|
|
|
@ -18,6 +18,10 @@ from imaginairy.utils import get_device
|
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def to_torch(x):
|
|
|
|
|
return x.clone().detach().to(torch.float32).to(get_device())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class DDIMSchedule:
|
|
|
|
|
def __init__(
|
|
|
|
|
self,
|
|
|
|
@ -26,32 +30,31 @@ class DDIMSchedule:
|
|
|
|
|
ddim_num_steps,
|
|
|
|
|
ddim_discretize="uniform",
|
|
|
|
|
ddim_eta=0.0,
|
|
|
|
|
device=get_device(),
|
|
|
|
|
):
|
|
|
|
|
device = get_device()
|
|
|
|
|
if not model_alphas_cumprod.shape[0] == model_num_timesteps:
|
|
|
|
|
raise ValueError("alphas have to be defined for each timestep")
|
|
|
|
|
|
|
|
|
|
self.alphas_cumprod = to_torch(model_alphas_cumprod)
|
|
|
|
|
|
|
|
|
|
ddim_timesteps = make_ddim_timesteps(
|
|
|
|
|
ddim_discr_method=ddim_discretize,
|
|
|
|
|
num_ddim_timesteps=ddim_num_steps,
|
|
|
|
|
num_ddpm_timesteps=model_num_timesteps,
|
|
|
|
|
)
|
|
|
|
|
alphas_cumprod = model_alphas_cumprod
|
|
|
|
|
if not alphas_cumprod.shape[0] == model_num_timesteps:
|
|
|
|
|
raise ValueError("alphas have to be defined for each timestep")
|
|
|
|
|
|
|
|
|
|
def to_torch(x):
|
|
|
|
|
return x.clone().detach().to(torch.float32).to(device)
|
|
|
|
|
|
|
|
|
|
# ddim sampling parameters
|
|
|
|
|
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(
|
|
|
|
|
alphacums=alphas_cumprod.cpu(),
|
|
|
|
|
alphacums=model_alphas_cumprod.cpu(),
|
|
|
|
|
ddim_timesteps=ddim_timesteps,
|
|
|
|
|
eta=ddim_eta,
|
|
|
|
|
)
|
|
|
|
|
self.ddim_timesteps = ddim_timesteps
|
|
|
|
|
self.alphas_cumprod = to_torch(alphas_cumprod)
|
|
|
|
|
|
|
|
|
|
# calculations for diffusion q(x_t | x_{t-1}) and others
|
|
|
|
|
self.sqrt_alphas_cumprod = to_torch(np.sqrt(alphas_cumprod.cpu()))
|
|
|
|
|
self.sqrt_alphas_cumprod = to_torch(np.sqrt(model_alphas_cumprod.cpu()))
|
|
|
|
|
self.sqrt_one_minus_alphas_cumprod = to_torch(
|
|
|
|
|
np.sqrt(1.0 - alphas_cumprod.cpu())
|
|
|
|
|
np.sqrt(1.0 - model_alphas_cumprod.cpu())
|
|
|
|
|
)
|
|
|
|
|
self.ddim_sigmas = ddim_sigmas.to(torch.float32).to(device)
|
|
|
|
|
self.ddim_alphas = ddim_alphas.to(torch.float32).to(device)
|
|
|
|
@ -70,186 +73,138 @@ class DDIMSampler:
|
|
|
|
|
|
|
|
|
|
def __init__(self, model):
|
|
|
|
|
self.model = model
|
|
|
|
|
self.device = get_device()
|
|
|
|
|
|
|
|
|
|
@torch.no_grad()
|
|
|
|
|
def sample(
|
|
|
|
|
self,
|
|
|
|
|
num_steps,
|
|
|
|
|
batch_size,
|
|
|
|
|
shape,
|
|
|
|
|
conditioning,
|
|
|
|
|
callback=None,
|
|
|
|
|
normals_sequence=None,
|
|
|
|
|
img_callback=None,
|
|
|
|
|
quantize_x0=False,
|
|
|
|
|
eta=0.0,
|
|
|
|
|
neutral_conditioning,
|
|
|
|
|
positive_conditioning,
|
|
|
|
|
guidance_scale=1.0,
|
|
|
|
|
batch_size=1,
|
|
|
|
|
mask=None,
|
|
|
|
|
x0=None,
|
|
|
|
|
orig_latent=None,
|
|
|
|
|
temperature=1.0,
|
|
|
|
|
noise_dropout=0.0,
|
|
|
|
|
x_T=None,
|
|
|
|
|
unconditional_guidance_scale=1.0,
|
|
|
|
|
unconditional_conditioning=None,
|
|
|
|
|
**kwargs,
|
|
|
|
|
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
|
|
|
|
|
initial_latent=None,
|
|
|
|
|
quantize_x0=False,
|
|
|
|
|
):
|
|
|
|
|
if conditioning.shape[0] != batch_size:
|
|
|
|
|
logger.warning(
|
|
|
|
|
f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}"
|
|
|
|
|
if positive_conditioning.shape[0] != batch_size:
|
|
|
|
|
raise ValueError(
|
|
|
|
|
f"Got {positive_conditioning.shape[0]} conditionings but batch-size is {batch_size}"
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
schedule = DDIMSchedule(
|
|
|
|
|
model_num_timesteps=self.model.num_timesteps,
|
|
|
|
|
model_alphas_cumprod=self.model.alphas_cumprod,
|
|
|
|
|
ddim_num_steps=num_steps,
|
|
|
|
|
ddim_discretize="uniform",
|
|
|
|
|
ddim_eta=0.0,
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
samples = self.ddim_sampling(
|
|
|
|
|
conditioning,
|
|
|
|
|
shape=shape,
|
|
|
|
|
schedule=schedule,
|
|
|
|
|
callback=callback,
|
|
|
|
|
img_callback=img_callback,
|
|
|
|
|
quantize_denoised=quantize_x0,
|
|
|
|
|
mask=mask,
|
|
|
|
|
x0=x0,
|
|
|
|
|
noise_dropout=noise_dropout,
|
|
|
|
|
temperature=temperature,
|
|
|
|
|
x_T=x_T,
|
|
|
|
|
unconditional_guidance_scale=unconditional_guidance_scale,
|
|
|
|
|
unconditional_conditioning=unconditional_conditioning,
|
|
|
|
|
)
|
|
|
|
|
return samples
|
|
|
|
|
if initial_latent is None:
|
|
|
|
|
initial_latent = torch.randn(shape, device="cpu").to(self.device)
|
|
|
|
|
|
|
|
|
|
@torch.no_grad()
|
|
|
|
|
def ddim_sampling(
|
|
|
|
|
self,
|
|
|
|
|
cond,
|
|
|
|
|
shape,
|
|
|
|
|
schedule,
|
|
|
|
|
x_T=None,
|
|
|
|
|
callback=None,
|
|
|
|
|
timesteps=None,
|
|
|
|
|
quantize_denoised=False,
|
|
|
|
|
mask=None,
|
|
|
|
|
x0=None,
|
|
|
|
|
img_callback=None,
|
|
|
|
|
temperature=1.0,
|
|
|
|
|
noise_dropout=0.0,
|
|
|
|
|
unconditional_guidance_scale=1.0,
|
|
|
|
|
unconditional_conditioning=None,
|
|
|
|
|
):
|
|
|
|
|
device = self.model.betas.device
|
|
|
|
|
b = shape[0]
|
|
|
|
|
if x_T is None:
|
|
|
|
|
# run on CPU for seed consistency. M1/mps runs were not consistent otherwise
|
|
|
|
|
img = torch.randn(shape, device="cpu").to(device)
|
|
|
|
|
else:
|
|
|
|
|
img = x_T
|
|
|
|
|
log_latent(img, "initial noise")
|
|
|
|
|
log_latent(initial_latent, "initial latent")
|
|
|
|
|
|
|
|
|
|
if timesteps is None:
|
|
|
|
|
timesteps = schedule.ddim_timesteps
|
|
|
|
|
else:
|
|
|
|
|
subset_end = (
|
|
|
|
|
int(
|
|
|
|
|
min(timesteps / schedule.ddim_timesteps.shape[0], 1)
|
|
|
|
|
* schedule.ddim_timesteps.shape[0]
|
|
|
|
|
)
|
|
|
|
|
- 1
|
|
|
|
|
)
|
|
|
|
|
timesteps = schedule.ddim_timesteps[:subset_end]
|
|
|
|
|
|
|
|
|
|
time_range = np.flip(timesteps)
|
|
|
|
|
total_steps = timesteps.shape[0]
|
|
|
|
|
logger.info(f"Running DDIM Sampling with {total_steps} timesteps")
|
|
|
|
|
|
|
|
|
|
iterator = tqdm(time_range, desc="DDIM Sampler", total=total_steps)
|
|
|
|
|
noisy_latent = initial_latent
|
|
|
|
|
|
|
|
|
|
for i, step in enumerate(iterator):
|
|
|
|
|
for i, step in enumerate(tqdm(time_range, total=total_steps)):
|
|
|
|
|
index = total_steps - i - 1
|
|
|
|
|
ts = torch.full((b,), step, device=device, dtype=torch.long)
|
|
|
|
|
ts = torch.full((batch_size,), step, device=self.device, dtype=torch.long)
|
|
|
|
|
|
|
|
|
|
if mask is not None:
|
|
|
|
|
assert x0 is not None
|
|
|
|
|
img_orig = self.model.q_sample(
|
|
|
|
|
x0, ts
|
|
|
|
|
) # TODO: deterministic forward pass?
|
|
|
|
|
img = img_orig * mask + (1.0 - mask) * img
|
|
|
|
|
|
|
|
|
|
img, pred_x0 = self.p_sample_ddim(
|
|
|
|
|
img,
|
|
|
|
|
cond,
|
|
|
|
|
ts,
|
|
|
|
|
assert orig_latent is not None
|
|
|
|
|
img_orig = self.model.q_sample(orig_latent, ts)
|
|
|
|
|
noisy_latent = img_orig * mask + (1.0 - mask) * noisy_latent
|
|
|
|
|
|
|
|
|
|
noisy_latent, predicted_latent = self.p_sample_ddim(
|
|
|
|
|
noisy_latent=noisy_latent,
|
|
|
|
|
neutral_conditioning=neutral_conditioning,
|
|
|
|
|
positive_conditioning=positive_conditioning,
|
|
|
|
|
guidance_scale=guidance_scale,
|
|
|
|
|
time_encoding=ts,
|
|
|
|
|
index=index,
|
|
|
|
|
schedule=schedule,
|
|
|
|
|
quantize_denoised=quantize_denoised,
|
|
|
|
|
quantize_denoised=quantize_x0,
|
|
|
|
|
temperature=temperature,
|
|
|
|
|
noise_dropout=noise_dropout,
|
|
|
|
|
unconditional_guidance_scale=unconditional_guidance_scale,
|
|
|
|
|
unconditional_conditioning=unconditional_conditioning,
|
|
|
|
|
)
|
|
|
|
|
if callback:
|
|
|
|
|
callback(i)
|
|
|
|
|
|
|
|
|
|
log_latent(img, "img")
|
|
|
|
|
log_latent(pred_x0, "pred_x0")
|
|
|
|
|
log_latent(noisy_latent, "noisy_latent")
|
|
|
|
|
log_latent(predicted_latent, "predicted_latent")
|
|
|
|
|
|
|
|
|
|
return img
|
|
|
|
|
return noisy_latent
|
|
|
|
|
|
|
|
|
|
def p_sample_ddim(
|
|
|
|
|
self,
|
|
|
|
|
x,
|
|
|
|
|
c,
|
|
|
|
|
t,
|
|
|
|
|
noisy_latent,
|
|
|
|
|
neutral_conditioning,
|
|
|
|
|
positive_conditioning,
|
|
|
|
|
guidance_scale,
|
|
|
|
|
time_encoding,
|
|
|
|
|
index,
|
|
|
|
|
schedule,
|
|
|
|
|
repeat_noise=False,
|
|
|
|
|
quantize_denoised=False,
|
|
|
|
|
temperature=1.0,
|
|
|
|
|
noise_dropout=0.0,
|
|
|
|
|
unconditional_guidance_scale=1.0,
|
|
|
|
|
unconditional_conditioning=None,
|
|
|
|
|
loss_function=None,
|
|
|
|
|
):
|
|
|
|
|
assert unconditional_guidance_scale >= 1
|
|
|
|
|
assert guidance_scale >= 1
|
|
|
|
|
noise_pred = get_noise_prediction(
|
|
|
|
|
denoise_func=self.model.apply_model,
|
|
|
|
|
noisy_latent=x,
|
|
|
|
|
time_encoding=t,
|
|
|
|
|
neutral_conditioning=unconditional_conditioning,
|
|
|
|
|
positive_conditioning=c,
|
|
|
|
|
signal_amplification=unconditional_guidance_scale,
|
|
|
|
|
noisy_latent=noisy_latent,
|
|
|
|
|
time_encoding=time_encoding,
|
|
|
|
|
neutral_conditioning=neutral_conditioning,
|
|
|
|
|
positive_conditioning=positive_conditioning,
|
|
|
|
|
signal_amplification=guidance_scale,
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
b = x.shape[0]
|
|
|
|
|
log_latent(noise_pred, "noise prediction")
|
|
|
|
|
batch_size = noisy_latent.shape[0]
|
|
|
|
|
|
|
|
|
|
# select parameters corresponding to the currently considered timestep
|
|
|
|
|
a_t = torch.full((b, 1, 1, 1), schedule.ddim_alphas[index], device=x.device)
|
|
|
|
|
a_t = torch.full(
|
|
|
|
|
(batch_size, 1, 1, 1),
|
|
|
|
|
schedule.ddim_alphas[index],
|
|
|
|
|
device=noisy_latent.device,
|
|
|
|
|
)
|
|
|
|
|
a_prev = torch.full(
|
|
|
|
|
(b, 1, 1, 1), schedule.ddim_alphas_prev[index], device=x.device
|
|
|
|
|
(batch_size, 1, 1, 1),
|
|
|
|
|
schedule.ddim_alphas_prev[index],
|
|
|
|
|
device=noisy_latent.device,
|
|
|
|
|
)
|
|
|
|
|
sigma_t = torch.full(
|
|
|
|
|
(batch_size, 1, 1, 1),
|
|
|
|
|
schedule.ddim_sigmas[index],
|
|
|
|
|
device=noisy_latent.device,
|
|
|
|
|
)
|
|
|
|
|
sigma_t = torch.full((b, 1, 1, 1), schedule.ddim_sigmas[index], device=x.device)
|
|
|
|
|
sqrt_one_minus_at = torch.full(
|
|
|
|
|
(b, 1, 1, 1), schedule.ddim_sqrt_one_minus_alphas[index], device=x.device
|
|
|
|
|
(batch_size, 1, 1, 1),
|
|
|
|
|
schedule.ddim_sqrt_one_minus_alphas[index],
|
|
|
|
|
device=noisy_latent.device,
|
|
|
|
|
)
|
|
|
|
|
return self._p_sample_ddim_formula(
|
|
|
|
|
x,
|
|
|
|
|
noise_pred,
|
|
|
|
|
sqrt_one_minus_at,
|
|
|
|
|
a_t,
|
|
|
|
|
sigma_t,
|
|
|
|
|
a_prev,
|
|
|
|
|
noise_dropout,
|
|
|
|
|
repeat_noise,
|
|
|
|
|
temperature,
|
|
|
|
|
noisy_latent, predicted_latent = self._p_sample_ddim_formula(
|
|
|
|
|
noisy_latent=noisy_latent,
|
|
|
|
|
noise_pred=noise_pred,
|
|
|
|
|
sqrt_one_minus_at=sqrt_one_minus_at,
|
|
|
|
|
a_t=a_t,
|
|
|
|
|
sigma_t=sigma_t,
|
|
|
|
|
a_prev=a_prev,
|
|
|
|
|
noise_dropout=noise_dropout,
|
|
|
|
|
repeat_noise=repeat_noise,
|
|
|
|
|
temperature=temperature,
|
|
|
|
|
)
|
|
|
|
|
return noisy_latent, predicted_latent
|
|
|
|
|
|
|
|
|
|
@staticmethod
|
|
|
|
|
def _p_sample_ddim_formula(
|
|
|
|
|
x,
|
|
|
|
|
noisy_latent,
|
|
|
|
|
noise_pred,
|
|
|
|
|
sqrt_one_minus_at,
|
|
|
|
|
a_t,
|
|
|
|
@ -259,15 +214,18 @@ class DDIMSampler:
|
|
|
|
|
repeat_noise,
|
|
|
|
|
temperature,
|
|
|
|
|
):
|
|
|
|
|
# current prediction for x_0
|
|
|
|
|
pred_x0 = (x - sqrt_one_minus_at * noise_pred) / a_t.sqrt()
|
|
|
|
|
predicted_latent = (noisy_latent - sqrt_one_minus_at * noise_pred) / a_t.sqrt()
|
|
|
|
|
# direction pointing to x_t
|
|
|
|
|
dir_xt = (1.0 - a_prev - sigma_t**2).sqrt() * noise_pred
|
|
|
|
|
noise = sigma_t * noise_like(x.shape, x.device, repeat_noise) * temperature
|
|
|
|
|
noise = (
|
|
|
|
|
sigma_t
|
|
|
|
|
* noise_like(noisy_latent.shape, noisy_latent.device, repeat_noise)
|
|
|
|
|
* temperature
|
|
|
|
|
)
|
|
|
|
|
if noise_dropout > 0.0:
|
|
|
|
|
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
|
|
|
|
x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
|
|
|
|
|
return x_prev, pred_x0
|
|
|
|
|
x_prev = a_prev.sqrt() * predicted_latent + dir_xt + noise
|
|
|
|
|
return x_prev, predicted_latent
|
|
|
|
|
|
|
|
|
|
@torch.no_grad()
|
|
|
|
|
def noise_an_image(self, init_latent, t, schedule, noise=None):
|
|
|
|
@ -288,12 +246,11 @@ class DDIMSampler:
|
|
|
|
|
def decode(
|
|
|
|
|
self,
|
|
|
|
|
initial_latent,
|
|
|
|
|
cond,
|
|
|
|
|
neutral_conditioning,
|
|
|
|
|
positive_conditioning,
|
|
|
|
|
guidance_scale,
|
|
|
|
|
t_start,
|
|
|
|
|
schedule,
|
|
|
|
|
unconditional_guidance_scale=1.0,
|
|
|
|
|
unconditional_conditioning=None,
|
|
|
|
|
img_callback=None,
|
|
|
|
|
temperature=1.0,
|
|
|
|
|
mask=None,
|
|
|
|
|
orig_latent=None,
|
|
|
|
@ -303,12 +260,10 @@ class DDIMSampler:
|
|
|
|
|
|
|
|
|
|
time_range = np.flip(timesteps)
|
|
|
|
|
total_steps = timesteps.shape[0]
|
|
|
|
|
logger.debug(f"Running DDIM Sampling with {total_steps} timesteps")
|
|
|
|
|
|
|
|
|
|
iterator = tqdm(time_range, desc="Decoding image", total=total_steps)
|
|
|
|
|
x_dec = initial_latent
|
|
|
|
|
noisy_latent = initial_latent
|
|
|
|
|
|
|
|
|
|
for i, step in enumerate(iterator):
|
|
|
|
|
for i, step in enumerate(tqdm(time_range, total=total_steps)):
|
|
|
|
|
index = total_steps - i - 1
|
|
|
|
|
ts = torch.full(
|
|
|
|
|
(initial_latent.shape[0],),
|
|
|
|
@ -329,20 +284,20 @@ class DDIMSampler:
|
|
|
|
|
)
|
|
|
|
|
else:
|
|
|
|
|
xdec_orig_with_hints = xdec_orig
|
|
|
|
|
x_dec = xdec_orig_with_hints * mask + (1.0 - mask) * x_dec
|
|
|
|
|
log_latent(x_dec, "x_dec")
|
|
|
|
|
noisy_latent = xdec_orig_with_hints * mask + (1.0 - mask) * noisy_latent
|
|
|
|
|
log_latent(noisy_latent, "noisy_latent")
|
|
|
|
|
|
|
|
|
|
x_dec, pred_x0 = self.p_sample_ddim(
|
|
|
|
|
x_dec,
|
|
|
|
|
cond,
|
|
|
|
|
ts,
|
|
|
|
|
noisy_latent, predicted_latent = self.p_sample_ddim(
|
|
|
|
|
noisy_latent=noisy_latent,
|
|
|
|
|
positive_conditioning=positive_conditioning,
|
|
|
|
|
time_encoding=ts,
|
|
|
|
|
schedule=schedule,
|
|
|
|
|
index=index,
|
|
|
|
|
unconditional_guidance_scale=unconditional_guidance_scale,
|
|
|
|
|
unconditional_conditioning=unconditional_conditioning,
|
|
|
|
|
guidance_scale=guidance_scale,
|
|
|
|
|
neutral_conditioning=neutral_conditioning,
|
|
|
|
|
temperature=temperature,
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
log_latent(x_dec, f"x_dec {i}")
|
|
|
|
|
log_latent(pred_x0, f"pred_x0 {i}")
|
|
|
|
|
return x_dec
|
|
|
|
|
log_latent(noisy_latent, f"noisy_latent {i}")
|
|
|
|
|
log_latent(predicted_latent, f"predicted_latent {i}")
|
|
|
|
|
return noisy_latent
|
|
|
|
|