Spaces:
Running
on
Zero
Running
on
Zero
File size: 13,017 Bytes
0a63786 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 |
import torch
from typing import List, Union, Tuple
from tqdm import tqdm
from .inference import Inference
class InferenceDAMO(Inference):
@torch.no_grad()
def __call__(
self,
latent: torch.Tensor,
context: torch.Tensor,
uncond_context: torch.Tensor=None,
start_time: int = 0,
null_embedding: List[torch.Tensor]=None,
):
all_latent = []
all_pred = [] # x0_hat
do_classifier_free_guidance = self.guidance_scale > 1 and ((uncond_context is not None) or (null_embedding is not None))
for i, t in enumerate(tqdm(self.scheduler.timesteps[start_time:])):
t = int(t)
if do_classifier_free_guidance:
latent_input = torch.cat([latent, latent], dim=0)
if null_embedding is not None:
context_input = torch.cat([null_embedding[i], context], dim=0)
else:
context_input = torch.cat([uncond_context, context], dim=0)
else:
latent_input = latent
context_input = context
noise_pred = self.unet(
latent_input,
torch.full((len(latent_input),), t, device=latent_input.device, dtype=torch.long),
context_input,
)
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2, dim=0)
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
pred_samples = self.scheduler.step(noise_pred, t, latent)
latent = pred_samples.prev_sample
pred = pred_samples.pred_original_sample
all_latent.append(latent.detach())
all_pred.append(pred.detach())
return {
'latent': latent,
'all_latent': all_latent,
'all_pred': all_pred
}
class InferenceDAMO_PTP(Inference):
def infer_old_context(self, latent, context, t, uncond_context=None):
do_classifier_free_guidance = self.guidance_scale > 1 and (uncond_context is not None)
if do_classifier_free_guidance:
latent_input = torch.cat([latent, latent], dim=0)
context_input = torch.cat([uncond_context, context], dim=0)
else:
latent_input = latent
context_input = context
noise_pred = self.unet(
latent_input,
torch.full((len(latent_input),), t, device=latent_input.device, dtype=torch.long),
context_input,
)
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2, dim=0)
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
pred_samples = self.scheduler.step(noise_pred, t, latent)
latent = pred_samples.prev_sample
pred = pred_samples.pred_original_sample
return latent, pred
def infer_new_context(self, latent, context, t, uncond_context=None):
do_classifier_free_guidance = self.guidance_scale > 1 and (uncond_context is not None)
if do_classifier_free_guidance:
latent_input = torch.cat([latent, latent], dim=0)
if isinstance(context, (list, tuple)):
context_input = (
torch.cat([uncond_context, context[0]], dim=0),
torch.cat([uncond_context, context[1]], dim=0),
)
else:
context_input = torch.cat([uncond_context, context], dim=0)
else:
latent_input = latent
context_input = context
noise_pred = self.unet(
latent_input,
torch.full((len(latent_input),), t, device=latent_input.device, dtype=torch.long),
context_input,
)
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2, dim=0)
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
pred_samples = self.scheduler.step(noise_pred, t, latent)
latent = pred_samples.prev_sample
pred = pred_samples.pred_original_sample
return latent, pred
@torch.no_grad()
def __call__(
self,
latent: torch.Tensor,
context: torch.Tensor, # used when > ca_end_time
old_context: torch.Tensor=None, # used when < sa_end_time
old_to_new_context: Union[Tuple, List]=None, # used when sa_end_time < t < ca_end_time
uncond_context: torch.Tensor=None,
sa_end_time: float=0.3,
ca_end_time: float=0.8,
start_time: int = 0,
):
assert sa_end_time < ca_end_time, f"sa_end_time must be less than ca_end_time, got {sa_end_time} and {ca_end_time} respectively"
all_latent = []
all_pred = []
all_latent_old = []
all_pred_old = []
old_latent = latent.clone()
new_latent = latent.clone()
for i, t in enumerate(tqdm(self.scheduler.timesteps[start_time:])):
t = int(t)
old_latent_next_t, pred_old = self.infer_old_context(old_latent, old_context, t, uncond_context)
if i < sa_end_time * self.num_ddim_steps:
new_latent_next_t, pred_new = old_latent_next_t, pred_old
elif sa_end_time * self.num_ddim_steps <= i < ca_end_time * self.num_ddim_steps:
new_latent_next_t, pred_new = self.infer_new_context(
new_latent, old_to_new_context, t, uncond_context
)
else:
new_latent_next_t, pred_new = self.infer_new_context(
new_latent, context, t, uncond_context
)
old_latent = old_latent_next_t
new_latent = new_latent_next_t
all_latent.append(new_latent_next_t.detach())
all_pred.append(pred_new.detach())
all_latent_old.append(old_latent_next_t.detach())
all_pred_old.append(pred_old.detach())
return {
'latent': new_latent,
'latent_old': old_latent,
'all_latent': all_latent,
'all_pred': all_pred,
'all_latent_old': all_latent_old,
'all_pred_old': all_pred_old,
}
class InferenceDAMO_PTP_v2(Inference):
def set_ptp_in_xattn_layers(self, prompt_to_prompt: bool, num_frames=1):
for m in self.unet.modules():
if m.__class__.__name__ == 'CrossAttention':
m.ptp_sa_replace = prompt_to_prompt
m.num_frames = num_frames
def infer_both_with_sa_replace(self, old_latent, new_latent, old_context, new_context, t, uncond_context=None):
do_classifier_free_guidance = self.guidance_scale > 1 and (uncond_context is not None)
if do_classifier_free_guidance:
latent_input = torch.cat([old_latent, new_latent, old_latent, new_latent], dim=0)
context_input = torch.cat([uncond_context, uncond_context, old_context, new_context], dim=0)
else:
latent_input = torch.cat([old_latent, new_latent], dim=0)
context_input = torch.cat([old_context, new_context], dim=0)
noise_pred = self.unet(
latent_input,
torch.full((len(latent_input),), t, device=latent_input.device, dtype=torch.long),
context_input,
)
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2, dim=0)
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
noise_pred_old, noise_pred_new = noise_pred.chunk(2, dim=0)
pred_samples_old = self.scheduler.step(noise_pred_old, t, old_latent)
pred_samples_new = self.scheduler.step(noise_pred_new, t, new_latent)
old_latent = pred_samples_old.prev_sample
new_latent = pred_samples_new.prev_sample
old_pred = pred_samples_old.pred_original_sample
new_pred = pred_samples_new.pred_original_sample
return old_latent, new_latent, old_pred, new_pred
def infer_old_context(self, latent, context, t, uncond_context=None):
do_classifier_free_guidance = self.guidance_scale > 1 and (uncond_context is not None)
if do_classifier_free_guidance:
latent_input = torch.cat([latent, latent], dim=0)
context_input = torch.cat([uncond_context, context], dim=0)
else:
latent_input = latent
context_input = context
noise_pred = self.unet(
latent_input,
torch.full((len(latent_input),), t, device=latent_input.device, dtype=torch.long),
context_input,
)
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2, dim=0)
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
pred_samples = self.scheduler.step(noise_pred, t, latent)
latent = pred_samples.prev_sample
pred = pred_samples.pred_original_sample
return latent, pred
def infer_new_context(self, latent, context, t, uncond_context=None):
do_classifier_free_guidance = self.guidance_scale > 1 and (uncond_context is not None)
if do_classifier_free_guidance:
latent_input = torch.cat([latent, latent], dim=0)
if isinstance(context, (list, tuple)):
context_input = (
torch.cat([uncond_context, context[0]], dim=0),
torch.cat([uncond_context, context[1]], dim=0),
)
else:
context_input = torch.cat([uncond_context, context], dim=0)
else:
latent_input = latent
context_input = context
noise_pred = self.unet(
latent_input,
torch.full((len(latent_input),), t, device=latent_input.device, dtype=torch.long),
context_input,
)
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2, dim=0)
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
pred_samples = self.scheduler.step(noise_pred, t, latent)
latent = pred_samples.prev_sample
pred = pred_samples.pred_original_sample
return latent, pred
@torch.no_grad()
def __call__(
self,
latent: torch.Tensor,
context: torch.Tensor, # used when > ca_end_time
old_context: torch.Tensor=None, # used when < sa_end_time
old_to_new_context: Union[Tuple, List]=None, # used when sa_end_time < t < ca_end_time
uncond_context: torch.Tensor=None,
sa_end_time: float=0.3,
ca_end_time: float=0.8,
start_time: int = 0,
):
assert sa_end_time < ca_end_time, f"sa_end_time must be less than ca_end_time, got {sa_end_time} and {ca_end_time} respectively"
all_latent = []
all_pred = []
all_latent_old = []
all_pred_old = []
old_latent = latent.clone()
new_latent = latent.clone()
for i, t in enumerate(tqdm(self.scheduler.timesteps[start_time:])):
t = int(t)
if i < sa_end_time * self.num_ddim_steps:
self.set_ptp_in_xattn_layers(True, num_frames=latent.shape[2])
old_latent_next_t, new_latent_next_t, pred_old, pred_new = self.infer_both_with_sa_replace(
old_latent, new_latent, old_context, context, t, uncond_context
)
elif sa_end_time * self.num_ddim_steps <= i < ca_end_time * self.num_ddim_steps:
self.set_ptp_in_xattn_layers(False)
old_latent_next_t, pred_old = self.infer_old_context(old_latent, old_context, t, uncond_context)
new_latent_next_t, pred_new = self.infer_new_context(
new_latent, old_to_new_context, t, uncond_context
)
else:
self.set_ptp_in_xattn_layers(False)
old_latent_next_t, pred_old = self.infer_old_context(old_latent, old_context, t, uncond_context)
new_latent_next_t, pred_new = self.infer_new_context(
new_latent, context, t, uncond_context
)
old_latent = old_latent_next_t
new_latent = new_latent_next_t
all_latent.append(new_latent_next_t.detach())
all_pred.append(pred_new.detach())
all_latent_old.append(old_latent_next_t.detach())
all_pred_old.append(pred_old.detach())
return {
'latent': new_latent,
'latent_old': old_latent,
'all_latent': all_latent,
'all_pred': all_pred,
'all_latent_old': all_latent_old,
'all_pred_old': all_pred_old,
} |