Spaces:
Runtime error
Runtime error
modify the rendertext_tool
Browse files- app.py +3 -3
- cldm/cldm.py +47 -342
- scripts/rendertext_tool.py +33 -3
app.py
CHANGED
|
@@ -93,7 +93,7 @@ def load_ckpt(model_ckpt = "LAION-Glyph-10M-Epoch-5"):
|
|
| 93 |
elif model_ckpt == "TextCaps-5K-Epoch-40":
|
| 94 |
model = load_model_ckpt(model, "textcaps5K_epoch_40_model_wo_ema.ckpt")
|
| 95 |
|
| 96 |
-
render_tool = Render_Text(model)
|
| 97 |
output_str = f"already change the model checkpoint to {model_ckpt}"
|
| 98 |
print(output_str)
|
| 99 |
if torch.cuda.is_available():
|
|
@@ -104,14 +104,14 @@ def load_ckpt(model_ckpt = "LAION-Glyph-10M-Epoch-5"):
|
|
| 104 |
allow_run_generation = False
|
| 105 |
return output_str, None, allow_run_generation
|
| 106 |
|
| 107 |
-
|
| 108 |
cfg = OmegaConf.load("config.yaml")
|
| 109 |
model = load_model_from_config(cfg, "laion10M_epoch_6_model_wo_ema.ckpt", verbose=True)
|
| 110 |
# model = load_model_from_config(cfg, "model_wo_ema.ckpt", verbose=True)
|
| 111 |
# model = load_model_from_config(cfg, "model_states.pt", verbose=True)
|
| 112 |
# model = load_model_from_config(cfg, "model.ckpt", verbose=True)
|
| 113 |
# ddim_sampler = DDIMSampler(model)
|
| 114 |
-
render_tool = Render_Text(model)
|
| 115 |
|
| 116 |
|
| 117 |
description = """
|
|
|
|
| 93 |
elif model_ckpt == "TextCaps-5K-Epoch-40":
|
| 94 |
model = load_model_ckpt(model, "textcaps5K_epoch_40_model_wo_ema.ckpt")
|
| 95 |
|
| 96 |
+
render_tool = Render_Text(model, save_memory = SAVE_MEMORY)
|
| 97 |
output_str = f"already change the model checkpoint to {model_ckpt}"
|
| 98 |
print(output_str)
|
| 99 |
if torch.cuda.is_available():
|
|
|
|
| 104 |
allow_run_generation = False
|
| 105 |
return output_str, None, allow_run_generation
|
| 106 |
|
| 107 |
+
SAVE_MEMORY = True
|
| 108 |
cfg = OmegaConf.load("config.yaml")
|
| 109 |
model = load_model_from_config(cfg, "laion10M_epoch_6_model_wo_ema.ckpt", verbose=True)
|
| 110 |
# model = load_model_from_config(cfg, "model_wo_ema.ckpt", verbose=True)
|
| 111 |
# model = load_model_from_config(cfg, "model_states.pt", verbose=True)
|
| 112 |
# model = load_model_from_config(cfg, "model.ckpt", verbose=True)
|
| 113 |
# ddim_sampler = DDIMSampler(model)
|
| 114 |
+
render_tool = Render_Text(model, save_memory = SAVE_MEMORY)
|
| 115 |
|
| 116 |
|
| 117 |
description = """
|
cldm/cldm.py
CHANGED
|
@@ -28,7 +28,7 @@ def disabled_train(self, mode=True):
|
|
| 28 |
return self
|
| 29 |
|
| 30 |
class ControlledUnetModel(UNetModel):
|
| 31 |
-
def forward(self, x, timesteps=None, context=None, control=None, only_mid_control=False,
|
| 32 |
hs = []
|
| 33 |
with torch.no_grad():
|
| 34 |
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
|
|
@@ -47,7 +47,7 @@ class ControlledUnetModel(UNetModel):
|
|
| 47 |
h = torch.cat([h, hs.pop()], dim=1)
|
| 48 |
else:
|
| 49 |
h = torch.cat([h, hs.pop() + control.pop()], dim=1)
|
| 50 |
-
h = module(h, emb, context
|
| 51 |
|
| 52 |
h = h.type(x.dtype)
|
| 53 |
return self.out(h)
|
|
@@ -317,16 +317,12 @@ class ControlLDM(LatentDiffusion):
|
|
| 317 |
|
| 318 |
def __init__(self,
|
| 319 |
control_stage_config,
|
| 320 |
-
control_key, only_mid_control,
|
| 321 |
-
sd_locked = True, concat_textemb = False,
|
| 322 |
-
trans_textemb=False, trans_textemb_config = None,
|
| 323 |
learnable_conscale = False, guess_mode=False,
|
| 324 |
-
sep_lr = False, decoder_lr = 1.0**-4,
|
| 325 |
-
|
| 326 |
-
glyph_control_key = "centered_hint", sep_cond_txt = False, exchange_cond_txt = False,
|
| 327 |
-
max_step = None, multiple_optimizers = False, deepspeed = False, trans_glyph_lr = 1.0**-5,
|
| 328 |
*args, **kwargs
|
| 329 |
-
):
|
| 330 |
use_ema = kwargs.pop("use_ema", False)
|
| 331 |
ckpt_path = kwargs.pop("ckpt_path", None)
|
| 332 |
reset_ema = kwargs.pop("reset_ema", False)
|
|
@@ -336,90 +332,53 @@ class ControlLDM(LatentDiffusion):
|
|
| 336 |
ignore_keys = kwargs.pop("ignore_keys", [])
|
| 337 |
|
| 338 |
super().__init__(*args, use_ema=False, **kwargs)
|
|
|
|
|
|
|
| 339 |
self.control_model = instantiate_from_config(control_stage_config)
|
| 340 |
self.control_key = control_key
|
| 341 |
self.only_mid_control = only_mid_control
|
|
|
|
| 342 |
self.learnable_conscale = learnable_conscale
|
| 343 |
conscale_init = [1.0] * 13 if not guess_mode else [(0.825 ** float(12 - i)) for i in range(13)]
|
| 344 |
if learnable_conscale:
|
| 345 |
# self.control_scales = nn.Parameter(torch.ones(13), requires_grad=True)
|
| 346 |
self.control_scales = nn.Parameter(torch.Tensor(conscale_init), requires_grad=True)
|
| 347 |
-
else:
|
| 348 |
self.control_scales = conscale_init #[1.0] * 13
|
|
|
|
|
|
|
|
|
|
| 349 |
self.sd_locked = sd_locked
|
| 350 |
-
self.concat_textemb = concat_textemb
|
| 351 |
-
# update
|
| 352 |
-
self.trans_textemb = False
|
| 353 |
-
if trans_textemb and trans_textemb_config is not None:
|
| 354 |
-
self.trans_textemb = True
|
| 355 |
-
self.instantiate_trans_textemb_model(trans_textemb_config)
|
| 356 |
-
# self.sep_cap_for_2b = sep_cap_for_2b
|
| 357 |
-
|
| 358 |
self.sep_lr = sep_lr
|
| 359 |
self.decoder_lr = decoder_lr
|
|
|
|
|
|
|
| 360 |
self.sep_cond_txt = sep_cond_txt
|
|
|
|
| 361 |
self.exchange_cond_txt = exchange_cond_txt
|
| 362 |
-
# update (4.18)
|
| 363 |
-
self.multiple_optimizers = multiple_optimizers
|
| 364 |
-
self.add_glyph_control = False
|
| 365 |
-
self.glyph_control_key = glyph_control_key
|
| 366 |
-
self.freeze_glyph_image_encoder = True
|
| 367 |
-
self.glyph_image_encoder_type = "CLIP"
|
| 368 |
-
self.max_step = max_step
|
| 369 |
-
self.trans_glyph_embed = False
|
| 370 |
-
self.trans_glyph_lr = trans_glyph_lr
|
| 371 |
-
if deepspeed:
|
| 372 |
-
try:
|
| 373 |
-
from deepspeed.ops.adam import FusedAdam, DeepSpeedCPUAdam
|
| 374 |
-
self.optimizer = DeepSpeedCPUAdam #FusedAdam
|
| 375 |
-
except:
|
| 376 |
-
print("could not import FuseAdam from deepspeed")
|
| 377 |
-
self.optimizer = torch.optim.AdamW
|
| 378 |
-
else:
|
| 379 |
-
self.optimizer = torch.optim.AdamW
|
| 380 |
|
| 381 |
-
|
| 382 |
-
self.add_glyph_control = True
|
| 383 |
-
self.glycon_wd = glycon_wd
|
| 384 |
-
self.glycon_lr = glycon_lr
|
| 385 |
-
self.glycon_sched = glycon_sched
|
| 386 |
-
self.instantiate_glyph_control_model(glyph_control_config)
|
| 387 |
-
if self.glyph_control_model.trans_glyph_emb_model is not None:
|
| 388 |
-
self.trans_glyph_embed = True
|
| 389 |
-
|
| 390 |
self.use_ema = use_ema
|
| 391 |
-
if self.use_ema:
|
| 392 |
-
# assert self.sd_locked == True
|
| 393 |
self.model_ema = LitEma(self.control_model, init_num_updates= 0)
|
| 394 |
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
|
| 395 |
-
if not self.sd_locked:
|
| 396 |
self.model_diffoutblock_ema = LitEma(self.model.diffusion_model.output_blocks, init_num_updates= 0)
|
| 397 |
print(f"Keeping diffoutblock EMAs of {len(list(self.model_diffoutblock_ema.buffers()))}.")
|
| 398 |
self.model_diffout_ema = LitEma(self.model.diffusion_model.out, init_num_updates= 0)
|
| 399 |
print(f"Keeping diffout EMAs of {len(list(self.model_diffout_ema.buffers()))}.")
|
| 400 |
-
if not self.freeze_glyph_image_encoder:
|
| 401 |
-
self.model_glyphcon_ema = LitEma(self.glyph_control_model.image_encoder, init_num_updates=0)
|
| 402 |
-
print(f"Keeping glyphcon EMAs of {len(list(self.model_glyphcon_ema.buffers()))}.")
|
| 403 |
-
if self.trans_glyph_embed:
|
| 404 |
-
self.model_transglyph_ema = LitEma(self.glyph_control_model.trans_glyph_emb_model, init_num_updates=0)
|
| 405 |
-
print(f"Keeping glyphcon EMAs of {len(list(self.model_transglyph_ema.buffers()))}.")
|
| 406 |
|
|
|
|
| 407 |
if ckpt_path is not None:
|
| 408 |
ema_num_updates = self.init_from_ckpt(ckpt_path, ignore_keys, only_model=only_model)
|
| 409 |
self.restarted_from_ckpt = True
|
| 410 |
-
# if reset_ema:
|
| 411 |
-
# assert self.use_ema
|
| 412 |
if self.use_ema and reset_ema:
|
| 413 |
print(
|
| 414 |
f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.")
|
| 415 |
self.model_ema = LitEma(self.control_model, init_num_updates= ema_num_updates if keep_num_ema_updates else 0)
|
| 416 |
-
if not self.sd_locked:
|
| 417 |
self.model_diffoutblock_ema = LitEma(self.model.diffusion_model.output_blocks, init_num_updates= ema_num_updates if keep_num_ema_updates else 0)
|
| 418 |
self.model_diffout_ema = LitEma(self.model.diffusion_model.out, init_num_updates= ema_num_updates if keep_num_ema_updates else 0)
|
| 419 |
-
if not self.freeze_glyph_image_encoder:
|
| 420 |
-
self.model_glyphcon_ema = LitEma(self.glyph_control_model.image_encoder, init_num_updates= ema_num_updates if keep_num_ema_updates else 0)
|
| 421 |
-
if self.trans_glyph_embed:
|
| 422 |
-
self.model_transglyph_ema = LitEma(self.glyph_control_model.trans_glyph_emb_model, init_num_updates= ema_num_updates if keep_num_ema_updates else 0)
|
| 423 |
|
| 424 |
if reset_num_ema_updates:
|
| 425 |
print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ")
|
|
@@ -428,13 +387,6 @@ class ControlLDM(LatentDiffusion):
|
|
| 428 |
if not self.sd_locked: # Update
|
| 429 |
self.model_diffoutblock_ema.reset_num_updates()
|
| 430 |
self.model_diffout_ema.reset_num_updates()
|
| 431 |
-
if not self.freeze_glyph_image_encoder:
|
| 432 |
-
self.model_glyphcon_ema.reset_num_updates()
|
| 433 |
-
if self.trans_glyph_embed:
|
| 434 |
-
self.model_transglyph_ema.reset_num_updates()
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
# self.freeze_unet()
|
| 438 |
|
| 439 |
@contextmanager
|
| 440 |
def ema_scope(self, context=None):
|
|
@@ -446,12 +398,6 @@ class ControlLDM(LatentDiffusion):
|
|
| 446 |
self.model_diffoutblock_ema.copy_to(self.model.diffusion_model.output_blocks)
|
| 447 |
self.model_diffout_ema.store(self.model.diffusion_model.out.parameters())
|
| 448 |
self.model_diffout_ema.copy_to(self.model.diffusion_model.out)
|
| 449 |
-
if not self.freeze_glyph_image_encoder:
|
| 450 |
-
self.model_glyphcon_ema.store(self.glyph_control_model.image_encoder.parameters())
|
| 451 |
-
self.model_glyphcon_ema.copy_to(self.glyph_control_model.image_encoder)
|
| 452 |
-
if self.trans_glyph_embed:
|
| 453 |
-
self.model_transglyph_ema.store(self.glyph_control_model.trans_glyph_emb_model.parameters())
|
| 454 |
-
self.model_transglyph_ema.copy_to(self.glyph_control_model.trans_glyph_emb_model)
|
| 455 |
|
| 456 |
if context is not None:
|
| 457 |
print(f"{context}: Switched ControlNet to EMA weights")
|
|
@@ -463,10 +409,6 @@ class ControlLDM(LatentDiffusion):
|
|
| 463 |
if not self.sd_locked: # Update
|
| 464 |
self.model_diffoutblock_ema.restore(self.model.diffusion_model.output_blocks.parameters())
|
| 465 |
self.model_diffout_ema.restore(self.model.diffusion_model.out.parameters())
|
| 466 |
-
if not self.freeze_glyph_image_encoder:
|
| 467 |
-
self.model_glyphcon_ema.restore(self.glyph_control_model.image_encoder.parameters())
|
| 468 |
-
if self.trans_glyph_embed:
|
| 469 |
-
self.model_transglyph_ema.restore(self.glyph_control_model.trans_glyph_emb_model.parameters())
|
| 470 |
if context is not None:
|
| 471 |
print(f"{context}: Restored training weights of ControlNet")
|
| 472 |
|
|
@@ -493,14 +435,8 @@ class ControlLDM(LatentDiffusion):
|
|
| 493 |
if k.startswith(ik):
|
| 494 |
print("Deleting key {} from state_dict.".format(k))
|
| 495 |
del sd[k]
|
| 496 |
-
|
| 497 |
-
|
| 498 |
-
if not only_model:
|
| 499 |
-
missing, unexpected = self.load_state_dict(sd, strict=False)
|
| 500 |
-
elif path.endswith(".bin"):
|
| 501 |
-
missing, unexpected = self.model.diffusion_model.load_state_dict(sd, strict=False)
|
| 502 |
-
elif path.endswith(".ckpt"):
|
| 503 |
-
missing, unexpected = self.model.load_state_dict(sd, strict=False)
|
| 504 |
|
| 505 |
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
| 506 |
if len(missing) > 0:
|
|
@@ -513,28 +449,6 @@ class ControlLDM(LatentDiffusion):
|
|
| 513 |
else:
|
| 514 |
return 0
|
| 515 |
|
| 516 |
-
def instantiate_trans_textemb_model(self, config):
|
| 517 |
-
model = instantiate_from_config(config)
|
| 518 |
-
params = []
|
| 519 |
-
for i in range(model.emb_num):
|
| 520 |
-
if model.trans_trainable[i]:
|
| 521 |
-
params += list(model.trans_list[i].parameters())
|
| 522 |
-
else:
|
| 523 |
-
for param in model.trans_list[i].parameters():
|
| 524 |
-
param.requires_grad = False
|
| 525 |
-
self.trans_textemb_model = model
|
| 526 |
-
self.trans_textemb_params = params
|
| 527 |
-
|
| 528 |
-
# add
|
| 529 |
-
def instantiate_glyph_control_model(self, config):
|
| 530 |
-
model = instantiate_from_config(config)
|
| 531 |
-
# params = []
|
| 532 |
-
self.freeze_glyph_image_encoder = model.freeze_image_encoder #image_encoder.freeze_model
|
| 533 |
-
self.glyph_control_model = model
|
| 534 |
-
self.glyph_image_encoder_type = model.image_encoder_type
|
| 535 |
-
|
| 536 |
-
|
| 537 |
-
|
| 538 |
@torch.no_grad()
|
| 539 |
def get_input(self, batch, k, bs=None, *args, **kwargs):
|
| 540 |
x, c = super().get_input(batch, self.first_stage_key, *args, **kwargs)
|
|
@@ -544,80 +458,42 @@ class ControlLDM(LatentDiffusion):
|
|
| 544 |
control = control.to(self.device)
|
| 545 |
control = einops.rearrange(control, 'b h w c -> b c h w')
|
| 546 |
control = control.to(memory_format=torch.contiguous_format).float()
|
| 547 |
-
|
| 548 |
-
if self.add_glyph_control:
|
| 549 |
-
assert self.glyph_control_key in batch.keys()
|
| 550 |
-
glyph_control = batch[self.glyph_control_key]
|
| 551 |
-
if bs is not None:
|
| 552 |
-
glyph_control = glyph_control[:bs]
|
| 553 |
-
glycon_samples = []
|
| 554 |
-
for glycon_sample in glyph_control:
|
| 555 |
-
glycon_sample = glycon_sample.to(self.device)
|
| 556 |
-
glycon_sample = einops.rearrange(glycon_sample, 'b h w c -> b c h w')
|
| 557 |
-
glycon_sample = glycon_sample.to(memory_format=torch.contiguous_format).float()
|
| 558 |
-
glycon_samples.append(glycon_sample)
|
| 559 |
-
# return x, dict(c_crossattn=[c], c_concat=[control])
|
| 560 |
-
return x, dict(c_crossattn=[c] if not isinstance(c, list) else c, c_concat=[control], c_glyph=glycon_samples)
|
| 561 |
return x, dict(c_crossattn=[c] if not isinstance(c, list) else c, c_concat=[control])
|
| 562 |
|
| 563 |
def apply_model(self, x_noisy, t, cond, *args, **kwargs):
|
| 564 |
assert isinstance(cond, dict)
|
| 565 |
diffusion_model = self.model.diffusion_model
|
| 566 |
-
|
| 567 |
-
#update
|
| 568 |
-
embdim_list = []
|
| 569 |
-
for c in cond["c_crossattn"]:
|
| 570 |
-
embdim_list.append(c.shape[-1])
|
| 571 |
-
embdim_list = np.array(embdim_list)
|
| 572 |
-
if np.sum(embdim_list != diffusion_model.context_dim):
|
| 573 |
-
assert self.trans_textemb
|
| 574 |
-
|
| 575 |
-
if self.trans_textemb:
|
| 576 |
-
assert self.trans_textemb_model
|
| 577 |
-
cond_txt_list = self.trans_textemb_model(cond["c_crossattn"])
|
| 578 |
-
# if len(cond_txt_list) == 2:
|
| 579 |
-
# print("cond_txt_2 max: {}".format(torch.max(torch.abs(cond_txt_list[1]))))
|
| 580 |
-
else:
|
| 581 |
-
cond_txt_list = cond["c_crossattn"]
|
| 582 |
-
|
| 583 |
|
| 584 |
assert len(cond_txt_list) > 0
|
| 585 |
-
|
| 586 |
-
|
| 587 |
-
|
|
|
|
|
|
|
| 588 |
else:
|
| 589 |
-
if
|
| 590 |
-
|
| 591 |
-
|
| 592 |
-
|
| 593 |
-
|
|
|
|
|
|
|
| 594 |
else:
|
| 595 |
cond_txt_2 = torch.cat(cond_txt_list, 1)
|
| 596 |
-
|
| 597 |
-
|
| 598 |
-
|
| 599 |
-
|
| 600 |
-
|
| 601 |
-
|
| 602 |
-
cond_txt = torch.cat(cond_txt_list, 1)
|
| 603 |
-
cond_txt_2 = None
|
| 604 |
-
|
| 605 |
-
context_glyph = None
|
| 606 |
-
if self.add_glyph_control:
|
| 607 |
-
assert "c_glyph" in cond.keys()
|
| 608 |
-
if cond["c_glyph"] is not None:
|
| 609 |
-
context_glyph = self.glyph_control_model(cond["c_glyph"], text_embed = cond_txt_list[-1] if len(cond_txt_list) == 3 else cond_txt)
|
| 610 |
-
else:
|
| 611 |
-
context_glyph = cond_txt_list[-1] if len(cond_txt_list) == 3 else cond_txt
|
| 612 |
-
# if cond_txt_2 is None:
|
| 613 |
-
# print("cond_txt_2 is None")
|
| 614 |
|
| 615 |
if cond['c_concat'] is None:
|
| 616 |
-
eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=None, only_mid_control=self.only_mid_control
|
| 617 |
else:
|
| 618 |
control = self.control_model(x=x_noisy, hint=torch.cat(cond['c_concat'], 1), timesteps=t, context=cond_txt if cond_txt_2 is None else cond_txt_2)
|
| 619 |
control = [c * scale for c, scale in zip(control, self.control_scales)]
|
| 620 |
-
eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=control, only_mid_control=self.only_mid_control
|
| 621 |
|
| 622 |
return eps
|
| 623 |
|
|
@@ -625,96 +501,16 @@ class ControlLDM(LatentDiffusion):
|
|
| 625 |
def get_unconditional_conditioning(self, N):
|
| 626 |
return self.get_learned_conditioning([""] * N)
|
| 627 |
|
| 628 |
-
# Maybe not useful: modify the codes to fit the separate input captions
|
| 629 |
-
# @torch.no_grad()
|
| 630 |
-
# def get_unconditional_conditioning(self, N):
|
| 631 |
-
# return self.get_learned_conditioning([""] * N) if not self.sep_cap_for_2b else self.get_learned_conditioning([[""] * N, [""] * N])
|
| 632 |
-
# TODO: adapt to new model
|
| 633 |
-
@torch.no_grad()
|
| 634 |
-
def log_images(self, batch, N=4, n_row=2, sample=False, ddim_steps=50, ddim_eta=0.0, return_keys=None,
|
| 635 |
-
quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
|
| 636 |
-
plot_diffusion_rows=False, unconditional_guidance_scale=9.0, unconditional_guidance_label=None,
|
| 637 |
-
use_ema_scope=True,
|
| 638 |
-
**kwargs):
|
| 639 |
-
use_ddim = ddim_steps is not None
|
| 640 |
-
|
| 641 |
-
log = dict()
|
| 642 |
-
z, c = self.get_input(batch, self.first_stage_key, bs=N)
|
| 643 |
-
c_cat, c = c["c_concat"][0][:N], c["c_crossattn"][0][:N]
|
| 644 |
-
N = min(z.shape[0], N)
|
| 645 |
-
n_row = min(z.shape[0], n_row)
|
| 646 |
-
log["reconstruction"] = self.decode_first_stage(z)
|
| 647 |
-
log["control"] = c_cat * 2.0 - 1.0
|
| 648 |
-
log["conditioning"] = log_txt_as_img((512, 512), batch[self.cond_stage_key], size=16)
|
| 649 |
-
|
| 650 |
-
if plot_diffusion_rows:
|
| 651 |
-
# get diffusion row
|
| 652 |
-
diffusion_row = list()
|
| 653 |
-
z_start = z[:n_row]
|
| 654 |
-
for t in range(self.num_timesteps):
|
| 655 |
-
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
|
| 656 |
-
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
|
| 657 |
-
t = t.to(self.device).long()
|
| 658 |
-
noise = torch.randn_like(z_start)
|
| 659 |
-
z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
|
| 660 |
-
diffusion_row.append(self.decode_first_stage(z_noisy))
|
| 661 |
-
|
| 662 |
-
diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
|
| 663 |
-
diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
|
| 664 |
-
diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
|
| 665 |
-
diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
|
| 666 |
-
log["diffusion_row"] = diffusion_grid
|
| 667 |
-
|
| 668 |
-
if sample:
|
| 669 |
-
# get denoise row
|
| 670 |
-
samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
|
| 671 |
-
batch_size=N, ddim=use_ddim,
|
| 672 |
-
ddim_steps=ddim_steps, eta=ddim_eta)
|
| 673 |
-
x_samples = self.decode_first_stage(samples)
|
| 674 |
-
log["samples"] = x_samples
|
| 675 |
-
if plot_denoise_rows:
|
| 676 |
-
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
|
| 677 |
-
log["denoise_row"] = denoise_grid
|
| 678 |
-
|
| 679 |
-
if unconditional_guidance_scale > 1.0:
|
| 680 |
-
uc_cross = self.get_unconditional_conditioning(N)
|
| 681 |
-
uc_cat = c_cat # torch.zeros_like(c_cat)
|
| 682 |
-
uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]}
|
| 683 |
-
samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
|
| 684 |
-
batch_size=N, ddim=use_ddim,
|
| 685 |
-
ddim_steps=ddim_steps, eta=ddim_eta,
|
| 686 |
-
unconditional_guidance_scale=unconditional_guidance_scale,
|
| 687 |
-
unconditional_conditioning=uc_full,
|
| 688 |
-
)
|
| 689 |
-
x_samples_cfg = self.decode_first_stage(samples_cfg)
|
| 690 |
-
log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
|
| 691 |
-
|
| 692 |
-
return log
|
| 693 |
-
# TODO: adapt to new model
|
| 694 |
-
@torch.no_grad()
|
| 695 |
-
def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs):
|
| 696 |
-
ddim_sampler = DDIMSampler(self)
|
| 697 |
-
b, c, h, w = cond["c_concat"][0].shape
|
| 698 |
-
shape = (self.channels, h // 8, w // 8)
|
| 699 |
-
samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, shape, cond, verbose=False, **kwargs)
|
| 700 |
-
return samples, intermediates
|
| 701 |
-
# add
|
| 702 |
def training_step(self, batch, batch_idx, optimizer_idx=0):
|
| 703 |
loss = super().training_step(batch, batch_idx, optimizer_idx)
|
| 704 |
if self.use_scheduler and not self.sd_locked and self.sep_lr:
|
| 705 |
decoder_lr = self.optimizers().param_groups[1]["lr"]
|
| 706 |
self.log('decoder_lr_abs', decoder_lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
| 707 |
-
if self.trans_glyph_embed and self.freeze_glyph_image_encoder:
|
| 708 |
-
trans_glyph_embed_lr = self.optimizers().param_groups[2]["lr"]
|
| 709 |
-
self.log('trans_glyph_embed_lr_abs', trans_glyph_embed_lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
| 710 |
return loss
|
| 711 |
|
| 712 |
def configure_optimizers(self):
|
| 713 |
lr = self.learning_rate
|
| 714 |
-
params = list(self.control_model.parameters())
|
| 715 |
-
if self.trans_textemb:
|
| 716 |
-
params += self.trans_textemb_params #list(self.trans_textemb_model.parameters())
|
| 717 |
-
|
| 718 |
if self.learnable_conscale:
|
| 719 |
params += [self.control_scales]
|
| 720 |
|
|
@@ -731,34 +527,9 @@ class ControlLDM(LatentDiffusion):
|
|
| 731 |
if decoder_params is not None:
|
| 732 |
params_wlr.append({"params": decoder_params, "lr": self.decoder_lr})
|
| 733 |
|
| 734 |
-
if not self.freeze_glyph_image_encoder:
|
| 735 |
-
if self.glyph_image_encoder_type == "CLIP":
|
| 736 |
-
# assert self.sep_lr
|
| 737 |
-
# follow the training codes in the OpenClip repo
|
| 738 |
-
# https://github.com/mlfoundations/open_clip/blob/main/src/training/main.py#L303
|
| 739 |
-
exclude = lambda n, p: p.ndim < 2 or "bn" in n or "ln" in n or "bias" in n or 'logit_scale' in n
|
| 740 |
-
include = lambda n, p: not exclude(n, p)
|
| 741 |
-
|
| 742 |
-
# named_parameters = list(model.image_encoder.named_parameters())
|
| 743 |
-
named_parameters = list(self.glyph_control_model.image_encoder.named_parameters())
|
| 744 |
-
gain_or_bias_params = [p for n, p in named_parameters if exclude(n, p) and p.requires_grad]
|
| 745 |
-
rest_params = [p for n, p in named_parameters if include(n, p) and p.requires_grad]
|
| 746 |
-
self.glyph_control_params_wlr = [
|
| 747 |
-
{"params": gain_or_bias_params, "weight_decay": 0., "lr": self.glycon_lr},
|
| 748 |
-
{"params": rest_params, "weight_decay": self.glycon_wd, "lr": self.glycon_lr},
|
| 749 |
-
]
|
| 750 |
-
if not self.freeze_glyph_image_encoder and not self.multiple_optimizers:
|
| 751 |
-
params_wlr.extend(self.glyph_control_params_wlr)
|
| 752 |
-
|
| 753 |
-
if self.trans_glyph_embed:
|
| 754 |
-
trans_glyph_params = list(self.glyph_control_model.trans_glyph_emb_model.parameters())
|
| 755 |
-
params_wlr.append({"params": trans_glyph_params, "lr": self.trans_glyph_lr})
|
| 756 |
# opt = torch.optim.AdamW(params_wlr)
|
| 757 |
opt = self.optimizer(params_wlr)
|
| 758 |
opts = [opt]
|
| 759 |
-
if not self.freeze_glyph_image_encoder and self.multiple_optimizers:
|
| 760 |
-
glyph_control_opt = self.optimizer(self.glyph_control_params_wlr) #torch.optim.AdamW(self.glyph_control_params_wlr)
|
| 761 |
-
opts.append(glyph_control_opt)
|
| 762 |
|
| 763 |
# updated
|
| 764 |
schedulers = []
|
|
@@ -776,33 +547,8 @@ class ControlLDM(LatentDiffusion):
|
|
| 776 |
'frequency': 1
|
| 777 |
}]
|
| 778 |
|
| 779 |
-
if not self.freeze_glyph_image_encoder and self.multiple_optimizers:
|
| 780 |
-
if self.glycon_sched == "cosine" and self.max_step is not None:
|
| 781 |
-
glyph_scheduler = CosineAnnealingLR(glyph_control_opt, T_max=self.max_step) #: max_step
|
| 782 |
-
elif self.glycon_sched == "onecycle" and self.max_step is not None:
|
| 783 |
-
glyph_scheduler = OneCycleLR(
|
| 784 |
-
glyph_control_opt,
|
| 785 |
-
max_lr=self.glycon_lr,
|
| 786 |
-
total_steps=self.max_step, #: max_step
|
| 787 |
-
pct_start=0.0001,
|
| 788 |
-
anneal_strategy="cos" #'linear'
|
| 789 |
-
)
|
| 790 |
-
# elif self.glycon_sched == "lambda":
|
| 791 |
-
else:
|
| 792 |
-
glyph_scheduler = LambdaLR(
|
| 793 |
-
glyph_control_opt,
|
| 794 |
-
lr_lambda = [scheduler_func.schedule] * len(self.glyph_control_params_wlr)
|
| 795 |
-
)
|
| 796 |
-
schedulers.append(
|
| 797 |
-
{
|
| 798 |
-
"scheduler": glyph_scheduler,
|
| 799 |
-
"interval": 'step',
|
| 800 |
-
'frequency': 1
|
| 801 |
-
}
|
| 802 |
-
)
|
| 803 |
return opts, schedulers
|
| 804 |
-
|
| 805 |
-
# TODO: adapt to new model
|
| 806 |
def low_vram_shift(self, is_diffusing):
|
| 807 |
if is_diffusing:
|
| 808 |
self.model = self.model.cuda()
|
|
@@ -822,10 +568,6 @@ class ControlLDM(LatentDiffusion):
|
|
| 822 |
if not self.sd_locked: # Update
|
| 823 |
self.model_diffoutblock_ema(self.model.diffusion_model.output_blocks)
|
| 824 |
self.model_diffout_ema(self.model.diffusion_model.out)
|
| 825 |
-
if not self.freeze_glyph_image_encoder:
|
| 826 |
-
self.model_glyphcon_ema(self.glyph_control_model.image_encoder)
|
| 827 |
-
if self.trans_glyph_embed:
|
| 828 |
-
self.model_transglyph_ema(self.glyph_control_model.trans_glyph_emb_model)
|
| 829 |
if self.log_all_grad_norm:
|
| 830 |
zeroconvs = list(self.control_model.input_hint_block.named_parameters())[-2:]
|
| 831 |
zeroconvs.extend(
|
|
@@ -867,43 +609,6 @@ class ControlLDM(LatentDiffusion):
|
|
| 867 |
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 868 |
)
|
| 869 |
|
| 870 |
-
if self.trans_textemb:
|
| 871 |
-
for name, p in self.trans_textemb_model.named_parameters():
|
| 872 |
-
if p.requires_grad and p.grad is not None:
|
| 873 |
-
self.log(
|
| 874 |
-
"trans_textemb_gradient_norm/{}".format(name),
|
| 875 |
-
p.grad.cpu().detach().norm().item(),
|
| 876 |
-
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 877 |
-
)
|
| 878 |
-
self.log(
|
| 879 |
-
"trans_textemb_params/{}_norm".format(name),
|
| 880 |
-
p.cpu().detach().norm().item(),
|
| 881 |
-
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 882 |
-
)
|
| 883 |
-
self.log(
|
| 884 |
-
"trans_textemb_params/{}_abs_max".format(name),
|
| 885 |
-
torch.max(torch.abs(p.cpu().detach())).item(),
|
| 886 |
-
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 887 |
-
)
|
| 888 |
-
if self.trans_glyph_embed:
|
| 889 |
-
for name, p in self.glyph_control_model.trans_glyph_emb_model.named_parameters():
|
| 890 |
-
if p.requires_grad and p.grad is not None:
|
| 891 |
-
self.log(
|
| 892 |
-
"trans_glyph_embed_gradient_norm/{}".format(name),
|
| 893 |
-
p.grad.cpu().detach().norm().item(),
|
| 894 |
-
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 895 |
-
)
|
| 896 |
-
self.log(
|
| 897 |
-
"trans_glyph_embed_params/{}_norm".format(name),
|
| 898 |
-
p.cpu().detach().norm().item(),
|
| 899 |
-
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 900 |
-
)
|
| 901 |
-
self.log(
|
| 902 |
-
"trans_glyph_embed_params/{}_abs_max".format(name),
|
| 903 |
-
torch.max(torch.abs(p.cpu().detach())).item(),
|
| 904 |
-
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 905 |
-
)
|
| 906 |
-
|
| 907 |
if self.learnable_conscale:
|
| 908 |
for i in range(len(self.control_scales)):
|
| 909 |
self.log(
|
|
@@ -912,4 +617,4 @@ class ControlLDM(LatentDiffusion):
|
|
| 912 |
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 913 |
)
|
| 914 |
del gradnorm_list
|
| 915 |
-
del zeroconvs
|
|
|
|
| 28 |
return self
|
| 29 |
|
| 30 |
class ControlledUnetModel(UNetModel):
|
| 31 |
+
def forward(self, x, timesteps=None, context=None, control=None, only_mid_control=False, **kwargs):
|
| 32 |
hs = []
|
| 33 |
with torch.no_grad():
|
| 34 |
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
|
|
|
|
| 47 |
h = torch.cat([h, hs.pop()], dim=1)
|
| 48 |
else:
|
| 49 |
h = torch.cat([h, hs.pop() + control.pop()], dim=1)
|
| 50 |
+
h = module(h, emb, context)
|
| 51 |
|
| 52 |
h = h.type(x.dtype)
|
| 53 |
return self.out(h)
|
|
|
|
| 317 |
|
| 318 |
def __init__(self,
|
| 319 |
control_stage_config,
|
| 320 |
+
control_key, only_mid_control,
|
|
|
|
|
|
|
| 321 |
learnable_conscale = False, guess_mode=False,
|
| 322 |
+
sd_locked = True, sep_lr = False, decoder_lr = 1.0**-4,
|
| 323 |
+
sep_cond_txt = True, exchange_cond_txt = False, concat_all_textemb = False,
|
|
|
|
|
|
|
| 324 |
*args, **kwargs
|
| 325 |
+
):
|
| 326 |
use_ema = kwargs.pop("use_ema", False)
|
| 327 |
ckpt_path = kwargs.pop("ckpt_path", None)
|
| 328 |
reset_ema = kwargs.pop("reset_ema", False)
|
|
|
|
| 332 |
ignore_keys = kwargs.pop("ignore_keys", [])
|
| 333 |
|
| 334 |
super().__init__(*args, use_ema=False, **kwargs)
|
| 335 |
+
|
| 336 |
+
# Glyph ControlNet
|
| 337 |
self.control_model = instantiate_from_config(control_stage_config)
|
| 338 |
self.control_key = control_key
|
| 339 |
self.only_mid_control = only_mid_control
|
| 340 |
+
|
| 341 |
self.learnable_conscale = learnable_conscale
|
| 342 |
conscale_init = [1.0] * 13 if not guess_mode else [(0.825 ** float(12 - i)) for i in range(13)]
|
| 343 |
if learnable_conscale:
|
| 344 |
# self.control_scales = nn.Parameter(torch.ones(13), requires_grad=True)
|
| 345 |
self.control_scales = nn.Parameter(torch.Tensor(conscale_init), requires_grad=True)
|
| 346 |
+
else:
|
| 347 |
self.control_scales = conscale_init #[1.0] * 13
|
| 348 |
+
|
| 349 |
+
self.optimizer = torch.optim.AdamW
|
| 350 |
+
# whether to unlock (fine-tune) the decoder parts of SD U-Net
|
| 351 |
self.sd_locked = sd_locked
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 352 |
self.sep_lr = sep_lr
|
| 353 |
self.decoder_lr = decoder_lr
|
| 354 |
+
|
| 355 |
+
# specify the input text embedding of two branches (SD branch and Glyph ControlNet branch)
|
| 356 |
self.sep_cond_txt = sep_cond_txt
|
| 357 |
+
self.concat_all_textemb = concat_all_textemb
|
| 358 |
self.exchange_cond_txt = exchange_cond_txt
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 359 |
|
| 360 |
+
# ema
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 361 |
self.use_ema = use_ema
|
| 362 |
+
if self.use_ema:
|
|
|
|
| 363 |
self.model_ema = LitEma(self.control_model, init_num_updates= 0)
|
| 364 |
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
|
| 365 |
+
if not self.sd_locked:
|
| 366 |
self.model_diffoutblock_ema = LitEma(self.model.diffusion_model.output_blocks, init_num_updates= 0)
|
| 367 |
print(f"Keeping diffoutblock EMAs of {len(list(self.model_diffoutblock_ema.buffers()))}.")
|
| 368 |
self.model_diffout_ema = LitEma(self.model.diffusion_model.out, init_num_updates= 0)
|
| 369 |
print(f"Keeping diffout EMAs of {len(list(self.model_diffout_ema.buffers()))}.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 370 |
|
| 371 |
+
# initialize the model from the checkpoint
|
| 372 |
if ckpt_path is not None:
|
| 373 |
ema_num_updates = self.init_from_ckpt(ckpt_path, ignore_keys, only_model=only_model)
|
| 374 |
self.restarted_from_ckpt = True
|
|
|
|
|
|
|
| 375 |
if self.use_ema and reset_ema:
|
| 376 |
print(
|
| 377 |
f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.")
|
| 378 |
self.model_ema = LitEma(self.control_model, init_num_updates= ema_num_updates if keep_num_ema_updates else 0)
|
| 379 |
+
if not self.sd_locked:
|
| 380 |
self.model_diffoutblock_ema = LitEma(self.model.diffusion_model.output_blocks, init_num_updates= ema_num_updates if keep_num_ema_updates else 0)
|
| 381 |
self.model_diffout_ema = LitEma(self.model.diffusion_model.out, init_num_updates= ema_num_updates if keep_num_ema_updates else 0)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 382 |
|
| 383 |
if reset_num_ema_updates:
|
| 384 |
print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ")
|
|
|
|
| 387 |
if not self.sd_locked: # Update
|
| 388 |
self.model_diffoutblock_ema.reset_num_updates()
|
| 389 |
self.model_diffout_ema.reset_num_updates()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 390 |
|
| 391 |
@contextmanager
|
| 392 |
def ema_scope(self, context=None):
|
|
|
|
| 398 |
self.model_diffoutblock_ema.copy_to(self.model.diffusion_model.output_blocks)
|
| 399 |
self.model_diffout_ema.store(self.model.diffusion_model.out.parameters())
|
| 400 |
self.model_diffout_ema.copy_to(self.model.diffusion_model.out)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 401 |
|
| 402 |
if context is not None:
|
| 403 |
print(f"{context}: Switched ControlNet to EMA weights")
|
|
|
|
| 409 |
if not self.sd_locked: # Update
|
| 410 |
self.model_diffoutblock_ema.restore(self.model.diffusion_model.output_blocks.parameters())
|
| 411 |
self.model_diffout_ema.restore(self.model.diffusion_model.out.parameters())
|
|
|
|
|
|
|
|
|
|
|
|
|
| 412 |
if context is not None:
|
| 413 |
print(f"{context}: Restored training weights of ControlNet")
|
| 414 |
|
|
|
|
| 435 |
if k.startswith(ik):
|
| 436 |
print("Deleting key {} from state_dict.".format(k))
|
| 437 |
del sd[k]
|
| 438 |
+
missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
|
| 439 |
+
sd, strict=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 440 |
|
| 441 |
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
|
| 442 |
if len(missing) > 0:
|
|
|
|
| 449 |
else:
|
| 450 |
return 0
|
| 451 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 452 |
@torch.no_grad()
|
| 453 |
def get_input(self, batch, k, bs=None, *args, **kwargs):
|
| 454 |
x, c = super().get_input(batch, self.first_stage_key, *args, **kwargs)
|
|
|
|
| 458 |
control = control.to(self.device)
|
| 459 |
control = einops.rearrange(control, 'b h w c -> b c h w')
|
| 460 |
control = control.to(memory_format=torch.contiguous_format).float()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 461 |
return x, dict(c_crossattn=[c] if not isinstance(c, list) else c, c_concat=[control])
|
| 462 |
|
| 463 |
def apply_model(self, x_noisy, t, cond, *args, **kwargs):
|
| 464 |
assert isinstance(cond, dict)
|
| 465 |
diffusion_model = self.model.diffusion_model
|
| 466 |
+
cond_txt_list = cond["c_crossattn"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 467 |
|
| 468 |
assert len(cond_txt_list) > 0
|
| 469 |
+
# cond_txt: input text embedding of the pretrained SD branch
|
| 470 |
+
# cond_txt_2: input text embedding of the Glyph ControlNet branch
|
| 471 |
+
cond_txt = cond_txt_list[0]
|
| 472 |
+
if len(cond_txt_list) == 1:
|
| 473 |
+
cond_txt_2 = None
|
| 474 |
else:
|
| 475 |
+
if self.sep_cond_txt:
|
| 476 |
+
# use each embedding for each branch separately
|
| 477 |
+
cond_txt_2 = cond_txt_list[1]
|
| 478 |
+
else:
|
| 479 |
+
# concat the embedding for Glyph ControlNet branch
|
| 480 |
+
if not self.concat_all_textemb:
|
| 481 |
+
cond_txt_2 = torch.cat(cond_txt_list[1:], 1)
|
| 482 |
else:
|
| 483 |
cond_txt_2 = torch.cat(cond_txt_list, 1)
|
| 484 |
+
|
| 485 |
+
if self.exchange_cond_txt:
|
| 486 |
+
# exchange the input text embedding of two branches
|
| 487 |
+
txt_buffer = cond_txt
|
| 488 |
+
cond_txt = cond_txt_2
|
| 489 |
+
cond_txt_2 = txt_buffer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 490 |
|
| 491 |
if cond['c_concat'] is None:
|
| 492 |
+
eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=None, only_mid_control=self.only_mid_control)
|
| 493 |
else:
|
| 494 |
control = self.control_model(x=x_noisy, hint=torch.cat(cond['c_concat'], 1), timesteps=t, context=cond_txt if cond_txt_2 is None else cond_txt_2)
|
| 495 |
control = [c * scale for c, scale in zip(control, self.control_scales)]
|
| 496 |
+
eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=control, only_mid_control=self.only_mid_control)
|
| 497 |
|
| 498 |
return eps
|
| 499 |
|
|
|
|
| 501 |
def get_unconditional_conditioning(self, N):
|
| 502 |
return self.get_learned_conditioning([""] * N)
|
| 503 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 504 |
def training_step(self, batch, batch_idx, optimizer_idx=0):
|
| 505 |
loss = super().training_step(batch, batch_idx, optimizer_idx)
|
| 506 |
if self.use_scheduler and not self.sd_locked and self.sep_lr:
|
| 507 |
decoder_lr = self.optimizers().param_groups[1]["lr"]
|
| 508 |
self.log('decoder_lr_abs', decoder_lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
|
|
|
|
|
|
|
|
|
|
| 509 |
return loss
|
| 510 |
|
| 511 |
def configure_optimizers(self):
|
| 512 |
lr = self.learning_rate
|
| 513 |
+
params = list(self.control_model.parameters())
|
|
|
|
|
|
|
|
|
|
| 514 |
if self.learnable_conscale:
|
| 515 |
params += [self.control_scales]
|
| 516 |
|
|
|
|
| 527 |
if decoder_params is not None:
|
| 528 |
params_wlr.append({"params": decoder_params, "lr": self.decoder_lr})
|
| 529 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 530 |
# opt = torch.optim.AdamW(params_wlr)
|
| 531 |
opt = self.optimizer(params_wlr)
|
| 532 |
opts = [opt]
|
|
|
|
|
|
|
|
|
|
| 533 |
|
| 534 |
# updated
|
| 535 |
schedulers = []
|
|
|
|
| 547 |
'frequency': 1
|
| 548 |
}]
|
| 549 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 550 |
return opts, schedulers
|
| 551 |
+
|
|
|
|
| 552 |
def low_vram_shift(self, is_diffusing):
|
| 553 |
if is_diffusing:
|
| 554 |
self.model = self.model.cuda()
|
|
|
|
| 568 |
if not self.sd_locked: # Update
|
| 569 |
self.model_diffoutblock_ema(self.model.diffusion_model.output_blocks)
|
| 570 |
self.model_diffout_ema(self.model.diffusion_model.out)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 571 |
if self.log_all_grad_norm:
|
| 572 |
zeroconvs = list(self.control_model.input_hint_block.named_parameters())[-2:]
|
| 573 |
zeroconvs.extend(
|
|
|
|
| 609 |
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 610 |
)
|
| 611 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 612 |
if self.learnable_conscale:
|
| 613 |
for i in range(len(self.control_scales)):
|
| 614 |
self.log(
|
|
|
|
| 617 |
prog_bar=False, logger=True, on_step=True, on_epoch=False
|
| 618 |
)
|
| 619 |
del gradnorm_list
|
| 620 |
+
del zeroconvs
|
scripts/rendertext_tool.py
CHANGED
|
@@ -75,12 +75,14 @@ class Render_Text:
|
|
| 75 |
def __init__(self,
|
| 76 |
model,
|
| 77 |
precision_scope=nullcontext,
|
| 78 |
-
transform=ToTensor()
|
|
|
|
| 79 |
):
|
| 80 |
self.model = model
|
| 81 |
self.precision_scope = precision_scope
|
| 82 |
self.transform = transform
|
| 83 |
self.ddim_sampler = DDIMSampler(model)
|
|
|
|
| 84 |
|
| 85 |
def process_multi(self,
|
| 86 |
rendered_txt_values, shared_prompt,
|
|
@@ -138,11 +140,36 @@ class Render_Text:
|
|
| 138 |
shared_seed = random.randint(0, 65535)
|
| 139 |
seed_everything(shared_seed)
|
| 140 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
print("control is None: {}".format(control is None))
|
| 142 |
-
|
| 143 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
un_cond_cross = self.model.get_learned_conditioning([shared_n_prompt] * shared_num_samples)
|
| 145 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
cond = {"c_concat": control, "c_crossattn": [cond_c_cross] if not isinstance(cond_c_cross, list) else cond_c_cross}
|
| 147 |
un_cond = {"c_concat": None if shared_guess_mode else control, "c_crossattn": [un_cond_cross] if not isinstance(un_cond_cross, list) else un_cond_cross}
|
| 148 |
shape = (4, H // 8, W // 8)
|
|
@@ -155,6 +182,9 @@ class Render_Text:
|
|
| 155 |
shape, cond, verbose=False, eta=shared_eta,
|
| 156 |
unconditional_guidance_scale=shared_scale,
|
| 157 |
unconditional_conditioning=un_cond)
|
|
|
|
|
|
|
|
|
|
| 158 |
|
| 159 |
x_samples = self.model.decode_first_stage(samples)
|
| 160 |
x_samples = (einops.rearrange(x_samples, 'b c h w -> b h w c') * 127.5 + 127.5).cpu().numpy().clip(0, 255).astype(np.uint8)
|
|
|
|
| 75 |
def __init__(self,
|
| 76 |
model,
|
| 77 |
precision_scope=nullcontext,
|
| 78 |
+
transform=ToTensor(),
|
| 79 |
+
save_memory = False,
|
| 80 |
):
|
| 81 |
self.model = model
|
| 82 |
self.precision_scope = precision_scope
|
| 83 |
self.transform = transform
|
| 84 |
self.ddim_sampler = DDIMSampler(model)
|
| 85 |
+
self.save_memory = save_memory
|
| 86 |
|
| 87 |
def process_multi(self,
|
| 88 |
rendered_txt_values, shared_prompt,
|
|
|
|
| 140 |
shared_seed = random.randint(0, 65535)
|
| 141 |
seed_everything(shared_seed)
|
| 142 |
|
| 143 |
+
if torch.cuda.is_available() and self.save_memory:
|
| 144 |
+
print("low_vram_shift: is_diffusing", False)
|
| 145 |
+
self.model.low_vram_shift(is_diffusing=False)
|
| 146 |
+
|
| 147 |
print("control is None: {}".format(control is None))
|
| 148 |
+
if shared_prompt.endswith("."):
|
| 149 |
+
if shared_a_prompt == "":
|
| 150 |
+
c_prompt = shared_prompt
|
| 151 |
+
else:
|
| 152 |
+
c_prompt = shared_prompt + " " + shared_a_prompt
|
| 153 |
+
elif shared_prompt.endswith(","):
|
| 154 |
+
if shared_a_prompt == "":
|
| 155 |
+
c_prompt = shared_prompt[:-1] + "."
|
| 156 |
+
else:
|
| 157 |
+
c_prompt = shared_prompt + " " + shared_a_prompt
|
| 158 |
+
else:
|
| 159 |
+
if shared_a_prompt == "":
|
| 160 |
+
c_prompt = shared_prompt + "."
|
| 161 |
+
else:
|
| 162 |
+
c_prompt = shared_prompt + ", " + shared_a_prompt
|
| 163 |
+
|
| 164 |
+
# cond_c_cross = self.model.get_learned_conditioning([shared_prompt + ', ' + shared_a_prompt] * shared_num_samples)
|
| 165 |
+
cond_c_cross = self.model.get_learned_conditioning([c_prompt] * shared_num_samples)
|
| 166 |
+
print("prompt:", c_prompt)
|
| 167 |
un_cond_cross = self.model.get_learned_conditioning([shared_n_prompt] * shared_num_samples)
|
| 168 |
|
| 169 |
+
if torch.cuda.is_available() and self.save_memory:
|
| 170 |
+
print("low_vram_shift: is_diffusing", True)
|
| 171 |
+
self.model.low_vram_shift(is_diffusing=True)
|
| 172 |
+
|
| 173 |
cond = {"c_concat": control, "c_crossattn": [cond_c_cross] if not isinstance(cond_c_cross, list) else cond_c_cross}
|
| 174 |
un_cond = {"c_concat": None if shared_guess_mode else control, "c_crossattn": [un_cond_cross] if not isinstance(un_cond_cross, list) else un_cond_cross}
|
| 175 |
shape = (4, H // 8, W // 8)
|
|
|
|
| 182 |
shape, cond, verbose=False, eta=shared_eta,
|
| 183 |
unconditional_guidance_scale=shared_scale,
|
| 184 |
unconditional_conditioning=un_cond)
|
| 185 |
+
if torch.cuda.is_available() and self.save_memory:
|
| 186 |
+
print("low_vram_shift: is_diffusing", False)
|
| 187 |
+
self.model.low_vram_shift(is_diffusing=False)
|
| 188 |
|
| 189 |
x_samples = self.model.decode_first_stage(samples)
|
| 190 |
x_samples = (einops.rearrange(x_samples, 'b c h w -> b h w c') * 127.5 + 127.5).cpu().numpy().clip(0, 255).astype(np.uint8)
|