
|
981 |
+
else:
|
982 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
|
983 |
+
|
984 |
+
@torch.no_grad()
|
985 |
+
def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
|
986 |
+
img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
|
987 |
+
score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
|
988 |
+
log_every_t=None):
|
989 |
+
if not log_every_t:
|
990 |
+
log_every_t = self.log_every_t
|
991 |
+
timesteps = self.num_timesteps
|
992 |
+
if batch_size is not None:
|
993 |
+
b = batch_size if batch_size is not None else shape[0]
|
994 |
+
shape = [batch_size] + list(shape)
|
995 |
+
else:
|
996 |
+
b = batch_size = shape[0]
|
997 |
+
if x_T is None:
|
998 |
+
img = torch.randn(shape, device=self.device)
|
999 |
+
else:
|
1000 |
+
img = x_T
|
1001 |
+
intermediates = []
|
1002 |
+
if cond is not None:
|
1003 |
+
if isinstance(cond, dict):
|
1004 |
+
cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
|
1005 |
+
list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
|
1006 |
+
else:
|
1007 |
+
cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
|
1008 |
+
|
1009 |
+
if start_T is not None:
|
1010 |
+
timesteps = min(timesteps, start_T)
|
1011 |
+
iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
|
1012 |
+
total=timesteps) if verbose else reversed(
|
1013 |
+
range(0, timesteps))
|
1014 |
+
if type(temperature) == float:
|
1015 |
+
temperature = [temperature] * timesteps
|
1016 |
+
|
1017 |
+
for i in iterator:
|
1018 |
+
ts = torch.full((b,), i, device=self.device, dtype=torch.long)
|
1019 |
+
if self.shorten_cond_schedule:
|
1020 |
+
assert self.model.conditioning_key != 'hybrid'
|
1021 |
+
tc = self.cond_ids[ts].to(cond.device)
|
1022 |
+
cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
|
1023 |
+
|
1024 |
+
img, x0_partial = self.p_sample(img, cond, ts,
|
1025 |
+
clip_denoised=self.clip_denoised,
|
1026 |
+
quantize_denoised=quantize_denoised, return_x0=True,
|
1027 |
+
temperature=temperature[i], noise_dropout=noise_dropout,
|
1028 |
+
score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
|
1029 |
+
if mask is not None:
|
1030 |
+
assert x0 is not None
|
1031 |
+
img_orig = self.q_sample(x0, ts)
|
1032 |
+
img = img_orig * mask + (1. - mask) * img
|
1033 |
+
|
1034 |
+
if i % log_every_t == 0 or i == timesteps - 1:
|
1035 |
+
intermediates.append(x0_partial)
|
1036 |
+
if callback: callback(i)
|
1037 |
+
if img_callback: img_callback(img, i)
|
1038 |
+
return img, intermediates
|
1039 |
+
|
1040 |
+
@torch.no_grad()
|
1041 |
+
def p_sample_loop(self, cond, shape, return_intermediates=False,
|
1042 |
+
x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
|
1043 |
+
mask=None, x0=None, img_callback=None, start_T=None,
|
1044 |
+
log_every_t=None):
|
1045 |
+
|
1046 |
+
if not log_every_t:
|
1047 |
+
log_every_t = self.log_every_t
|
1048 |
+
device = self.betas.device
|
1049 |
+
b = shape[0]
|
1050 |
+
if x_T is None:
|
1051 |
+
img = torch.randn(shape, device=device)
|
1052 |
+
else:
|
1053 |
+
img = x_T
|
1054 |
+
|
1055 |
+
intermediates = [img]
|
1056 |
+
if timesteps is None:
|
1057 |
+
timesteps = self.num_timesteps
|
1058 |
+
|
1059 |
+
if start_T is not None:
|
1060 |
+
timesteps = min(timesteps, start_T)
|
1061 |
+
iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
|
1062 |
+
range(0, timesteps))
|
1063 |
+
|
1064 |
+
if mask is not None:
|
1065 |
+
assert x0 is not None
|
1066 |
+
assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match
|
1067 |
+
|
1068 |
+
for i in iterator:
|
1069 |
+
ts = torch.full((b,), i, device=device, dtype=torch.long)
|
1070 |
+
if self.shorten_cond_schedule:
|
1071 |
+
assert self.model.conditioning_key != 'hybrid'
|
1072 |
+
tc = self.cond_ids[ts].to(cond.device)
|
1073 |
+
cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
|
1074 |
+
|
1075 |
+
img = self.p_sample(img, cond, ts,
|
1076 |
+
clip_denoised=self.clip_denoised,
|
1077 |
+
quantize_denoised=quantize_denoised)
|
1078 |
+
if mask is not None:
|
1079 |
+
img_orig = self.q_sample(x0, ts)
|
1080 |
+
img = img_orig * mask + (1. - mask) * img
|
1081 |
+
|
1082 |
+
if i % log_every_t == 0 or i == timesteps - 1:
|
1083 |
+
intermediates.append(img)
|
1084 |
+
if callback: callback(i)
|
1085 |
+
if img_callback: img_callback(img, i)
|
1086 |
+
|
1087 |
+
if return_intermediates:
|
1088 |
+
return img, intermediates
|
1089 |
+
return img
|
1090 |
+
|
1091 |
+
@torch.no_grad()
|
1092 |
+
def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
|
1093 |
+
verbose=True, timesteps=None, quantize_denoised=False,
|
1094 |
+
mask=None, x0=None, shape=None, **kwargs):
|
1095 |
+
if shape is None:
|
1096 |
+
shape = (batch_size, self.channels, self.image_size, self.image_size)
|
1097 |
+
if cond is not None:
|
1098 |
+
if isinstance(cond, dict):
|
1099 |
+
cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
|
1100 |
+
list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
|
1101 |
+
else:
|
1102 |
+
cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
|
1103 |
+
return self.p_sample_loop(cond,
|
1104 |
+
shape,
|
1105 |
+
return_intermediates=return_intermediates, x_T=x_T,
|
1106 |
+
verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
|
1107 |
+
mask=mask, x0=x0)
|
1108 |
+
|
1109 |
+
@torch.no_grad()
|
1110 |
+
def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs):
|
1111 |
+
if ddim:
|
1112 |
+
ddim_sampler = DDIMSampler(self)
|
1113 |
+
shape = (self.channels, self.image_size, self.image_size)
|
1114 |
+
samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size,
|
1115 |
+
shape, cond, verbose=False, **kwargs)
|
1116 |
+
|
1117 |
+
else:
|
1118 |
+
samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
|
1119 |
+
return_intermediates=True, **kwargs)
|
1120 |
+
|
1121 |
+
return samples, intermediates
|
1122 |
+
|
1123 |
+
@torch.no_grad()
|
1124 |
+
def get_unconditional_conditioning(self, batch_size, null_label=None):
|
1125 |
+
if null_label is not None:
|
1126 |
+
xc = null_label
|
1127 |
+
if isinstance(xc, ListConfig):
|
1128 |
+
xc = list(xc)
|
1129 |
+
if isinstance(xc, dict) or isinstance(xc, list):
|
1130 |
+
c = self.get_learned_conditioning(xc)
|
1131 |
+
else:
|
1132 |
+
if hasattr(xc, "to"):
|
1133 |
+
xc = xc.to(self.device)
|
1134 |
+
c = self.get_learned_conditioning(xc)
|
1135 |
+
else:
|
1136 |
+
if self.cond_stage_key in ["class_label", "cls"]:
|
1137 |
+
xc = self.cond_stage_model.get_unconditional_conditioning(batch_size, device=self.device)
|
1138 |
+
return self.get_learned_conditioning(xc)
|
1139 |
+
else:
|
1140 |
+
raise NotImplementedError("todo")
|
1141 |
+
if isinstance(c, list): # in case the encoder gives us a list
|
1142 |
+
for i in range(len(c)):
|
1143 |
+
c[i] = repeat(c[i], '1 ... -> b ...', b=batch_size).to(self.device)
|
1144 |
+
else:
|
1145 |
+
c = repeat(c, '1 ... -> b ...', b=batch_size).to(self.device)
|
1146 |
+
return c
|
1147 |
+
|
1148 |
+
@torch.no_grad()
|
1149 |
+
def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=50, ddim_eta=0., return_keys=None,
|
1150 |
+
quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
|
1151 |
+
plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None,
|
1152 |
+
use_ema_scope=True,
|
1153 |
+
**kwargs):
|
1154 |
+
ema_scope = self.ema_scope if use_ema_scope else nullcontext
|
1155 |
+
use_ddim = ddim_steps is not None
|
1156 |
+
|
1157 |
+
log = dict()
|
1158 |
+
z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
|
1159 |
+
return_first_stage_outputs=True,
|
1160 |
+
force_c_encode=True,
|
1161 |
+
return_original_cond=True,
|
1162 |
+
bs=N)
|
1163 |
+
N = min(x.shape[0], N)
|
1164 |
+
n_row = min(x.shape[0], n_row)
|
1165 |
+
log["inputs"] = x
|
1166 |
+
log["reconstruction"] = xrec
|
1167 |
+
if self.model.conditioning_key is not None:
|
1168 |
+
if hasattr(self.cond_stage_model, "decode"):
|
1169 |
+
xc = self.cond_stage_model.decode(c)
|
1170 |
+
log["conditioning"] = xc
|
1171 |
+
elif self.cond_stage_key in ["caption", "txt"]:
|
1172 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25)
|
1173 |
+
log["conditioning"] = xc
|
1174 |
+
elif self.cond_stage_key in ['class_label', "cls"]:
|
1175 |
+
try:
|
1176 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25)
|
1177 |
+
log['conditioning'] = xc
|
1178 |
+
except KeyError:
|
1179 |
+
# probably no "human_label" in batch
|
1180 |
+
pass
|
1181 |
+
elif isimage(xc):
|
1182 |
+
log["conditioning"] = xc
|
1183 |
+
if ismap(xc):
|
1184 |
+
log["original_conditioning"] = self.to_rgb(xc)
|
1185 |
+
|
1186 |
+
if plot_diffusion_rows:
|
1187 |
+
# get diffusion row
|
1188 |
+
diffusion_row = list()
|
1189 |
+
z_start = z[:n_row]
|
1190 |
+
for t in range(self.num_timesteps):
|
1191 |
+
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
1192 |
+
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
1193 |
+
t = t.to(self.device).long()
|
1194 |
+
noise = torch.randn_like(z_start)
|
1195 |
+
z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
|
1196 |
+
diffusion_row.append(self.decode_first_stage(z_noisy))
|
1197 |
+
|
1198 |
+
diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
|
1199 |
+
diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
|
1200 |
+
diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
|
1201 |
+
diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
|
1202 |
+
log["diffusion_row"] = diffusion_grid
|
1203 |
+
|
1204 |
+
if sample:
|
1205 |
+
# get denoise row
|
1206 |
+
with ema_scope("Sampling"):
|
1207 |
+
samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
|
1208 |
+
ddim_steps=ddim_steps, eta=ddim_eta)
|
1209 |
+
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
|
1210 |
+
x_samples = self.decode_first_stage(samples)
|
1211 |
+
log["samples"] = x_samples
|
1212 |
+
if plot_denoise_rows:
|
1213 |
+
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
|
1214 |
+
log["denoise_row"] = denoise_grid
|
1215 |
+
|
1216 |
+
if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
|
1217 |
+
self.first_stage_model, IdentityFirstStage):
|
1218 |
+
# also display when quantizing x0 while sampling
|
1219 |
+
with ema_scope("Plotting Quantized Denoised"):
|
1220 |
+
samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
|
1221 |
+
ddim_steps=ddim_steps, eta=ddim_eta,
|
1222 |
+
quantize_denoised=True)
|
1223 |
+
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
|
1224 |
+
# quantize_denoised=True)
|
1225 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1226 |
+
log["samples_x0_quantized"] = x_samples
|
1227 |
+
|
1228 |
+
if unconditional_guidance_scale > 1.0:
|
1229 |
+
uc = self.get_unconditional_conditioning(N, unconditional_guidance_label)
|
1230 |
+
if self.model.conditioning_key == "crossattn-adm":
|
1231 |
+
uc = {"c_crossattn": [uc], "c_adm": c["c_adm"]}
|
1232 |
+
with ema_scope("Sampling with classifier-free guidance"):
|
1233 |
+
samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
|
1234 |
+
ddim_steps=ddim_steps, eta=ddim_eta,
|
1235 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
1236 |
+
unconditional_conditioning=uc,
|
1237 |
+
)
|
1238 |
+
x_samples_cfg = self.decode_first_stage(samples_cfg)
|
1239 |
+
log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
|
1240 |
+
|
1241 |
+
if inpaint:
|
1242 |
+
# make a simple center square
|
1243 |
+
b, h, w = z.shape[0], z.shape[2], z.shape[3]
|
1244 |
+
mask = torch.ones(N, h, w).to(self.device)
|
1245 |
+
# zeros will be filled in
|
1246 |
+
mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
|
1247 |
+
mask = mask[:, None, ...]
|
1248 |
+
with ema_scope("Plotting Inpaint"):
|
1249 |
+
samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta,
|
1250 |
+
ddim_steps=ddim_steps, x0=z[:N], mask=mask)
|
1251 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1252 |
+
log["samples_inpainting"] = x_samples
|
1253 |
+
log["mask"] = mask
|
1254 |
+
|
1255 |
+
# outpaint
|
1256 |
+
mask = 1. - mask
|
1257 |
+
with ema_scope("Plotting Outpaint"):
|
1258 |
+
samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta,
|
1259 |
+
ddim_steps=ddim_steps, x0=z[:N], mask=mask)
|
1260 |
+
x_samples = self.decode_first_stage(samples.to(self.device))
|
1261 |
+
log["samples_outpainting"] = x_samples
|
1262 |
+
|
1263 |
+
if plot_progressive_rows:
|
1264 |
+
with ema_scope("Plotting Progressives"):
|
1265 |
+
img, progressives = self.progressive_denoising(c,
|
1266 |
+
shape=(self.channels, self.image_size, self.image_size),
|
1267 |
+
batch_size=N)
|
1268 |
+
prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
|
1269 |
+
log["progressive_row"] = prog_row
|
1270 |
+
|
1271 |
+
if return_keys:
|
1272 |
+
if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
|
1273 |
+
return log
|
1274 |
+
else:
|
1275 |
+
return {key: log[key] for key in return_keys}
|
1276 |
+
return log
|
1277 |
+
|
1278 |
+
def configure_optimizers(self):
|
1279 |
+
lr = self.learning_rate
|
1280 |
+
params = list(self.model.parameters())
|
1281 |
+
if self.cond_stage_trainable:
|
1282 |
+
print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
|
1283 |
+
params = params + list(self.cond_stage_model.parameters())
|
1284 |
+
if self.learn_logvar:
|
1285 |
+
print('Diffusion model optimizing logvar')
|
1286 |
+
params.append(self.logvar)
|
1287 |
+
opt = torch.optim.AdamW(params, lr=lr)
|
1288 |
+
if self.use_scheduler:
|
1289 |
+
assert 'target' in self.scheduler_config
|
1290 |
+
scheduler = instantiate_from_config(self.scheduler_config)
|
1291 |
+
|
1292 |
+
print("Setting up LambdaLR scheduler...")
|
1293 |
+
scheduler = [
|
1294 |
+
{
|
1295 |
+
'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
|
1296 |
+
'interval': 'step',
|
1297 |
+
'frequency': 1
|
1298 |
+
}]
|
1299 |
+
return [opt], scheduler
|
1300 |
+
return opt
|
1301 |
+
|
1302 |
+
@torch.no_grad()
|
1303 |
+
def to_rgb(self, x):
|
1304 |
+
x = x.float()
|
1305 |
+
if not hasattr(self, "colorize"):
|
1306 |
+
self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
|
1307 |
+
x = nn.functional.conv2d(x, weight=self.colorize)
|
1308 |
+
x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
|
1309 |
+
return x
|
1310 |
+
|
1311 |
+
|
1312 |
+
class DiffusionWrapper(pl.LightningModule):
|
1313 |
+
def __init__(self, diff_model_config, conditioning_key):
|
1314 |
+
super().__init__()
|
1315 |
+
self.sequential_cross_attn = diff_model_config.pop("sequential_crossattn", False)
|
1316 |
+
self.diffusion_model = instantiate_from_config(diff_model_config)
|
1317 |
+
self.conditioning_key = conditioning_key
|
1318 |
+
assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm', 'hybrid-adm', 'crossattn-adm']
|
1319 |
+
|
1320 |
+
def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, c_adm=None):
|
1321 |
+
if self.conditioning_key is None:
|
1322 |
+
out = self.diffusion_model(x, t)
|
1323 |
+
elif self.conditioning_key == 'concat':
|
1324 |
+
xc = torch.cat([x] + c_concat, dim=1)
|
1325 |
+
out = self.diffusion_model(xc, t)
|
1326 |
+
elif self.conditioning_key == 'crossattn':
|
1327 |
+
if not self.sequential_cross_attn:
|
1328 |
+
cc = torch.cat(c_crossattn, 1)
|
1329 |
+
else:
|
1330 |
+
cc = c_crossattn
|
1331 |
+
out = self.diffusion_model(x, t, context=cc)
|
1332 |
+
elif self.conditioning_key == 'hybrid':
|
1333 |
+
xc = torch.cat([x] + c_concat, dim=1)
|
1334 |
+
cc = torch.cat(c_crossattn, 1)
|
1335 |
+
out = self.diffusion_model(xc, t, context=cc)
|
1336 |
+
elif self.conditioning_key == 'hybrid-adm':
|
1337 |
+
assert c_adm is not None
|
1338 |
+
xc = torch.cat([x] + c_concat, dim=1)
|
1339 |
+
cc = torch.cat(c_crossattn, 1)
|
1340 |
+
out = self.diffusion_model(xc, t, context=cc, y=c_adm)
|
1341 |
+
elif self.conditioning_key == 'crossattn-adm':
|
1342 |
+
assert c_adm is not None
|
1343 |
+
cc = torch.cat(c_crossattn, 1)
|
1344 |
+
out = self.diffusion_model(x, t, context=cc, y=c_adm)
|
1345 |
+
elif self.conditioning_key == 'adm':
|
1346 |
+
cc = c_crossattn[0]
|
1347 |
+
out = self.diffusion_model(x, t, y=cc)
|
1348 |
+
else:
|
1349 |
+
raise NotImplementedError()
|
1350 |
+
|
1351 |
+
return out
|
1352 |
+
|
1353 |
+
|
1354 |
+
class LatentUpscaleDiffusion(LatentDiffusion):
|
1355 |
+
def __init__(self, *args, low_scale_config, low_scale_key="LR", noise_level_key=None, **kwargs):
|
1356 |
+
super().__init__(*args, **kwargs)
|
1357 |
+
# assumes that neither the cond_stage nor the low_scale_model contain trainable params
|
1358 |
+
assert not self.cond_stage_trainable
|
1359 |
+
self.instantiate_low_stage(low_scale_config)
|
1360 |
+
self.low_scale_key = low_scale_key
|
1361 |
+
self.noise_level_key = noise_level_key
|
1362 |
+
|
1363 |
+
def instantiate_low_stage(self, config):
|
1364 |
+
model = instantiate_from_config(config)
|
1365 |
+
self.low_scale_model = model.eval()
|
1366 |
+
self.low_scale_model.train = disabled_train
|
1367 |
+
for param in self.low_scale_model.parameters():
|
1368 |
+
param.requires_grad = False
|
1369 |
+
|
1370 |
+
@torch.no_grad()
|
1371 |
+
def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False):
|
1372 |
+
if not log_mode:
|
1373 |
+
z, c = super().get_input(batch, k, force_c_encode=True, bs=bs)
|
1374 |
+
else:
|
1375 |
+
z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
|
1376 |
+
force_c_encode=True, return_original_cond=True, bs=bs)
|
1377 |
+
x_low = batch[self.low_scale_key][:bs]
|
1378 |
+
x_low = rearrange(x_low, 'b h w c -> b c h w')
|
1379 |
+
x_low = x_low.to(memory_format=torch.contiguous_format).float()
|
1380 |
+
zx, noise_level = self.low_scale_model(x_low)
|
1381 |
+
if self.noise_level_key is not None:
|
1382 |
+
# get noise level from batch instead, e.g. when extracting a custom noise level for bsr
|
1383 |
+
raise NotImplementedError('TODO')
|
1384 |
+
|
1385 |
+
all_conds = {"c_concat": [zx], "c_crossattn": [c], "c_adm": noise_level}
|
1386 |
+
if log_mode:
|
1387 |
+
# TODO: maybe disable if too expensive
|
1388 |
+
x_low_rec = self.low_scale_model.decode(zx)
|
1389 |
+
return z, all_conds, x, xrec, xc, x_low, x_low_rec, noise_level
|
1390 |
+
return z, all_conds
|
1391 |
+
|
1392 |
+
@torch.no_grad()
|
1393 |
+
def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
|
1394 |
+
plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True,
|
1395 |
+
unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True,
|
1396 |
+
**kwargs):
|
1397 |
+
ema_scope = self.ema_scope if use_ema_scope else nullcontext
|
1398 |
+
use_ddim = ddim_steps is not None
|
1399 |
+
|
1400 |
+
log = dict()
|
1401 |
+
z, c, x, xrec, xc, x_low, x_low_rec, noise_level = self.get_input(batch, self.first_stage_key, bs=N,
|
1402 |
+
log_mode=True)
|
1403 |
+
N = min(x.shape[0], N)
|
1404 |
+
n_row = min(x.shape[0], n_row)
|
1405 |
+
log["inputs"] = x
|
1406 |
+
log["reconstruction"] = xrec
|
1407 |
+
log["x_lr"] = x_low
|
1408 |
+
log[f"x_lr_rec_@noise_levels{'-'.join(map(lambda x: str(x), list(noise_level.cpu().numpy())))}"] = x_low_rec
|
1409 |
+
if self.model.conditioning_key is not None:
|
1410 |
+
if hasattr(self.cond_stage_model, "decode"):
|
1411 |
+
xc = self.cond_stage_model.decode(c)
|
1412 |
+
log["conditioning"] = xc
|
1413 |
+
elif self.cond_stage_key in ["caption", "txt"]:
|
1414 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25)
|
1415 |
+
log["conditioning"] = xc
|
1416 |
+
elif self.cond_stage_key in ['class_label', 'cls']:
|
1417 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25)
|
1418 |
+
log['conditioning'] = xc
|
1419 |
+
elif isimage(xc):
|
1420 |
+
log["conditioning"] = xc
|
1421 |
+
if ismap(xc):
|
1422 |
+
log["original_conditioning"] = self.to_rgb(xc)
|
1423 |
+
|
1424 |
+
if plot_diffusion_rows:
|
1425 |
+
# get diffusion row
|
1426 |
+
diffusion_row = list()
|
1427 |
+
z_start = z[:n_row]
|
1428 |
+
for t in range(self.num_timesteps):
|
1429 |
+
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
1430 |
+
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
1431 |
+
t = t.to(self.device).long()
|
1432 |
+
noise = torch.randn_like(z_start)
|
1433 |
+
z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
|
1434 |
+
diffusion_row.append(self.decode_first_stage(z_noisy))
|
1435 |
+
|
1436 |
+
diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
|
1437 |
+
diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
|
1438 |
+
diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
|
1439 |
+
diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
|
1440 |
+
log["diffusion_row"] = diffusion_grid
|
1441 |
+
|
1442 |
+
if sample:
|
1443 |
+
# get denoise row
|
1444 |
+
with ema_scope("Sampling"):
|
1445 |
+
samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
|
1446 |
+
ddim_steps=ddim_steps, eta=ddim_eta)
|
1447 |
+
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
|
1448 |
+
x_samples = self.decode_first_stage(samples)
|
1449 |
+
log["samples"] = x_samples
|
1450 |
+
if plot_denoise_rows:
|
1451 |
+
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
|
1452 |
+
log["denoise_row"] = denoise_grid
|
1453 |
+
|
1454 |
+
if unconditional_guidance_scale > 1.0:
|
1455 |
+
uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label)
|
1456 |
+
# TODO explore better "unconditional" choices for the other keys
|
1457 |
+
# maybe guide away from empty text label and highest noise level and maximally degraded zx?
|
1458 |
+
uc = dict()
|
1459 |
+
for k in c:
|
1460 |
+
if k == "c_crossattn":
|
1461 |
+
assert isinstance(c[k], list) and len(c[k]) == 1
|
1462 |
+
uc[k] = [uc_tmp]
|
1463 |
+
elif k == "c_adm": # todo: only run with text-based guidance?
|
1464 |
+
assert isinstance(c[k], torch.Tensor)
|
1465 |
+
#uc[k] = torch.ones_like(c[k]) * self.low_scale_model.max_noise_level
|
1466 |
+
uc[k] = c[k]
|
1467 |
+
elif isinstance(c[k], list):
|
1468 |
+
uc[k] = [c[k][i] for i in range(len(c[k]))]
|
1469 |
+
else:
|
1470 |
+
uc[k] = c[k]
|
1471 |
+
|
1472 |
+
with ema_scope("Sampling with classifier-free guidance"):
|
1473 |
+
samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,
|
1474 |
+
ddim_steps=ddim_steps, eta=ddim_eta,
|
1475 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
1476 |
+
unconditional_conditioning=uc,
|
1477 |
+
)
|
1478 |
+
x_samples_cfg = self.decode_first_stage(samples_cfg)
|
1479 |
+
log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
|
1480 |
+
|
1481 |
+
if plot_progressive_rows:
|
1482 |
+
with ema_scope("Plotting Progressives"):
|
1483 |
+
img, progressives = self.progressive_denoising(c,
|
1484 |
+
shape=(self.channels, self.image_size, self.image_size),
|
1485 |
+
batch_size=N)
|
1486 |
+
prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
|
1487 |
+
log["progressive_row"] = prog_row
|
1488 |
+
|
1489 |
+
return log
|
1490 |
+
|
1491 |
+
|
1492 |
+
class LatentFinetuneDiffusion(LatentDiffusion):
|
1493 |
+
"""
|
1494 |
+
Basis for different finetunas, such as inpainting or depth2image
|
1495 |
+
To disable finetuning mode, set finetune_keys to None
|
1496 |
+
"""
|
1497 |
+
|
1498 |
+
def __init__(self,
|
1499 |
+
concat_keys: tuple,
|
1500 |
+
finetune_keys=("model.diffusion_model.input_blocks.0.0.weight",
|
1501 |
+
"model_ema.diffusion_modelinput_blocks00weight"
|
1502 |
+
),
|
1503 |
+
keep_finetune_dims=4,
|
1504 |
+
# if model was trained without concat mode before and we would like to keep these channels
|
1505 |
+
c_concat_log_start=None, # to log reconstruction of c_concat codes
|
1506 |
+
c_concat_log_end=None,
|
1507 |
+
*args, **kwargs
|
1508 |
+
):
|
1509 |
+
ckpt_path = kwargs.pop("ckpt_path", None)
|
1510 |
+
ignore_keys = kwargs.pop("ignore_keys", list())
|
1511 |
+
super().__init__(*args, **kwargs)
|
1512 |
+
self.finetune_keys = finetune_keys
|
1513 |
+
self.concat_keys = concat_keys
|
1514 |
+
self.keep_dims = keep_finetune_dims
|
1515 |
+
self.c_concat_log_start = c_concat_log_start
|
1516 |
+
self.c_concat_log_end = c_concat_log_end
|
1517 |
+
if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint'
|
1518 |
+
if exists(ckpt_path):
|
1519 |
+
self.init_from_ckpt(ckpt_path, ignore_keys)
|
1520 |
+
|
1521 |
+
def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
|
1522 |
+
sd = torch.load(path, map_location="cpu")
|
1523 |
+
if "state_dict" in list(sd.keys()):
|
1524 |
+
sd = sd["state_dict"]
|
1525 |
+
keys = list(sd.keys())
|
1526 |
+
for k in keys:
|
1527 |
+
for ik in ignore_keys:
|
1528 |
+
if k.startswith(ik):
|
1529 |
+
print("Deleting key {} from state_dict.".format(k))
|
1530 |
+
del sd[k]
|
1531 |
+
|
1532 |
+
# make it explicit, finetune by including extra input channels
|
1533 |
+
if exists(self.finetune_keys) and k in self.finetune_keys:
|
1534 |
+
new_entry = None
|
1535 |
+
for name, param in self.named_parameters():
|
1536 |
+
if name in self.finetune_keys:
|
1537 |
+
print(
|
1538 |
+
f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only")
|
1539 |
+
new_entry = torch.zeros_like(param) # zero init
|
1540 |
+
assert exists(new_entry), 'did not find matching parameter to modify'
|
1541 |
+
new_entry[:, :self.keep_dims, ...] = sd[k]
|
1542 |
+
sd[k] = new_entry
|
1543 |
+
|
1544 |
+
missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
|
1545 |
+
sd, strict=False)
|
1546 |
+
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
1547 |
+
if len(missing) > 0:
|
1548 |
+
print(f"Missing Keys: {missing}")
|
1549 |
+
if len(unexpected) > 0:
|
1550 |
+
print(f"Unexpected Keys: {unexpected}")
|
1551 |
+
|
1552 |
+
@torch.no_grad()
|
1553 |
+
def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
|
1554 |
+
quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
|
1555 |
+
plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None,
|
1556 |
+
use_ema_scope=True,
|
1557 |
+
**kwargs):
|
1558 |
+
ema_scope = self.ema_scope if use_ema_scope else nullcontext
|
1559 |
+
use_ddim = ddim_steps is not None
|
1560 |
+
|
1561 |
+
log = dict()
|
1562 |
+
z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True)
|
1563 |
+
c_cat, c = c["c_concat"][0], c["c_crossattn"][0]
|
1564 |
+
N = min(x.shape[0], N)
|
1565 |
+
n_row = min(x.shape[0], n_row)
|
1566 |
+
log["inputs"] = x
|
1567 |
+
log["reconstruction"] = xrec
|
1568 |
+
if self.model.conditioning_key is not None:
|
1569 |
+
if hasattr(self.cond_stage_model, "decode"):
|
1570 |
+
xc = self.cond_stage_model.decode(c)
|
1571 |
+
log["conditioning"] = xc
|
1572 |
+
elif self.cond_stage_key in ["caption", "txt"]:
|
1573 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25)
|
1574 |
+
log["conditioning"] = xc
|
1575 |
+
elif self.cond_stage_key in ['class_label', 'cls']:
|
1576 |
+
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25)
|
1577 |
+
log['conditioning'] = xc
|
1578 |
+
elif isimage(xc):
|
1579 |
+
log["conditioning"] = xc
|
1580 |
+
if ismap(xc):
|
1581 |
+
log["original_conditioning"] = self.to_rgb(xc)
|
1582 |
+
|
1583 |
+
if not (self.c_concat_log_start is None and self.c_concat_log_end is None):
|
1584 |
+
log["c_concat_decoded"] = self.decode_first_stage(c_cat[:, self.c_concat_log_start:self.c_concat_log_end])
|
1585 |
+
|
1586 |
+
if plot_diffusion_rows:
|
1587 |
+
# get diffusion row
|
1588 |
+
diffusion_row = list()
|
1589 |
+
z_start = z[:n_row]
|
1590 |
+
for t in range(self.num_timesteps):
|
1591 |
+
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
1592 |
+
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
1593 |
+
t = t.to(self.device).long()
|
1594 |
+
noise = torch.randn_like(z_start)
|
1595 |
+
z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
|
1596 |
+
diffusion_row.append(self.decode_first_stage(z_noisy))
|
1597 |
+
|
1598 |
+
diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
|
1599 |
+
diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
|
1600 |
+
diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
|
1601 |
+
diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
|
1602 |
+
log["diffusion_row"] = diffusion_grid
|
1603 |
+
|
1604 |
+
if sample:
|
1605 |
+
# get denoise row
|
1606 |
+
with ema_scope("Sampling"):
|
1607 |
+
samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
|
1608 |
+
batch_size=N, ddim=use_ddim,
|
1609 |
+
ddim_steps=ddim_steps, eta=ddim_eta)
|
1610 |
+
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
|
1611 |
+
x_samples = self.decode_first_stage(samples)
|
1612 |
+
log["samples"] = x_samples
|
1613 |
+
if plot_denoise_rows:
|
1614 |
+
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
|
1615 |
+
log["denoise_row"] = denoise_grid
|
1616 |
+
|
1617 |
+
if unconditional_guidance_scale > 1.0:
|
1618 |
+
uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label)
|
1619 |
+
uc_cat = c_cat
|
1620 |
+
uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]}
|
1621 |
+
with ema_scope("Sampling with classifier-free guidance"):
|
1622 |
+
samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
|
1623 |
+
batch_size=N, ddim=use_ddim,
|
1624 |
+
ddim_steps=ddim_steps, eta=ddim_eta,
|
1625 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
1626 |
+
unconditional_conditioning=uc_full,
|
1627 |
+
)
|
1628 |
+
x_samples_cfg = self.decode_first_stage(samples_cfg)
|
1629 |
+
log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
|
1630 |
+
|
1631 |
+
return log
|
1632 |
+
|
1633 |
+
|
1634 |
+
class LatentInpaintDiffusion(LatentFinetuneDiffusion):
|
1635 |
+
"""
|
1636 |
+
can either run as pure inpainting model (only concat mode) or with mixed conditionings,
|
1637 |
+
e.g. mask as concat and text via cross-attn.
|
1638 |
+
To disable finetuning mode, set finetune_keys to None
|
1639 |
+
"""
|
1640 |
+
|
1641 |
+
def __init__(self,
|
1642 |
+
concat_keys=("mask", "masked_image"),
|
1643 |
+
masked_image_key="masked_image",
|
1644 |
+
*args, **kwargs
|
1645 |
+
):
|
1646 |
+
super().__init__(concat_keys, *args, **kwargs)
|
1647 |
+
self.masked_image_key = masked_image_key
|
1648 |
+
assert self.masked_image_key in concat_keys
|
1649 |
+
|
1650 |
+
@torch.no_grad()
|
1651 |
+
def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
|
1652 |
+
# note: restricted to non-trainable encoders currently
|
1653 |
+
assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpainting'
|
1654 |
+
z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
|
1655 |
+
force_c_encode=True, return_original_cond=True, bs=bs)
|
1656 |
+
|
1657 |
+
assert exists(self.concat_keys)
|
1658 |
+
c_cat = list()
|
1659 |
+
for ck in self.concat_keys:
|
1660 |
+
cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
|
1661 |
+
if bs is not None:
|
1662 |
+
cc = cc[:bs]
|
1663 |
+
cc = cc.to(self.device)
|
1664 |
+
bchw = z.shape
|
1665 |
+
if ck != self.masked_image_key:
|
1666 |
+
cc = torch.nn.functional.interpolate(cc, size=bchw[-2:])
|
1667 |
+
else:
|
1668 |
+
cc = self.get_first_stage_encoding(self.encode_first_stage(cc))
|
1669 |
+
c_cat.append(cc)
|
1670 |
+
c_cat = torch.cat(c_cat, dim=1)
|
1671 |
+
all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
|
1672 |
+
if return_first_stage_outputs:
|
1673 |
+
return z, all_conds, x, xrec, xc
|
1674 |
+
return z, all_conds
|
1675 |
+
|
1676 |
+
@torch.no_grad()
|
1677 |
+
def log_images(self, *args, **kwargs):
|
1678 |
+
log = super(LatentInpaintDiffusion, self).log_images(*args, **kwargs)
|
1679 |
+
log["masked_image"] = rearrange(args[0]["masked_image"],
|
1680 |
+
'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
|
1681 |
+
return log
|
1682 |
+
|
1683 |
+
|
1684 |
+
class LatentDepth2ImageDiffusion(LatentFinetuneDiffusion):
|
1685 |
+
"""
|
1686 |
+
condition on monocular depth estimation
|
1687 |
+
"""
|
1688 |
+
|
1689 |
+
def __init__(self, depth_stage_config, concat_keys=("midas_in",), *args, **kwargs):
|
1690 |
+
super().__init__(concat_keys=concat_keys, *args, **kwargs)
|
1691 |
+
self.depth_model = instantiate_from_config(depth_stage_config)
|
1692 |
+
self.depth_stage_key = concat_keys[0]
|
1693 |
+
|
1694 |
+
@torch.no_grad()
|
1695 |
+
def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
|
1696 |
+
# note: restricted to non-trainable encoders currently
|
1697 |
+
assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for depth2img'
|
1698 |
+
z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
|
1699 |
+
force_c_encode=True, return_original_cond=True, bs=bs)
|
1700 |
+
|
1701 |
+
assert exists(self.concat_keys)
|
1702 |
+
assert len(self.concat_keys) == 1
|
1703 |
+
c_cat = list()
|
1704 |
+
for ck in self.concat_keys:
|
1705 |
+
cc = batch[ck]
|
1706 |
+
if bs is not None:
|
1707 |
+
cc = cc[:bs]
|
1708 |
+
cc = cc.to(self.device)
|
1709 |
+
cc = self.depth_model(cc)
|
1710 |
+
cc = torch.nn.functional.interpolate(
|
1711 |
+
cc,
|
1712 |
+
size=z.shape[2:],
|
1713 |
+
mode="bicubic",
|
1714 |
+
align_corners=False,
|
1715 |
+
)
|
1716 |
+
|
1717 |
+
depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3],
|
1718 |
+
keepdim=True)
|
1719 |
+
cc = 2. * (cc - depth_min) / (depth_max - depth_min + 0.001) - 1.
|
1720 |
+
c_cat.append(cc)
|
1721 |
+
c_cat = torch.cat(c_cat, dim=1)
|
1722 |
+
all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
|
1723 |
+
if return_first_stage_outputs:
|
1724 |
+
return z, all_conds, x, xrec, xc
|
1725 |
+
return z, all_conds
|
1726 |
+
|
1727 |
+
@torch.no_grad()
|
1728 |
+
def log_images(self, *args, **kwargs):
|
1729 |
+
log = super().log_images(*args, **kwargs)
|
1730 |
+
depth = self.depth_model(args[0][self.depth_stage_key])
|
1731 |
+
depth_min, depth_max = torch.amin(depth, dim=[1, 2, 3], keepdim=True), \
|
1732 |
+
torch.amax(depth, dim=[1, 2, 3], keepdim=True)
|
1733 |
+
log["depth"] = 2. * (depth - depth_min) / (depth_max - depth_min) - 1.
|
1734 |
+
return log
|
1735 |
+
|
1736 |
+
|
1737 |
+
class LatentUpscaleFinetuneDiffusion(LatentFinetuneDiffusion):
|
1738 |
+
"""
|
1739 |
+
condition on low-res image (and optionally on some spatial noise augmentation)
|
1740 |
+
"""
|
1741 |
+
def __init__(self, concat_keys=("lr",), reshuffle_patch_size=None,
|
1742 |
+
low_scale_config=None, low_scale_key=None, *args, **kwargs):
|
1743 |
+
super().__init__(concat_keys=concat_keys, *args, **kwargs)
|
1744 |
+
self.reshuffle_patch_size = reshuffle_patch_size
|
1745 |
+
self.low_scale_model = None
|
1746 |
+
if low_scale_config is not None:
|
1747 |
+
print("Initializing a low-scale model")
|
1748 |
+
assert exists(low_scale_key)
|
1749 |
+
self.instantiate_low_stage(low_scale_config)
|
1750 |
+
self.low_scale_key = low_scale_key
|
1751 |
+
|
1752 |
+
def instantiate_low_stage(self, config):
|
1753 |
+
model = instantiate_from_config(config)
|
1754 |
+
self.low_scale_model = model.eval()
|
1755 |
+
self.low_scale_model.train = disabled_train
|
1756 |
+
for param in self.low_scale_model.parameters():
|
1757 |
+
param.requires_grad = False
|
1758 |
+
|
1759 |
+
@torch.no_grad()
|
1760 |
+
def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
|
1761 |
+
# note: restricted to non-trainable encoders currently
|
1762 |
+
assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for upscaling-ft'
|
1763 |
+
z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
|
1764 |
+
force_c_encode=True, return_original_cond=True, bs=bs)
|
1765 |
+
|
1766 |
+
assert exists(self.concat_keys)
|
1767 |
+
assert len(self.concat_keys) == 1
|
1768 |
+
# optionally make spatial noise_level here
|
1769 |
+
c_cat = list()
|
1770 |
+
noise_level = None
|
1771 |
+
for ck in self.concat_keys:
|
1772 |
+
cc = batch[ck]
|
1773 |
+
cc = rearrange(cc, 'b h w c -> b c h w')
|
1774 |
+
if exists(self.reshuffle_patch_size):
|
1775 |
+
assert isinstance(self.reshuffle_patch_size, int)
|
1776 |
+
cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w',
|
1777 |
+
p1=self.reshuffle_patch_size, p2=self.reshuffle_patch_size)
|
1778 |
+
if bs is not None:
|
1779 |
+
cc = cc[:bs]
|
1780 |
+
cc = cc.to(self.device)
|
1781 |
+
if exists(self.low_scale_model) and ck == self.low_scale_key:
|
1782 |
+
cc, noise_level = self.low_scale_model(cc)
|
1783 |
+
c_cat.append(cc)
|
1784 |
+
c_cat = torch.cat(c_cat, dim=1)
|
1785 |
+
if exists(noise_level):
|
1786 |
+
all_conds = {"c_concat": [c_cat], "c_crossattn": [c], "c_adm": noise_level}
|
1787 |
+
else:
|
1788 |
+
all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
|
1789 |
+
if return_first_stage_outputs:
|
1790 |
+
return z, all_conds, x, xrec, xc
|
1791 |
+
return z, all_conds
|
1792 |
+
|
1793 |
+
@torch.no_grad()
|
1794 |
+
def log_images(self, *args, **kwargs):
|
1795 |
+
log = super().log_images(*args, **kwargs)
|
1796 |
+
log["lr"] = rearrange(args[0]["lr"], 'b h w c -> b c h w')
|
1797 |
+
return log
|
ControlNet/ldm/models/diffusion/dpm_solver/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
from .sampler import DPMSolverSampler
|
ControlNet/ldm/models/diffusion/dpm_solver/dpm_solver.py
ADDED
@@ -0,0 +1,1154 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as F
|
3 |
+
import math
|
4 |
+
from tqdm import tqdm
|
5 |
+
|
6 |
+
|
7 |
+
class NoiseScheduleVP:
|
8 |
+
def __init__(
|
9 |
+
self,
|
10 |
+
schedule='discrete',
|
11 |
+
betas=None,
|
12 |
+
alphas_cumprod=None,
|
13 |
+
continuous_beta_0=0.1,
|
14 |
+
continuous_beta_1=20.,
|
15 |
+
):
|
16 |
+
"""Create a wrapper class for the forward SDE (VP type).
|
17 |
+
***
|
18 |
+
Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
|
19 |
+
We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
|
20 |
+
***
|
21 |
+
The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
|
22 |
+
We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
|
23 |
+
Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
|
24 |
+
log_alpha_t = self.marginal_log_mean_coeff(t)
|
25 |
+
sigma_t = self.marginal_std(t)
|
26 |
+
lambda_t = self.marginal_lambda(t)
|
27 |
+
Moreover, as lambda(t) is an invertible function, we also support its inverse function:
|
28 |
+
t = self.inverse_lambda(lambda_t)
|
29 |
+
===============================================================
|
30 |
+
We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
|
31 |
+
1. For discrete-time DPMs:
|
32 |
+
For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
|
33 |
+
t_i = (i + 1) / N
|
34 |
+
e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
|
35 |
+
We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
|
36 |
+
Args:
|
37 |
+
betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
|
38 |
+
alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
|
39 |
+
Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
|
40 |
+
**Important**: Please pay special attention for the args for `alphas_cumprod`:
|
41 |
+
The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
|
42 |
+
q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
|
43 |
+
Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
|
44 |
+
alpha_{t_n} = \sqrt{\hat{alpha_n}},
|
45 |
+
and
|
46 |
+
log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
|
47 |
+
2. For continuous-time DPMs:
|
48 |
+
We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
|
49 |
+
schedule are the default settings in DDPM and improved-DDPM:
|
50 |
+
Args:
|
51 |
+
beta_min: A `float` number. The smallest beta for the linear schedule.
|
52 |
+
beta_max: A `float` number. The largest beta for the linear schedule.
|
53 |
+
cosine_s: A `float` number. The hyperparameter in the cosine schedule.
|
54 |
+
cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
|
55 |
+
T: A `float` number. The ending time of the forward process.
|
56 |
+
===============================================================
|
57 |
+
Args:
|
58 |
+
schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
|
59 |
+
'linear' or 'cosine' for continuous-time DPMs.
|
60 |
+
Returns:
|
61 |
+
A wrapper object of the forward SDE (VP type).
|
62 |
+
|
63 |
+
===============================================================
|
64 |
+
Example:
|
65 |
+
# For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
|
66 |
+
>>> ns = NoiseScheduleVP('discrete', betas=betas)
|
67 |
+
# For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
|
68 |
+
>>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
|
69 |
+
# For continuous-time DPMs (VPSDE), linear schedule:
|
70 |
+
>>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
|
71 |
+
"""
|
72 |
+
|
73 |
+
if schedule not in ['discrete', 'linear', 'cosine']:
|
74 |
+
raise ValueError(
|
75 |
+
"Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(
|
76 |
+
schedule))
|
77 |
+
|
78 |
+
self.schedule = schedule
|
79 |
+
if schedule == 'discrete':
|
80 |
+
if betas is not None:
|
81 |
+
log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
|
82 |
+
else:
|
83 |
+
assert alphas_cumprod is not None
|
84 |
+
log_alphas = 0.5 * torch.log(alphas_cumprod)
|
85 |
+
self.total_N = len(log_alphas)
|
86 |
+
self.T = 1.
|
87 |
+
self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1))
|
88 |
+
self.log_alpha_array = log_alphas.reshape((1, -1,))
|
89 |
+
else:
|
90 |
+
self.total_N = 1000
|
91 |
+
self.beta_0 = continuous_beta_0
|
92 |
+
self.beta_1 = continuous_beta_1
|
93 |
+
self.cosine_s = 0.008
|
94 |
+
self.cosine_beta_max = 999.
|
95 |
+
self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * (
|
96 |
+
1. + self.cosine_s) / math.pi - self.cosine_s
|
97 |
+
self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.))
|
98 |
+
self.schedule = schedule
|
99 |
+
if schedule == 'cosine':
|
100 |
+
# For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
|
101 |
+
# Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
|
102 |
+
self.T = 0.9946
|
103 |
+
else:
|
104 |
+
self.T = 1.
|
105 |
+
|
106 |
+
def marginal_log_mean_coeff(self, t):
|
107 |
+
"""
|
108 |
+
Compute log(alpha_t) of a given continuous-time label t in [0, T].
|
109 |
+
"""
|
110 |
+
if self.schedule == 'discrete':
|
111 |
+
return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device),
|
112 |
+
self.log_alpha_array.to(t.device)).reshape((-1))
|
113 |
+
elif self.schedule == 'linear':
|
114 |
+
return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
|
115 |
+
elif self.schedule == 'cosine':
|
116 |
+
log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.))
|
117 |
+
log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
|
118 |
+
return log_alpha_t
|
119 |
+
|
120 |
+
def marginal_alpha(self, t):
|
121 |
+
"""
|
122 |
+
Compute alpha_t of a given continuous-time label t in [0, T].
|
123 |
+
"""
|
124 |
+
return torch.exp(self.marginal_log_mean_coeff(t))
|
125 |
+
|
126 |
+
def marginal_std(self, t):
|
127 |
+
"""
|
128 |
+
Compute sigma_t of a given continuous-time label t in [0, T].
|
129 |
+
"""
|
130 |
+
return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t)))
|
131 |
+
|
132 |
+
def marginal_lambda(self, t):
|
133 |
+
"""
|
134 |
+
Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
|
135 |
+
"""
|
136 |
+
log_mean_coeff = self.marginal_log_mean_coeff(t)
|
137 |
+
log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff))
|
138 |
+
return log_mean_coeff - log_std
|
139 |
+
|
140 |
+
def inverse_lambda(self, lamb):
|
141 |
+
"""
|
142 |
+
Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
|
143 |
+
"""
|
144 |
+
if self.schedule == 'linear':
|
145 |
+
tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
|
146 |
+
Delta = self.beta_0 ** 2 + tmp
|
147 |
+
return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
|
148 |
+
elif self.schedule == 'discrete':
|
149 |
+
log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb)
|
150 |
+
t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]),
|
151 |
+
torch.flip(self.t_array.to(lamb.device), [1]))
|
152 |
+
return t.reshape((-1,))
|
153 |
+
else:
|
154 |
+
log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
|
155 |
+
t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * (
|
156 |
+
1. + self.cosine_s) / math.pi - self.cosine_s
|
157 |
+
t = t_fn(log_alpha)
|
158 |
+
return t
|
159 |
+
|
160 |
+
|
161 |
+
def model_wrapper(
|
162 |
+
model,
|
163 |
+
noise_schedule,
|
164 |
+
model_type="noise",
|
165 |
+
model_kwargs={},
|
166 |
+
guidance_type="uncond",
|
167 |
+
condition=None,
|
168 |
+
unconditional_condition=None,
|
169 |
+
guidance_scale=1.,
|
170 |
+
classifier_fn=None,
|
171 |
+
classifier_kwargs={},
|
172 |
+
):
|
173 |
+
"""Create a wrapper function for the noise prediction model.
|
174 |
+
DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
|
175 |
+
firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
|
176 |
+
We support four types of the diffusion model by setting `model_type`:
|
177 |
+
1. "noise": noise prediction model. (Trained by predicting noise).
|
178 |
+
2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
|
179 |
+
3. "v": velocity prediction model. (Trained by predicting the velocity).
|
180 |
+
The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
|
181 |
+
[1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
|
182 |
+
arXiv preprint arXiv:2202.00512 (2022).
|
183 |
+
[2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
|
184 |
+
arXiv preprint arXiv:2210.02303 (2022).
|
185 |
+
|
186 |
+
4. "score": marginal score function. (Trained by denoising score matching).
|
187 |
+
Note that the score function and the noise prediction model follows a simple relationship:
|
188 |
+
```
|
189 |
+
noise(x_t, t) = -sigma_t * score(x_t, t)
|
190 |
+
```
|
191 |
+
We support three types of guided sampling by DPMs by setting `guidance_type`:
|
192 |
+
1. "uncond": unconditional sampling by DPMs.
|
193 |
+
The input `model` has the following format:
|
194 |
+
``
|
195 |
+
model(x, t_input, **model_kwargs) -> noise | x_start | v | score
|
196 |
+
``
|
197 |
+
2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
|
198 |
+
The input `model` has the following format:
|
199 |
+
``
|
200 |
+
model(x, t_input, **model_kwargs) -> noise | x_start | v | score
|
201 |
+
``
|
202 |
+
The input `classifier_fn` has the following format:
|
203 |
+
``
|
204 |
+
classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
|
205 |
+
``
|
206 |
+
[3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
|
207 |
+
in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
|
208 |
+
3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
|
209 |
+
The input `model` has the following format:
|
210 |
+
``
|
211 |
+
model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
|
212 |
+
``
|
213 |
+
And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
|
214 |
+
[4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
|
215 |
+
arXiv preprint arXiv:2207.12598 (2022).
|
216 |
+
|
217 |
+
The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
|
218 |
+
or continuous-time labels (i.e. epsilon to T).
|
219 |
+
We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
|
220 |
+
``
|
221 |
+
def model_fn(x, t_continuous) -> noise:
|
222 |
+
t_input = get_model_input_time(t_continuous)
|
223 |
+
return noise_pred(model, x, t_input, **model_kwargs)
|
224 |
+
``
|
225 |
+
where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
|
226 |
+
===============================================================
|
227 |
+
Args:
|
228 |
+
model: A diffusion model with the corresponding format described above.
|
229 |
+
noise_schedule: A noise schedule object, such as NoiseScheduleVP.
|
230 |
+
model_type: A `str`. The parameterization type of the diffusion model.
|
231 |
+
"noise" or "x_start" or "v" or "score".
|
232 |
+
model_kwargs: A `dict`. A dict for the other inputs of the model function.
|
233 |
+
guidance_type: A `str`. The type of the guidance for sampling.
|
234 |
+
"uncond" or "classifier" or "classifier-free".
|
235 |
+
condition: A pytorch tensor. The condition for the guided sampling.
|
236 |
+
Only used for "classifier" or "classifier-free" guidance type.
|
237 |
+
unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
|
238 |
+
Only used for "classifier-free" guidance type.
|
239 |
+
guidance_scale: A `float`. The scale for the guided sampling.
|
240 |
+
classifier_fn: A classifier function. Only used for the classifier guidance.
|
241 |
+
classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
|
242 |
+
Returns:
|
243 |
+
A noise prediction model that accepts the noised data and the continuous time as the inputs.
|
244 |
+
"""
|
245 |
+
|
246 |
+
def get_model_input_time(t_continuous):
|
247 |
+
"""
|
248 |
+
Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
|
249 |
+
For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
|
250 |
+
For continuous-time DPMs, we just use `t_continuous`.
|
251 |
+
"""
|
252 |
+
if noise_schedule.schedule == 'discrete':
|
253 |
+
return (t_continuous - 1. / noise_schedule.total_N) * 1000.
|
254 |
+
else:
|
255 |
+
return t_continuous
|
256 |
+
|
257 |
+
def noise_pred_fn(x, t_continuous, cond=None):
|
258 |
+
if t_continuous.reshape((-1,)).shape[0] == 1:
|
259 |
+
t_continuous = t_continuous.expand((x.shape[0]))
|
260 |
+
t_input = get_model_input_time(t_continuous)
|
261 |
+
if cond is None:
|
262 |
+
output = model(x, t_input, **model_kwargs)
|
263 |
+
else:
|
264 |
+
output = model(x, t_input, cond, **model_kwargs)
|
265 |
+
if model_type == "noise":
|
266 |
+
return output
|
267 |
+
elif model_type == "x_start":
|
268 |
+
alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
|
269 |
+
dims = x.dim()
|
270 |
+
return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims)
|
271 |
+
elif model_type == "v":
|
272 |
+
alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
|
273 |
+
dims = x.dim()
|
274 |
+
return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x
|
275 |
+
elif model_type == "score":
|
276 |
+
sigma_t = noise_schedule.marginal_std(t_continuous)
|
277 |
+
dims = x.dim()
|
278 |
+
return -expand_dims(sigma_t, dims) * output
|
279 |
+
|
280 |
+
def cond_grad_fn(x, t_input):
|
281 |
+
"""
|
282 |
+
Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
|
283 |
+
"""
|
284 |
+
with torch.enable_grad():
|
285 |
+
x_in = x.detach().requires_grad_(True)
|
286 |
+
log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
|
287 |
+
return torch.autograd.grad(log_prob.sum(), x_in)[0]
|
288 |
+
|
289 |
+
def model_fn(x, t_continuous):
|
290 |
+
"""
|
291 |
+
The noise predicition model function that is used for DPM-Solver.
|
292 |
+
"""
|
293 |
+
if t_continuous.reshape((-1,)).shape[0] == 1:
|
294 |
+
t_continuous = t_continuous.expand((x.shape[0]))
|
295 |
+
if guidance_type == "uncond":
|
296 |
+
return noise_pred_fn(x, t_continuous)
|
297 |
+
elif guidance_type == "classifier":
|
298 |
+
assert classifier_fn is not None
|
299 |
+
t_input = get_model_input_time(t_continuous)
|
300 |
+
cond_grad = cond_grad_fn(x, t_input)
|
301 |
+
sigma_t = noise_schedule.marginal_std(t_continuous)
|
302 |
+
noise = noise_pred_fn(x, t_continuous)
|
303 |
+
return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad
|
304 |
+
elif guidance_type == "classifier-free":
|
305 |
+
if guidance_scale == 1. or unconditional_condition is None:
|
306 |
+
return noise_pred_fn(x, t_continuous, cond=condition)
|
307 |
+
else:
|
308 |
+
x_in = torch.cat([x] * 2)
|
309 |
+
t_in = torch.cat([t_continuous] * 2)
|
310 |
+
c_in = torch.cat([unconditional_condition, condition])
|
311 |
+
noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
|
312 |
+
return noise_uncond + guidance_scale * (noise - noise_uncond)
|
313 |
+
|
314 |
+
assert model_type in ["noise", "x_start", "v"]
|
315 |
+
assert guidance_type in ["uncond", "classifier", "classifier-free"]
|
316 |
+
return model_fn
|
317 |
+
|
318 |
+
|
319 |
+
class DPM_Solver:
|
320 |
+
def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.):
|
321 |
+
"""Construct a DPM-Solver.
|
322 |
+
We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0").
|
323 |
+
If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver).
|
324 |
+
If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++).
|
325 |
+
In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True.
|
326 |
+
The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales.
|
327 |
+
Args:
|
328 |
+
model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
|
329 |
+
``
|
330 |
+
def model_fn(x, t_continuous):
|
331 |
+
return noise
|
332 |
+
``
|
333 |
+
noise_schedule: A noise schedule object, such as NoiseScheduleVP.
|
334 |
+
predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model.
|
335 |
+
thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1].
|
336 |
+
max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding.
|
337 |
+
|
338 |
+
[1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
|
339 |
+
"""
|
340 |
+
self.model = model_fn
|
341 |
+
self.noise_schedule = noise_schedule
|
342 |
+
self.predict_x0 = predict_x0
|
343 |
+
self.thresholding = thresholding
|
344 |
+
self.max_val = max_val
|
345 |
+
|
346 |
+
def noise_prediction_fn(self, x, t):
|
347 |
+
"""
|
348 |
+
Return the noise prediction model.
|
349 |
+
"""
|
350 |
+
return self.model(x, t)
|
351 |
+
|
352 |
+
def data_prediction_fn(self, x, t):
|
353 |
+
"""
|
354 |
+
Return the data prediction model (with thresholding).
|
355 |
+
"""
|
356 |
+
noise = self.noise_prediction_fn(x, t)
|
357 |
+
dims = x.dim()
|
358 |
+
alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t)
|
359 |
+
x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims)
|
360 |
+
if self.thresholding:
|
361 |
+
p = 0.995 # A hyperparameter in the paper of "Imagen" [1].
|
362 |
+
s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
|
363 |
+
s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims)
|
364 |
+
x0 = torch.clamp(x0, -s, s) / s
|
365 |
+
return x0
|
366 |
+
|
367 |
+
def model_fn(self, x, t):
|
368 |
+
"""
|
369 |
+
Convert the model to the noise prediction model or the data prediction model.
|
370 |
+
"""
|
371 |
+
if self.predict_x0:
|
372 |
+
return self.data_prediction_fn(x, t)
|
373 |
+
else:
|
374 |
+
return self.noise_prediction_fn(x, t)
|
375 |
+
|
376 |
+
def get_time_steps(self, skip_type, t_T, t_0, N, device):
|
377 |
+
"""Compute the intermediate time steps for sampling.
|
378 |
+
Args:
|
379 |
+
skip_type: A `str`. The type for the spacing of the time steps. We support three types:
|
380 |
+
- 'logSNR': uniform logSNR for the time steps.
|
381 |
+
- 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
|
382 |
+
- 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
|
383 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
384 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
385 |
+
N: A `int`. The total number of the spacing of the time steps.
|
386 |
+
device: A torch device.
|
387 |
+
Returns:
|
388 |
+
A pytorch tensor of the time steps, with the shape (N + 1,).
|
389 |
+
"""
|
390 |
+
if skip_type == 'logSNR':
|
391 |
+
lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
|
392 |
+
lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
|
393 |
+
logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device)
|
394 |
+
return self.noise_schedule.inverse_lambda(logSNR_steps)
|
395 |
+
elif skip_type == 'time_uniform':
|
396 |
+
return torch.linspace(t_T, t_0, N + 1).to(device)
|
397 |
+
elif skip_type == 'time_quadratic':
|
398 |
+
t_order = 2
|
399 |
+
t = torch.linspace(t_T ** (1. / t_order), t_0 ** (1. / t_order), N + 1).pow(t_order).to(device)
|
400 |
+
return t
|
401 |
+
else:
|
402 |
+
raise ValueError(
|
403 |
+
"Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type))
|
404 |
+
|
405 |
+
def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device):
|
406 |
+
"""
|
407 |
+
Get the order of each step for sampling by the singlestep DPM-Solver.
|
408 |
+
We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
|
409 |
+
Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
|
410 |
+
- If order == 1:
|
411 |
+
We take `steps` of DPM-Solver-1 (i.e. DDIM).
|
412 |
+
- If order == 2:
|
413 |
+
- Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
|
414 |
+
- If steps % 2 == 0, we use K steps of DPM-Solver-2.
|
415 |
+
- If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
|
416 |
+
- If order == 3:
|
417 |
+
- Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
|
418 |
+
- If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
|
419 |
+
- If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
|
420 |
+
- If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
|
421 |
+
============================================
|
422 |
+
Args:
|
423 |
+
order: A `int`. The max order for the solver (2 or 3).
|
424 |
+
steps: A `int`. The total number of function evaluations (NFE).
|
425 |
+
skip_type: A `str`. The type for the spacing of the time steps. We support three types:
|
426 |
+
- 'logSNR': uniform logSNR for the time steps.
|
427 |
+
- 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
|
428 |
+
- 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
|
429 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
430 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
431 |
+
device: A torch device.
|
432 |
+
Returns:
|
433 |
+
orders: A list of the solver order of each step.
|
434 |
+
"""
|
435 |
+
if order == 3:
|
436 |
+
K = steps // 3 + 1
|
437 |
+
if steps % 3 == 0:
|
438 |
+
orders = [3, ] * (K - 2) + [2, 1]
|
439 |
+
elif steps % 3 == 1:
|
440 |
+
orders = [3, ] * (K - 1) + [1]
|
441 |
+
else:
|
442 |
+
orders = [3, ] * (K - 1) + [2]
|
443 |
+
elif order == 2:
|
444 |
+
if steps % 2 == 0:
|
445 |
+
K = steps // 2
|
446 |
+
orders = [2, ] * K
|
447 |
+
else:
|
448 |
+
K = steps // 2 + 1
|
449 |
+
orders = [2, ] * (K - 1) + [1]
|
450 |
+
elif order == 1:
|
451 |
+
K = 1
|
452 |
+
orders = [1, ] * steps
|
453 |
+
else:
|
454 |
+
raise ValueError("'order' must be '1' or '2' or '3'.")
|
455 |
+
if skip_type == 'logSNR':
|
456 |
+
# To reproduce the results in DPM-Solver paper
|
457 |
+
timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
|
458 |
+
else:
|
459 |
+
timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[
|
460 |
+
torch.cumsum(torch.tensor([0, ] + orders)).to(device)]
|
461 |
+
return timesteps_outer, orders
|
462 |
+
|
463 |
+
def denoise_to_zero_fn(self, x, s):
|
464 |
+
"""
|
465 |
+
Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
|
466 |
+
"""
|
467 |
+
return self.data_prediction_fn(x, s)
|
468 |
+
|
469 |
+
def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
|
470 |
+
"""
|
471 |
+
DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
|
472 |
+
Args:
|
473 |
+
x: A pytorch tensor. The initial value at time `s`.
|
474 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
475 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
476 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
477 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
478 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`.
|
479 |
+
Returns:
|
480 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
481 |
+
"""
|
482 |
+
ns = self.noise_schedule
|
483 |
+
dims = x.dim()
|
484 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
485 |
+
h = lambda_t - lambda_s
|
486 |
+
log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t)
|
487 |
+
sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
|
488 |
+
alpha_t = torch.exp(log_alpha_t)
|
489 |
+
|
490 |
+
if self.predict_x0:
|
491 |
+
phi_1 = torch.expm1(-h)
|
492 |
+
if model_s is None:
|
493 |
+
model_s = self.model_fn(x, s)
|
494 |
+
x_t = (
|
495 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
496 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
497 |
+
)
|
498 |
+
if return_intermediate:
|
499 |
+
return x_t, {'model_s': model_s}
|
500 |
+
else:
|
501 |
+
return x_t
|
502 |
+
else:
|
503 |
+
phi_1 = torch.expm1(h)
|
504 |
+
if model_s is None:
|
505 |
+
model_s = self.model_fn(x, s)
|
506 |
+
x_t = (
|
507 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
508 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
509 |
+
)
|
510 |
+
if return_intermediate:
|
511 |
+
return x_t, {'model_s': model_s}
|
512 |
+
else:
|
513 |
+
return x_t
|
514 |
+
|
515 |
+
def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False,
|
516 |
+
solver_type='dpm_solver'):
|
517 |
+
"""
|
518 |
+
Singlestep solver DPM-Solver-2 from time `s` to time `t`.
|
519 |
+
Args:
|
520 |
+
x: A pytorch tensor. The initial value at time `s`.
|
521 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
522 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
523 |
+
r1: A `float`. The hyperparameter of the second-order solver.
|
524 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
525 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
526 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
|
527 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
528 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
529 |
+
Returns:
|
530 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
531 |
+
"""
|
532 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
533 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
534 |
+
if r1 is None:
|
535 |
+
r1 = 0.5
|
536 |
+
ns = self.noise_schedule
|
537 |
+
dims = x.dim()
|
538 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
539 |
+
h = lambda_t - lambda_s
|
540 |
+
lambda_s1 = lambda_s + r1 * h
|
541 |
+
s1 = ns.inverse_lambda(lambda_s1)
|
542 |
+
log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(
|
543 |
+
s1), ns.marginal_log_mean_coeff(t)
|
544 |
+
sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t)
|
545 |
+
alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
|
546 |
+
|
547 |
+
if self.predict_x0:
|
548 |
+
phi_11 = torch.expm1(-r1 * h)
|
549 |
+
phi_1 = torch.expm1(-h)
|
550 |
+
|
551 |
+
if model_s is None:
|
552 |
+
model_s = self.model_fn(x, s)
|
553 |
+
x_s1 = (
|
554 |
+
expand_dims(sigma_s1 / sigma_s, dims) * x
|
555 |
+
- expand_dims(alpha_s1 * phi_11, dims) * model_s
|
556 |
+
)
|
557 |
+
model_s1 = self.model_fn(x_s1, s1)
|
558 |
+
if solver_type == 'dpm_solver':
|
559 |
+
x_t = (
|
560 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
561 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
562 |
+
- (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s)
|
563 |
+
)
|
564 |
+
elif solver_type == 'taylor':
|
565 |
+
x_t = (
|
566 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
567 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
568 |
+
+ (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * (
|
569 |
+
model_s1 - model_s)
|
570 |
+
)
|
571 |
+
else:
|
572 |
+
phi_11 = torch.expm1(r1 * h)
|
573 |
+
phi_1 = torch.expm1(h)
|
574 |
+
|
575 |
+
if model_s is None:
|
576 |
+
model_s = self.model_fn(x, s)
|
577 |
+
x_s1 = (
|
578 |
+
expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
|
579 |
+
- expand_dims(sigma_s1 * phi_11, dims) * model_s
|
580 |
+
)
|
581 |
+
model_s1 = self.model_fn(x_s1, s1)
|
582 |
+
if solver_type == 'dpm_solver':
|
583 |
+
x_t = (
|
584 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
585 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
586 |
+
- (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s)
|
587 |
+
)
|
588 |
+
elif solver_type == 'taylor':
|
589 |
+
x_t = (
|
590 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
591 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
592 |
+
- (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s)
|
593 |
+
)
|
594 |
+
if return_intermediate:
|
595 |
+
return x_t, {'model_s': model_s, 'model_s1': model_s1}
|
596 |
+
else:
|
597 |
+
return x_t
|
598 |
+
|
599 |
+
def singlestep_dpm_solver_third_update(self, x, s, t, r1=1. / 3., r2=2. / 3., model_s=None, model_s1=None,
|
600 |
+
return_intermediate=False, solver_type='dpm_solver'):
|
601 |
+
"""
|
602 |
+
Singlestep solver DPM-Solver-3 from time `s` to time `t`.
|
603 |
+
Args:
|
604 |
+
x: A pytorch tensor. The initial value at time `s`.
|
605 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
606 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
607 |
+
r1: A `float`. The hyperparameter of the third-order solver.
|
608 |
+
r2: A `float`. The hyperparameter of the third-order solver.
|
609 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
610 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
611 |
+
model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
|
612 |
+
If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
|
613 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
|
614 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
615 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
616 |
+
Returns:
|
617 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
618 |
+
"""
|
619 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
620 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
621 |
+
if r1 is None:
|
622 |
+
r1 = 1. / 3.
|
623 |
+
if r2 is None:
|
624 |
+
r2 = 2. / 3.
|
625 |
+
ns = self.noise_schedule
|
626 |
+
dims = x.dim()
|
627 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
628 |
+
h = lambda_t - lambda_s
|
629 |
+
lambda_s1 = lambda_s + r1 * h
|
630 |
+
lambda_s2 = lambda_s + r2 * h
|
631 |
+
s1 = ns.inverse_lambda(lambda_s1)
|
632 |
+
s2 = ns.inverse_lambda(lambda_s2)
|
633 |
+
log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff(
|
634 |
+
s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t)
|
635 |
+
sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(
|
636 |
+
s2), ns.marginal_std(t)
|
637 |
+
alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t)
|
638 |
+
|
639 |
+
if self.predict_x0:
|
640 |
+
phi_11 = torch.expm1(-r1 * h)
|
641 |
+
phi_12 = torch.expm1(-r2 * h)
|
642 |
+
phi_1 = torch.expm1(-h)
|
643 |
+
phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.
|
644 |
+
phi_2 = phi_1 / h + 1.
|
645 |
+
phi_3 = phi_2 / h - 0.5
|
646 |
+
|
647 |
+
if model_s is None:
|
648 |
+
model_s = self.model_fn(x, s)
|
649 |
+
if model_s1 is None:
|
650 |
+
x_s1 = (
|
651 |
+
expand_dims(sigma_s1 / sigma_s, dims) * x
|
652 |
+
- expand_dims(alpha_s1 * phi_11, dims) * model_s
|
653 |
+
)
|
654 |
+
model_s1 = self.model_fn(x_s1, s1)
|
655 |
+
x_s2 = (
|
656 |
+
expand_dims(sigma_s2 / sigma_s, dims) * x
|
657 |
+
- expand_dims(alpha_s2 * phi_12, dims) * model_s
|
658 |
+
+ r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s)
|
659 |
+
)
|
660 |
+
model_s2 = self.model_fn(x_s2, s2)
|
661 |
+
if solver_type == 'dpm_solver':
|
662 |
+
x_t = (
|
663 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
664 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
665 |
+
+ (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s)
|
666 |
+
)
|
667 |
+
elif solver_type == 'taylor':
|
668 |
+
D1_0 = (1. / r1) * (model_s1 - model_s)
|
669 |
+
D1_1 = (1. / r2) * (model_s2 - model_s)
|
670 |
+
D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
|
671 |
+
D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
|
672 |
+
x_t = (
|
673 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
674 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
675 |
+
+ expand_dims(alpha_t * phi_2, dims) * D1
|
676 |
+
- expand_dims(alpha_t * phi_3, dims) * D2
|
677 |
+
)
|
678 |
+
else:
|
679 |
+
phi_11 = torch.expm1(r1 * h)
|
680 |
+
phi_12 = torch.expm1(r2 * h)
|
681 |
+
phi_1 = torch.expm1(h)
|
682 |
+
phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.
|
683 |
+
phi_2 = phi_1 / h - 1.
|
684 |
+
phi_3 = phi_2 / h - 0.5
|
685 |
+
|
686 |
+
if model_s is None:
|
687 |
+
model_s = self.model_fn(x, s)
|
688 |
+
if model_s1 is None:
|
689 |
+
x_s1 = (
|
690 |
+
expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
|
691 |
+
- expand_dims(sigma_s1 * phi_11, dims) * model_s
|
692 |
+
)
|
693 |
+
model_s1 = self.model_fn(x_s1, s1)
|
694 |
+
x_s2 = (
|
695 |
+
expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x
|
696 |
+
- expand_dims(sigma_s2 * phi_12, dims) * model_s
|
697 |
+
- r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s)
|
698 |
+
)
|
699 |
+
model_s2 = self.model_fn(x_s2, s2)
|
700 |
+
if solver_type == 'dpm_solver':
|
701 |
+
x_t = (
|
702 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
703 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
704 |
+
- (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s)
|
705 |
+
)
|
706 |
+
elif solver_type == 'taylor':
|
707 |
+
D1_0 = (1. / r1) * (model_s1 - model_s)
|
708 |
+
D1_1 = (1. / r2) * (model_s2 - model_s)
|
709 |
+
D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
|
710 |
+
D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
|
711 |
+
x_t = (
|
712 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
713 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
714 |
+
- expand_dims(sigma_t * phi_2, dims) * D1
|
715 |
+
- expand_dims(sigma_t * phi_3, dims) * D2
|
716 |
+
)
|
717 |
+
|
718 |
+
if return_intermediate:
|
719 |
+
return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2}
|
720 |
+
else:
|
721 |
+
return x_t
|
722 |
+
|
723 |
+
def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"):
|
724 |
+
"""
|
725 |
+
Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
|
726 |
+
Args:
|
727 |
+
x: A pytorch tensor. The initial value at time `s`.
|
728 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
729 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
730 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
731 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
732 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
733 |
+
Returns:
|
734 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
735 |
+
"""
|
736 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
737 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
738 |
+
ns = self.noise_schedule
|
739 |
+
dims = x.dim()
|
740 |
+
model_prev_1, model_prev_0 = model_prev_list
|
741 |
+
t_prev_1, t_prev_0 = t_prev_list
|
742 |
+
lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda(
|
743 |
+
t_prev_0), ns.marginal_lambda(t)
|
744 |
+
log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
|
745 |
+
sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
|
746 |
+
alpha_t = torch.exp(log_alpha_t)
|
747 |
+
|
748 |
+
h_0 = lambda_prev_0 - lambda_prev_1
|
749 |
+
h = lambda_t - lambda_prev_0
|
750 |
+
r0 = h_0 / h
|
751 |
+
D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
|
752 |
+
if self.predict_x0:
|
753 |
+
if solver_type == 'dpm_solver':
|
754 |
+
x_t = (
|
755 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
756 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
757 |
+
- 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0
|
758 |
+
)
|
759 |
+
elif solver_type == 'taylor':
|
760 |
+
x_t = (
|
761 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
762 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
763 |
+
+ expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0
|
764 |
+
)
|
765 |
+
else:
|
766 |
+
if solver_type == 'dpm_solver':
|
767 |
+
x_t = (
|
768 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
769 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
770 |
+
- 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0
|
771 |
+
)
|
772 |
+
elif solver_type == 'taylor':
|
773 |
+
x_t = (
|
774 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
775 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
776 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0
|
777 |
+
)
|
778 |
+
return x_t
|
779 |
+
|
780 |
+
def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'):
|
781 |
+
"""
|
782 |
+
Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
|
783 |
+
Args:
|
784 |
+
x: A pytorch tensor. The initial value at time `s`.
|
785 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
786 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
787 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
788 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
789 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
790 |
+
Returns:
|
791 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
792 |
+
"""
|
793 |
+
ns = self.noise_schedule
|
794 |
+
dims = x.dim()
|
795 |
+
model_prev_2, model_prev_1, model_prev_0 = model_prev_list
|
796 |
+
t_prev_2, t_prev_1, t_prev_0 = t_prev_list
|
797 |
+
lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda(
|
798 |
+
t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t)
|
799 |
+
log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
|
800 |
+
sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
|
801 |
+
alpha_t = torch.exp(log_alpha_t)
|
802 |
+
|
803 |
+
h_1 = lambda_prev_1 - lambda_prev_2
|
804 |
+
h_0 = lambda_prev_0 - lambda_prev_1
|
805 |
+
h = lambda_t - lambda_prev_0
|
806 |
+
r0, r1 = h_0 / h, h_1 / h
|
807 |
+
D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
|
808 |
+
D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2)
|
809 |
+
D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1)
|
810 |
+
D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1)
|
811 |
+
if self.predict_x0:
|
812 |
+
x_t = (
|
813 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
814 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
815 |
+
+ expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1
|
816 |
+
- expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h ** 2 - 0.5), dims) * D2
|
817 |
+
)
|
818 |
+
else:
|
819 |
+
x_t = (
|
820 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
821 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
822 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1
|
823 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h ** 2 - 0.5), dims) * D2
|
824 |
+
)
|
825 |
+
return x_t
|
826 |
+
|
827 |
+
def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None,
|
828 |
+
r2=None):
|
829 |
+
"""
|
830 |
+
Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
|
831 |
+
Args:
|
832 |
+
x: A pytorch tensor. The initial value at time `s`.
|
833 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
834 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
835 |
+
order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
|
836 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
|
837 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
838 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
839 |
+
r1: A `float`. The hyperparameter of the second-order or third-order solver.
|
840 |
+
r2: A `float`. The hyperparameter of the third-order solver.
|
841 |
+
Returns:
|
842 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
843 |
+
"""
|
844 |
+
if order == 1:
|
845 |
+
return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate)
|
846 |
+
elif order == 2:
|
847 |
+
return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate,
|
848 |
+
solver_type=solver_type, r1=r1)
|
849 |
+
elif order == 3:
|
850 |
+
return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate,
|
851 |
+
solver_type=solver_type, r1=r1, r2=r2)
|
852 |
+
else:
|
853 |
+
raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
|
854 |
+
|
855 |
+
def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'):
|
856 |
+
"""
|
857 |
+
Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
|
858 |
+
Args:
|
859 |
+
x: A pytorch tensor. The initial value at time `s`.
|
860 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
861 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
862 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
863 |
+
order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
|
864 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
865 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
866 |
+
Returns:
|
867 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
868 |
+
"""
|
869 |
+
if order == 1:
|
870 |
+
return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1])
|
871 |
+
elif order == 2:
|
872 |
+
return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
|
873 |
+
elif order == 3:
|
874 |
+
return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
|
875 |
+
else:
|
876 |
+
raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
|
877 |
+
|
878 |
+
def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5,
|
879 |
+
solver_type='dpm_solver'):
|
880 |
+
"""
|
881 |
+
The adaptive step size solver based on singlestep DPM-Solver.
|
882 |
+
Args:
|
883 |
+
x: A pytorch tensor. The initial value at time `t_T`.
|
884 |
+
order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
|
885 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
886 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
887 |
+
h_init: A `float`. The initial step size (for logSNR).
|
888 |
+
atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
|
889 |
+
rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
|
890 |
+
theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
|
891 |
+
t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
|
892 |
+
current time and `t_0` is less than `t_err`. The default setting is 1e-5.
|
893 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
894 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
895 |
+
Returns:
|
896 |
+
x_0: A pytorch tensor. The approximated solution at time `t_0`.
|
897 |
+
[1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
|
898 |
+
"""
|
899 |
+
ns = self.noise_schedule
|
900 |
+
s = t_T * torch.ones((x.shape[0],)).to(x)
|
901 |
+
lambda_s = ns.marginal_lambda(s)
|
902 |
+
lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
|
903 |
+
h = h_init * torch.ones_like(s).to(x)
|
904 |
+
x_prev = x
|
905 |
+
nfe = 0
|
906 |
+
if order == 2:
|
907 |
+
r1 = 0.5
|
908 |
+
lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True)
|
909 |
+
higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1,
|
910 |
+
solver_type=solver_type,
|
911 |
+
**kwargs)
|
912 |
+
elif order == 3:
|
913 |
+
r1, r2 = 1. / 3., 2. / 3.
|
914 |
+
lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1,
|
915 |
+
return_intermediate=True,
|
916 |
+
solver_type=solver_type)
|
917 |
+
higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2,
|
918 |
+
solver_type=solver_type,
|
919 |
+
**kwargs)
|
920 |
+
else:
|
921 |
+
raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order))
|
922 |
+
while torch.abs((s - t_0)).mean() > t_err:
|
923 |
+
t = ns.inverse_lambda(lambda_s + h)
|
924 |
+
x_lower, lower_noise_kwargs = lower_update(x, s, t)
|
925 |
+
x_higher = higher_update(x, s, t, **lower_noise_kwargs)
|
926 |
+
delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)))
|
927 |
+
norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True))
|
928 |
+
E = norm_fn((x_higher - x_lower) / delta).max()
|
929 |
+
if torch.all(E <= 1.):
|
930 |
+
x = x_higher
|
931 |
+
s = t
|
932 |
+
x_prev = x_lower
|
933 |
+
lambda_s = ns.marginal_lambda(s)
|
934 |
+
h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s)
|
935 |
+
nfe += order
|
936 |
+
print('adaptive solver nfe', nfe)
|
937 |
+
return x
|
938 |
+
|
939 |
+
def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform',
|
940 |
+
method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver',
|
941 |
+
atol=0.0078, rtol=0.05,
|
942 |
+
):
|
943 |
+
"""
|
944 |
+
Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
|
945 |
+
=====================================================
|
946 |
+
We support the following algorithms for both noise prediction model and data prediction model:
|
947 |
+
- 'singlestep':
|
948 |
+
Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
|
949 |
+
We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
|
950 |
+
The total number of function evaluations (NFE) == `steps`.
|
951 |
+
Given a fixed NFE == `steps`, the sampling procedure is:
|
952 |
+
- If `order` == 1:
|
953 |
+
- Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
|
954 |
+
- If `order` == 2:
|
955 |
+
- Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
|
956 |
+
- If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
|
957 |
+
- If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
|
958 |
+
- If `order` == 3:
|
959 |
+
- Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
|
960 |
+
- If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
|
961 |
+
- If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
|
962 |
+
- If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
|
963 |
+
- 'multistep':
|
964 |
+
Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
|
965 |
+
We initialize the first `order` values by lower order multistep solvers.
|
966 |
+
Given a fixed NFE == `steps`, the sampling procedure is:
|
967 |
+
Denote K = steps.
|
968 |
+
- If `order` == 1:
|
969 |
+
- We use K steps of DPM-Solver-1 (i.e. DDIM).
|
970 |
+
- If `order` == 2:
|
971 |
+
- We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
|
972 |
+
- If `order` == 3:
|
973 |
+
- We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
|
974 |
+
- 'singlestep_fixed':
|
975 |
+
Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
|
976 |
+
We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
|
977 |
+
- 'adaptive':
|
978 |
+
Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
|
979 |
+
We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
|
980 |
+
You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
|
981 |
+
(NFE) and the sample quality.
|
982 |
+
- If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
|
983 |
+
- If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
|
984 |
+
=====================================================
|
985 |
+
Some advices for choosing the algorithm:
|
986 |
+
- For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
|
987 |
+
Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`.
|
988 |
+
e.g.
|
989 |
+
>>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False)
|
990 |
+
>>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
|
991 |
+
skip_type='time_uniform', method='singlestep')
|
992 |
+
- For **guided sampling with large guidance scale** by DPMs:
|
993 |
+
Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`.
|
994 |
+
e.g.
|
995 |
+
>>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True)
|
996 |
+
>>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
|
997 |
+
skip_type='time_uniform', method='multistep')
|
998 |
+
We support three types of `skip_type`:
|
999 |
+
- 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
|
1000 |
+
- 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
|
1001 |
+
- 'time_quadratic': quadratic time for the time steps.
|
1002 |
+
=====================================================
|
1003 |
+
Args:
|
1004 |
+
x: A pytorch tensor. The initial value at time `t_start`
|
1005 |
+
e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
|
1006 |
+
steps: A `int`. The total number of function evaluations (NFE).
|
1007 |
+
t_start: A `float`. The starting time of the sampling.
|
1008 |
+
If `T` is None, we use self.noise_schedule.T (default is 1.0).
|
1009 |
+
t_end: A `float`. The ending time of the sampling.
|
1010 |
+
If `t_end` is None, we use 1. / self.noise_schedule.total_N.
|
1011 |
+
e.g. if total_N == 1000, we have `t_end` == 1e-3.
|
1012 |
+
For discrete-time DPMs:
|
1013 |
+
- We recommend `t_end` == 1. / self.noise_schedule.total_N.
|
1014 |
+
For continuous-time DPMs:
|
1015 |
+
- We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
|
1016 |
+
order: A `int`. The order of DPM-Solver.
|
1017 |
+
skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
|
1018 |
+
method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
|
1019 |
+
denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
|
1020 |
+
Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
|
1021 |
+
This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
|
1022 |
+
score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
|
1023 |
+
for diffusion models sampling by diffusion SDEs for low-resolutional images
|
1024 |
+
(such as CIFAR-10). However, we observed that such trick does not matter for
|
1025 |
+
high-resolutional images. As it needs an additional NFE, we do not recommend
|
1026 |
+
it for high-resolutional images.
|
1027 |
+
lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
|
1028 |
+
Only valid for `method=multistep` and `steps < 15`. We empirically find that
|
1029 |
+
this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
|
1030 |
+
(especially for steps <= 10). So we recommend to set it to be `True`.
|
1031 |
+
solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`.
|
1032 |
+
atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
|
1033 |
+
rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
|
1034 |
+
Returns:
|
1035 |
+
x_end: A pytorch tensor. The approximated solution at time `t_end`.
|
1036 |
+
"""
|
1037 |
+
t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end
|
1038 |
+
t_T = self.noise_schedule.T if t_start is None else t_start
|
1039 |
+
device = x.device
|
1040 |
+
if method == 'adaptive':
|
1041 |
+
with torch.no_grad():
|
1042 |
+
x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol,
|
1043 |
+
solver_type=solver_type)
|
1044 |
+
elif method == 'multistep':
|
1045 |
+
assert steps >= order
|
1046 |
+
timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device)
|
1047 |
+
assert timesteps.shape[0] - 1 == steps
|
1048 |
+
with torch.no_grad():
|
1049 |
+
vec_t = timesteps[0].expand((x.shape[0]))
|
1050 |
+
model_prev_list = [self.model_fn(x, vec_t)]
|
1051 |
+
t_prev_list = [vec_t]
|
1052 |
+
# Init the first `order` values by lower order multistep DPM-Solver.
|
1053 |
+
for init_order in tqdm(range(1, order), desc="DPM init order"):
|
1054 |
+
vec_t = timesteps[init_order].expand(x.shape[0])
|
1055 |
+
x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order,
|
1056 |
+
solver_type=solver_type)
|
1057 |
+
model_prev_list.append(self.model_fn(x, vec_t))
|
1058 |
+
t_prev_list.append(vec_t)
|
1059 |
+
# Compute the remaining values by `order`-th order multistep DPM-Solver.
|
1060 |
+
for step in tqdm(range(order, steps + 1), desc="DPM multistep"):
|
1061 |
+
vec_t = timesteps[step].expand(x.shape[0])
|
1062 |
+
if lower_order_final and steps < 15:
|
1063 |
+
step_order = min(order, steps + 1 - step)
|
1064 |
+
else:
|
1065 |
+
step_order = order
|
1066 |
+
x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, step_order,
|
1067 |
+
solver_type=solver_type)
|
1068 |
+
for i in range(order - 1):
|
1069 |
+
t_prev_list[i] = t_prev_list[i + 1]
|
1070 |
+
model_prev_list[i] = model_prev_list[i + 1]
|
1071 |
+
t_prev_list[-1] = vec_t
|
1072 |
+
# We do not need to evaluate the final model value.
|
1073 |
+
if step < steps:
|
1074 |
+
model_prev_list[-1] = self.model_fn(x, vec_t)
|
1075 |
+
elif method in ['singlestep', 'singlestep_fixed']:
|
1076 |
+
if method == 'singlestep':
|
1077 |
+
timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order,
|
1078 |
+
skip_type=skip_type,
|
1079 |
+
t_T=t_T, t_0=t_0,
|
1080 |
+
device=device)
|
1081 |
+
elif method == 'singlestep_fixed':
|
1082 |
+
K = steps // order
|
1083 |
+
orders = [order, ] * K
|
1084 |
+
timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device)
|
1085 |
+
for i, order in enumerate(orders):
|
1086 |
+
t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1]
|
1087 |
+
timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(),
|
1088 |
+
N=order, device=device)
|
1089 |
+
lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
|
1090 |
+
vec_s, vec_t = t_T_inner.tile(x.shape[0]), t_0_inner.tile(x.shape[0])
|
1091 |
+
h = lambda_inner[-1] - lambda_inner[0]
|
1092 |
+
r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
|
1093 |
+
r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
|
1094 |
+
x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2)
|
1095 |
+
if denoise_to_zero:
|
1096 |
+
x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0)
|
1097 |
+
return x
|
1098 |
+
|
1099 |
+
|
1100 |
+
#############################################################
|
1101 |
+
# other utility functions
|
1102 |
+
#############################################################
|
1103 |
+
|
1104 |
+
def interpolate_fn(x, xp, yp):
|
1105 |
+
"""
|
1106 |
+
A piecewise linear function y = f(x), using xp and yp as keypoints.
|
1107 |
+
We implement f(x) in a differentiable way (i.e. applicable for autograd).
|
1108 |
+
The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
|
1109 |
+
Args:
|
1110 |
+
x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
|
1111 |
+
xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
|
1112 |
+
yp: PyTorch tensor with shape [C, K].
|
1113 |
+
Returns:
|
1114 |
+
The function values f(x), with shape [N, C].
|
1115 |
+
"""
|
1116 |
+
N, K = x.shape[0], xp.shape[1]
|
1117 |
+
all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
|
1118 |
+
sorted_all_x, x_indices = torch.sort(all_x, dim=2)
|
1119 |
+
x_idx = torch.argmin(x_indices, dim=2)
|
1120 |
+
cand_start_idx = x_idx - 1
|
1121 |
+
start_idx = torch.where(
|
1122 |
+
torch.eq(x_idx, 0),
|
1123 |
+
torch.tensor(1, device=x.device),
|
1124 |
+
torch.where(
|
1125 |
+
torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
|
1126 |
+
),
|
1127 |
+
)
|
1128 |
+
end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1)
|
1129 |
+
start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
|
1130 |
+
end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
|
1131 |
+
start_idx2 = torch.where(
|
1132 |
+
torch.eq(x_idx, 0),
|
1133 |
+
torch.tensor(0, device=x.device),
|
1134 |
+
torch.where(
|
1135 |
+
torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
|
1136 |
+
),
|
1137 |
+
)
|
1138 |
+
y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
|
1139 |
+
start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2)
|
1140 |
+
end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2)
|
1141 |
+
cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
|
1142 |
+
return cand
|
1143 |
+
|
1144 |
+
|
1145 |
+
def expand_dims(v, dims):
|
1146 |
+
"""
|
1147 |
+
Expand the tensor `v` to the dim `dims`.
|
1148 |
+
Args:
|
1149 |
+
`v`: a PyTorch tensor with shape [N].
|
1150 |
+
`dim`: a `int`.
|
1151 |
+
Returns:
|
1152 |
+
a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
|
1153 |
+
"""
|
1154 |
+
return v[(...,) + (None,) * (dims - 1)]
|
ControlNet/ldm/models/diffusion/dpm_solver/sampler.py
ADDED
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""SAMPLING ONLY."""
|
2 |
+
import torch
|
3 |
+
|
4 |
+
from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver
|
5 |
+
|
6 |
+
|
7 |
+
MODEL_TYPES = {
|
8 |
+
"eps": "noise",
|
9 |
+
"v": "v"
|
10 |
+
}
|
11 |
+
|
12 |
+
|
13 |
+
class DPMSolverSampler(object):
|
14 |
+
def __init__(self, model, **kwargs):
|
15 |
+
super().__init__()
|
16 |
+
self.model = model
|
17 |
+
to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device)
|
18 |
+
self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod))
|
19 |
+
|
20 |
+
def register_buffer(self, name, attr):
|
21 |
+
if type(attr) == torch.Tensor:
|
22 |
+
if attr.device != torch.device("cuda"):
|
23 |
+
attr = attr.to(torch.device("cuda"))
|
24 |
+
setattr(self, name, attr)
|
25 |
+
|
26 |
+
@torch.no_grad()
|
27 |
+
def sample(self,
|
28 |
+
S,
|
29 |
+
batch_size,
|
30 |
+
shape,
|
31 |
+
conditioning=None,
|
32 |
+
callback=None,
|
33 |
+
normals_sequence=None,
|
34 |
+
img_callback=None,
|
35 |
+
quantize_x0=False,
|
36 |
+
eta=0.,
|
37 |
+
mask=None,
|
38 |
+
x0=None,
|
39 |
+
temperature=1.,
|
40 |
+
noise_dropout=0.,
|
41 |
+
score_corrector=None,
|
42 |
+
corrector_kwargs=None,
|
43 |
+
verbose=True,
|
44 |
+
x_T=None,
|
45 |
+
log_every_t=100,
|
46 |
+
unconditional_guidance_scale=1.,
|
47 |
+
unconditional_conditioning=None,
|
48 |
+
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
|
49 |
+
**kwargs
|
50 |
+
):
|
51 |
+
if conditioning is not None:
|
52 |
+
if isinstance(conditioning, dict):
|
53 |
+
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
|
54 |
+
if cbs != batch_size:
|
55 |
+
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
|
56 |
+
else:
|
57 |
+
if conditioning.shape[0] != batch_size:
|
58 |
+
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
|
59 |
+
|
60 |
+
# sampling
|
61 |
+
C, H, W = shape
|
62 |
+
size = (batch_size, C, H, W)
|
63 |
+
|
64 |
+
print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}')
|
65 |
+
|
66 |
+
device = self.model.betas.device
|
67 |
+
if x_T is None:
|
68 |
+
img = torch.randn(size, device=device)
|
69 |
+
else:
|
70 |
+
img = x_T
|
71 |
+
|
72 |
+
ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod)
|
73 |
+
|
74 |
+
model_fn = model_wrapper(
|
75 |
+
lambda x, t, c: self.model.apply_model(x, t, c),
|
76 |
+
ns,
|
77 |
+
model_type=MODEL_TYPES[self.model.parameterization],
|
78 |
+
guidance_type="classifier-free",
|
79 |
+
condition=conditioning,
|
80 |
+
unconditional_condition=unconditional_conditioning,
|
81 |
+
guidance_scale=unconditional_guidance_scale,
|
82 |
+
)
|
83 |
+
|
84 |
+
dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False)
|
85 |
+
x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, lower_order_final=True)
|
86 |
+
|
87 |
+
return x.to(device), None
|
ControlNet/ldm/models/diffusion/plms.py
ADDED
@@ -0,0 +1,244 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""SAMPLING ONLY."""
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import numpy as np
|
5 |
+
from tqdm import tqdm
|
6 |
+
from functools import partial
|
7 |
+
|
8 |
+
from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
|
9 |
+
from ldm.models.diffusion.sampling_util import norm_thresholding
|
10 |
+
|
11 |
+
|
12 |
+
class PLMSSampler(object):
|
13 |
+
def __init__(self, model, schedule="linear", **kwargs):
|
14 |
+
super().__init__()
|
15 |
+
self.model = model
|
16 |
+
self.ddpm_num_timesteps = model.num_timesteps
|
17 |
+
self.schedule = schedule
|
18 |
+
|
19 |
+
def register_buffer(self, name, attr):
|
20 |
+
if type(attr) == torch.Tensor:
|
21 |
+
if attr.device != torch.device("cuda"):
|
22 |
+
attr = attr.to(torch.device("cuda"))
|
23 |
+
setattr(self, name, attr)
|
24 |
+
|
25 |
+
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
|
26 |
+
if ddim_eta != 0:
|
27 |
+
raise ValueError('ddim_eta must be 0 for PLMS')
|
28 |
+
self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
|
29 |
+
num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
|
30 |
+
alphas_cumprod = self.model.alphas_cumprod
|
31 |
+
assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
|
32 |
+
to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
|
33 |
+
|
34 |
+
self.register_buffer('betas', to_torch(self.model.betas))
|
35 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
36 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
|
37 |
+
|
38 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
39 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
|
40 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
|
41 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
|
42 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
|
43 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
|
44 |
+
|
45 |
+
# ddim sampling parameters
|
46 |
+
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
|
47 |
+
ddim_timesteps=self.ddim_timesteps,
|
48 |
+
eta=ddim_eta,verbose=verbose)
|
49 |
+
self.register_buffer('ddim_sigmas', ddim_sigmas)
|
50 |
+
self.register_buffer('ddim_alphas', ddim_alphas)
|
51 |
+
self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
|
52 |
+
self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
|
53 |
+
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
|
54 |
+
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
|
55 |
+
1 - self.alphas_cumprod / self.alphas_cumprod_prev))
|
56 |
+
self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
|
57 |
+
|
58 |
+
@torch.no_grad()
|
59 |
+
def sample(self,
|
60 |
+
S,
|
61 |
+
batch_size,
|
62 |
+
shape,
|
63 |
+
conditioning=None,
|
64 |
+
callback=None,
|
65 |
+
normals_sequence=None,
|
66 |
+
img_callback=None,
|
67 |
+
quantize_x0=False,
|
68 |
+
eta=0.,
|
69 |
+
mask=None,
|
70 |
+
x0=None,
|
71 |
+
temperature=1.,
|
72 |
+
noise_dropout=0.,
|
73 |
+
score_corrector=None,
|
74 |
+
corrector_kwargs=None,
|
75 |
+
verbose=True,
|
76 |
+
x_T=None,
|
77 |
+
log_every_t=100,
|
78 |
+
unconditional_guidance_scale=1.,
|
79 |
+
unconditional_conditioning=None,
|
80 |
+
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
|
81 |
+
dynamic_threshold=None,
|
82 |
+
**kwargs
|
83 |
+
):
|
84 |
+
if conditioning is not None:
|
85 |
+
if isinstance(conditioning, dict):
|
86 |
+
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
|
87 |
+
if cbs != batch_size:
|
88 |
+
print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
|
89 |
+
else:
|
90 |
+
if conditioning.shape[0] != batch_size:
|
91 |
+
print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
|
92 |
+
|
93 |
+
self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
|
94 |
+
# sampling
|
95 |
+
C, H, W = shape
|
96 |
+
size = (batch_size, C, H, W)
|
97 |
+
print(f'Data shape for PLMS sampling is {size}')
|
98 |
+
|
99 |
+
samples, intermediates = self.plms_sampling(conditioning, size,
|
100 |
+
callback=callback,
|
101 |
+
img_callback=img_callback,
|
102 |
+
quantize_denoised=quantize_x0,
|
103 |
+
mask=mask, x0=x0,
|
104 |
+
ddim_use_original_steps=False,
|
105 |
+
noise_dropout=noise_dropout,
|
106 |
+
temperature=temperature,
|
107 |
+
score_corrector=score_corrector,
|
108 |
+
corrector_kwargs=corrector_kwargs,
|
109 |
+
x_T=x_T,
|
110 |
+
log_every_t=log_every_t,
|
111 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
112 |
+
unconditional_conditioning=unconditional_conditioning,
|
113 |
+
dynamic_threshold=dynamic_threshold,
|
114 |
+
)
|
115 |
+
return samples, intermediates
|
116 |
+
|
117 |
+
@torch.no_grad()
|
118 |
+
def plms_sampling(self, cond, shape,
|
119 |
+
x_T=None, ddim_use_original_steps=False,
|
120 |
+
callback=None, timesteps=None, quantize_denoised=False,
|
121 |
+
mask=None, x0=None, img_callback=None, log_every_t=100,
|
122 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
|
123 |
+
unconditional_guidance_scale=1., unconditional_conditioning=None,
|
124 |
+
dynamic_threshold=None):
|
125 |
+
device = self.model.betas.device
|
126 |
+
b = shape[0]
|
127 |
+
if x_T is None:
|
128 |
+
img = torch.randn(shape, device=device)
|
129 |
+
else:
|
130 |
+
img = x_T
|
131 |
+
|
132 |
+
if timesteps is None:
|
133 |
+
timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
|
134 |
+
elif timesteps is not None and not ddim_use_original_steps:
|
135 |
+
subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
|
136 |
+
timesteps = self.ddim_timesteps[:subset_end]
|
137 |
+
|
138 |
+
intermediates = {'x_inter': [img], 'pred_x0': [img]}
|
139 |
+
time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
|
140 |
+
total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
|
141 |
+
print(f"Running PLMS Sampling with {total_steps} timesteps")
|
142 |
+
|
143 |
+
iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
|
144 |
+
old_eps = []
|
145 |
+
|
146 |
+
for i, step in enumerate(iterator):
|
147 |
+
index = total_steps - i - 1
|
148 |
+
ts = torch.full((b,), step, device=device, dtype=torch.long)
|
149 |
+
ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
|
150 |
+
|
151 |
+
if mask is not None:
|
152 |
+
assert x0 is not None
|
153 |
+
img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?
|
154 |
+
img = img_orig * mask + (1. - mask) * img
|
155 |
+
|
156 |
+
outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
|
157 |
+
quantize_denoised=quantize_denoised, temperature=temperature,
|
158 |
+
noise_dropout=noise_dropout, score_corrector=score_corrector,
|
159 |
+
corrector_kwargs=corrector_kwargs,
|
160 |
+
unconditional_guidance_scale=unconditional_guidance_scale,
|
161 |
+
unconditional_conditioning=unconditional_conditioning,
|
162 |
+
old_eps=old_eps, t_next=ts_next,
|
163 |
+
dynamic_threshold=dynamic_threshold)
|
164 |
+
img, pred_x0, e_t = outs
|
165 |
+
old_eps.append(e_t)
|
166 |
+
if len(old_eps) >= 4:
|
167 |
+
old_eps.pop(0)
|
168 |
+
if callback: callback(i)
|
169 |
+
if img_callback: img_callback(pred_x0, i)
|
170 |
+
|
171 |
+
if index % log_every_t == 0 or index == total_steps - 1:
|
172 |
+
intermediates['x_inter'].append(img)
|
173 |
+
intermediates['pred_x0'].append(pred_x0)
|
174 |
+
|
175 |
+
return img, intermediates
|
176 |
+
|
177 |
+
@torch.no_grad()
|
178 |
+
def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
|
179 |
+
temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
|
180 |
+
unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None,
|
181 |
+
dynamic_threshold=None):
|
182 |
+
b, *_, device = *x.shape, x.device
|
183 |
+
|
184 |
+
def get_model_output(x, t):
|
185 |
+
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
|
186 |
+
e_t = self.model.apply_model(x, t, c)
|
187 |
+
else:
|
188 |
+
x_in = torch.cat([x] * 2)
|
189 |
+
t_in = torch.cat([t] * 2)
|
190 |
+
c_in = torch.cat([unconditional_conditioning, c])
|
191 |
+
e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
|
192 |
+
e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
|
193 |
+
|
194 |
+
if score_corrector is not None:
|
195 |
+
assert self.model.parameterization == "eps"
|
196 |
+
e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
|
197 |
+
|
198 |
+
return e_t
|
199 |
+
|
200 |
+
alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
|
201 |
+
alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
|
202 |
+
sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
|
203 |
+
sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
|
204 |
+
|
205 |
+
def get_x_prev_and_pred_x0(e_t, index):
|
206 |
+
# select parameters corresponding to the currently considered timestep
|
207 |
+
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
|
208 |
+
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
|
209 |
+
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
|
210 |
+
sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
|
211 |
+
|
212 |
+
# current prediction for x_0
|
213 |
+
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
|
214 |
+
if quantize_denoised:
|
215 |
+
pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
|
216 |
+
if dynamic_threshold is not None:
|
217 |
+
pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)
|
218 |
+
# direction pointing to x_t
|
219 |
+
dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
|
220 |
+
noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
|
221 |
+
if noise_dropout > 0.:
|
222 |
+
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
223 |
+
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
|
224 |
+
return x_prev, pred_x0
|
225 |
+
|
226 |
+
e_t = get_model_output(x, t)
|
227 |
+
if len(old_eps) == 0:
|
228 |
+
# Pseudo Improved Euler (2nd order)
|
229 |
+
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
|
230 |
+
e_t_next = get_model_output(x_prev, t_next)
|
231 |
+
e_t_prime = (e_t + e_t_next) / 2
|
232 |
+
elif len(old_eps) == 1:
|
233 |
+
# 2nd order Pseudo Linear Multistep (Adams-Bashforth)
|
234 |
+
e_t_prime = (3 * e_t - old_eps[-1]) / 2
|
235 |
+
elif len(old_eps) == 2:
|
236 |
+
# 3nd order Pseudo Linear Multistep (Adams-Bashforth)
|
237 |
+
e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
|
238 |
+
elif len(old_eps) >= 3:
|
239 |
+
# 4nd order Pseudo Linear Multistep (Adams-Bashforth)
|
240 |
+
e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
|
241 |
+
|
242 |
+
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
|
243 |
+
|
244 |
+
return x_prev, pred_x0, e_t
|
ControlNet/ldm/models/diffusion/sampling_util.py
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import numpy as np
|
3 |
+
|
4 |
+
|
5 |
+
def append_dims(x, target_dims):
|
6 |
+
"""Appends dimensions to the end of a tensor until it has target_dims dimensions.
|
7 |
+
From https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/utils.py"""
|
8 |
+
dims_to_append = target_dims - x.ndim
|
9 |
+
if dims_to_append < 0:
|
10 |
+
raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less')
|
11 |
+
return x[(...,) + (None,) * dims_to_append]
|
12 |
+
|
13 |
+
|
14 |
+
def norm_thresholding(x0, value):
|
15 |
+
s = append_dims(x0.pow(2).flatten(1).mean(1).sqrt().clamp(min=value), x0.ndim)
|
16 |
+
return x0 * (value / s)
|
17 |
+
|
18 |
+
|
19 |
+
def spatial_norm_thresholding(x0, value):
|
20 |
+
# b c h w
|
21 |
+
s = x0.pow(2).mean(1, keepdim=True).sqrt().clamp(min=value)
|
22 |
+
return x0 * (value / s)
|