490 lines
20 KiB
Python
490 lines
20 KiB
Python
import numpy as np
|
|
import torch
|
|
import copy
|
|
|
|
from unifolm_wma.utils.diffusion import make_ddim_sampling_parameters, make_ddim_timesteps, rescale_noise_cfg
|
|
from unifolm_wma.utils.common import noise_like
|
|
from unifolm_wma.utils.common import extract_into_tensor
|
|
from tqdm import tqdm
|
|
|
|
|
|
class DDIMSampler(object):
|
|
|
|
def __init__(self, model, schedule="linear", **kwargs):
|
|
super().__init__()
|
|
self.model = model
|
|
self.ddpm_num_timesteps = model.num_timesteps
|
|
self.schedule = schedule
|
|
self.counter = 0
|
|
|
|
def register_buffer(self, name, attr):
|
|
if type(attr) == torch.Tensor:
|
|
if attr.device != torch.device("cuda"):
|
|
attr = attr.to(torch.device("cuda"))
|
|
setattr(self, name, attr)
|
|
|
|
def make_schedule(self,
|
|
ddim_num_steps,
|
|
ddim_discretize="uniform",
|
|
ddim_eta=0.,
|
|
verbose=True):
|
|
device = self.model.betas.device
|
|
cache_key = (ddim_num_steps, ddim_discretize, float(ddim_eta),
|
|
str(device))
|
|
if getattr(self, "_schedule_cache", None) == cache_key:
|
|
return
|
|
self.ddim_timesteps = make_ddim_timesteps(
|
|
ddim_discr_method=ddim_discretize,
|
|
num_ddim_timesteps=ddim_num_steps,
|
|
num_ddpm_timesteps=self.ddpm_num_timesteps,
|
|
verbose=verbose)
|
|
alphas_cumprod = self.model.alphas_cumprod
|
|
assert alphas_cumprod.shape[
|
|
0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
|
|
to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model
|
|
.device)
|
|
|
|
if self.model.use_dynamic_rescale:
|
|
self.ddim_scale_arr = self.model.scale_arr[self.ddim_timesteps]
|
|
self.ddim_scale_arr_prev = torch.cat(
|
|
[self.ddim_scale_arr[0:1], self.ddim_scale_arr[:-1]])
|
|
|
|
self.register_buffer('betas', to_torch(self.model.betas))
|
|
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
|
self.register_buffer('alphas_cumprod_prev',
|
|
to_torch(self.model.alphas_cumprod_prev))
|
|
|
|
# Calculations for diffusion q(x_t | x_{t-1}) and others
|
|
# Computed directly on GPU to avoid CPU↔GPU transfers
|
|
ac = to_torch(alphas_cumprod)
|
|
self.register_buffer('sqrt_alphas_cumprod', ac.sqrt())
|
|
self.register_buffer('sqrt_one_minus_alphas_cumprod', (1. - ac).sqrt())
|
|
self.register_buffer('log_one_minus_alphas_cumprod', (1. - ac).log())
|
|
self.register_buffer('sqrt_recip_alphas_cumprod', ac.rsqrt())
|
|
self.register_buffer('sqrt_recipm1_alphas_cumprod', (1. / ac - 1).sqrt())
|
|
|
|
# DDIM sampling parameters
|
|
ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(
|
|
alphacums=alphas_cumprod.cpu(),
|
|
ddim_timesteps=self.ddim_timesteps,
|
|
eta=ddim_eta,
|
|
verbose=verbose)
|
|
ddim_sigmas = torch.as_tensor(ddim_sigmas,
|
|
device=self.model.device,
|
|
dtype=torch.float32)
|
|
ddim_alphas = torch.as_tensor(ddim_alphas,
|
|
device=self.model.device,
|
|
dtype=torch.float32)
|
|
ddim_alphas_prev = torch.as_tensor(ddim_alphas_prev,
|
|
device=self.model.device,
|
|
dtype=torch.float32)
|
|
self.register_buffer('ddim_sigmas', ddim_sigmas)
|
|
self.register_buffer('ddim_alphas', ddim_alphas)
|
|
self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
|
|
self.register_buffer('ddim_sqrt_one_minus_alphas',
|
|
torch.sqrt(1. - ddim_alphas))
|
|
# Precomputed coefficients for DDIM update formula
|
|
self.register_buffer('ddim_sqrt_alphas', ddim_alphas.sqrt())
|
|
self.register_buffer('ddim_sqrt_alphas_prev', ddim_alphas_prev.sqrt())
|
|
self.register_buffer('ddim_dir_coeff',
|
|
(1. - ddim_alphas_prev - ddim_sigmas**2).sqrt())
|
|
sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
|
|
(1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) *
|
|
(1 - self.alphas_cumprod / self.alphas_cumprod_prev))
|
|
self.register_buffer('ddim_sigmas_for_original_num_steps',
|
|
sigmas_for_original_sampling_steps)
|
|
self._schedule_cache = cache_key
|
|
|
|
@torch.no_grad()
|
|
def sample(
|
|
self,
|
|
S,
|
|
batch_size,
|
|
shape,
|
|
conditioning=None,
|
|
callback=None,
|
|
normals_sequence=None,
|
|
img_callback=None,
|
|
quantize_x0=False,
|
|
eta=0.,
|
|
mask=None,
|
|
x0=None,
|
|
temperature=1.,
|
|
noise_dropout=0.,
|
|
score_corrector=None,
|
|
corrector_kwargs=None,
|
|
verbose=True,
|
|
schedule_verbose=False,
|
|
x_T=None,
|
|
log_every_t=100,
|
|
unconditional_guidance_scale=1.,
|
|
unconditional_conditioning=None,
|
|
precision=None,
|
|
fs=None,
|
|
timestep_spacing='uniform', #uniform_trailing for starting from last timestep
|
|
guidance_rescale=0.0,
|
|
**kwargs):
|
|
|
|
# Check condition bs
|
|
if conditioning is not None:
|
|
if isinstance(conditioning, dict):
|
|
try:
|
|
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
|
|
except:
|
|
cbs = conditioning[list(
|
|
conditioning.keys())[0]][0].shape[0]
|
|
|
|
if cbs != batch_size:
|
|
print(
|
|
f"Warning: Got {cbs} conditionings but batch-size is {batch_size}"
|
|
)
|
|
else:
|
|
if conditioning.shape[0] != batch_size:
|
|
print(
|
|
f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}"
|
|
)
|
|
|
|
self.make_schedule(ddim_num_steps=S,
|
|
ddim_discretize=timestep_spacing,
|
|
ddim_eta=eta,
|
|
verbose=schedule_verbose)
|
|
|
|
# Make shape
|
|
if len(shape) == 3:
|
|
C, H, W = shape
|
|
size = (batch_size, C, H, W)
|
|
elif len(shape) == 4:
|
|
C, T, H, W = shape
|
|
size = (batch_size, C, T, H, W)
|
|
|
|
samples, actions, states, intermediates = self.ddim_sampling(
|
|
conditioning,
|
|
size,
|
|
callback=callback,
|
|
img_callback=img_callback,
|
|
quantize_denoised=quantize_x0,
|
|
mask=mask,
|
|
x0=x0,
|
|
ddim_use_original_steps=False,
|
|
noise_dropout=noise_dropout,
|
|
temperature=temperature,
|
|
score_corrector=score_corrector,
|
|
corrector_kwargs=corrector_kwargs,
|
|
x_T=x_T,
|
|
log_every_t=log_every_t,
|
|
unconditional_guidance_scale=unconditional_guidance_scale,
|
|
unconditional_conditioning=unconditional_conditioning,
|
|
verbose=verbose,
|
|
precision=precision,
|
|
fs=fs,
|
|
guidance_rescale=guidance_rescale,
|
|
**kwargs)
|
|
return samples, actions, states, intermediates
|
|
|
|
@torch.no_grad()
|
|
def ddim_sampling(self,
|
|
cond,
|
|
shape,
|
|
x_T=None,
|
|
ddim_use_original_steps=False,
|
|
callback=None,
|
|
timesteps=None,
|
|
quantize_denoised=False,
|
|
mask=None,
|
|
x0=None,
|
|
img_callback=None,
|
|
log_every_t=100,
|
|
temperature=1.,
|
|
noise_dropout=0.,
|
|
score_corrector=None,
|
|
corrector_kwargs=None,
|
|
unconditional_guidance_scale=1.,
|
|
unconditional_conditioning=None,
|
|
verbose=True,
|
|
precision=None,
|
|
fs=None,
|
|
guidance_rescale=0.0,
|
|
**kwargs):
|
|
device = self.model.betas.device
|
|
dp_ddim_scheduler_action = self.model.dp_noise_scheduler_action
|
|
dp_ddim_scheduler_state = self.model.dp_noise_scheduler_state
|
|
|
|
b = shape[0]
|
|
action = torch.randn((b, 16, self.model.agent_action_dim),
|
|
device=device)
|
|
state = torch.randn((b, 16, self.model.agent_state_dim),
|
|
device=device)
|
|
img = torch.randn(shape, device=device) if x_T is None else x_T
|
|
|
|
if precision is not None:
|
|
if precision == 16:
|
|
img = img.to(dtype=torch.float16)
|
|
action = action.to(dtype=torch.float16)
|
|
state = state.to(dtype=torch.float16)
|
|
|
|
if timesteps is None:
|
|
timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
|
|
elif timesteps is not None and not ddim_use_original_steps:
|
|
subset_end = int(
|
|
min(timesteps / self.ddim_timesteps.shape[0], 1) *
|
|
self.ddim_timesteps.shape[0]) - 1
|
|
timesteps = self.ddim_timesteps[:subset_end]
|
|
|
|
intermediates = {
|
|
'x_inter': [img],
|
|
'pred_x0': [img],
|
|
'x_inter_action': [action],
|
|
'pred_x0_action': [action],
|
|
'x_inter_state': [state],
|
|
'pred_x0_state': [state],
|
|
}
|
|
if ddim_use_original_steps:
|
|
time_range = np.arange(timesteps - 1, -1, -1)
|
|
else:
|
|
time_range = np.flip(timesteps)
|
|
time_range = np.ascontiguousarray(time_range)
|
|
total_steps = int(time_range.shape[0])
|
|
t_seq = torch.as_tensor(time_range, device=device, dtype=torch.long)
|
|
ts_batch = t_seq.unsqueeze(1).expand(total_steps, b).contiguous()
|
|
if verbose:
|
|
iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
|
|
else:
|
|
iterator = time_range
|
|
|
|
clean_cond = kwargs.pop("clean_cond", False)
|
|
|
|
dp_ddim_scheduler_action.set_timesteps(len(timesteps))
|
|
dp_ddim_scheduler_state.set_timesteps(len(timesteps))
|
|
for i, step in enumerate(iterator):
|
|
index = total_steps - i - 1
|
|
ts = ts_batch[i]
|
|
|
|
# Use mask to blend noised original latent (img_orig) & new sampled latent (img)
|
|
if mask is not None:
|
|
assert x0 is not None
|
|
if clean_cond:
|
|
img_orig = x0
|
|
else:
|
|
img_orig = self.model.q_sample(x0, ts)
|
|
img = img_orig * mask + (1. - mask) * img
|
|
|
|
outs = self.p_sample_ddim(
|
|
img,
|
|
action,
|
|
state,
|
|
cond,
|
|
ts,
|
|
index=index,
|
|
use_original_steps=ddim_use_original_steps,
|
|
quantize_denoised=quantize_denoised,
|
|
temperature=temperature,
|
|
noise_dropout=noise_dropout,
|
|
score_corrector=score_corrector,
|
|
corrector_kwargs=corrector_kwargs,
|
|
unconditional_guidance_scale=unconditional_guidance_scale,
|
|
unconditional_conditioning=unconditional_conditioning,
|
|
mask=mask,
|
|
x0=x0,
|
|
fs=fs,
|
|
guidance_rescale=guidance_rescale,
|
|
**kwargs)
|
|
|
|
img, pred_x0, model_output_action, model_output_state = outs
|
|
|
|
action = dp_ddim_scheduler_action.step(
|
|
model_output_action,
|
|
step,
|
|
action,
|
|
generator=None,
|
|
).prev_sample
|
|
state = dp_ddim_scheduler_state.step(
|
|
model_output_state,
|
|
step,
|
|
state,
|
|
generator=None,
|
|
).prev_sample
|
|
|
|
if callback: callback(i)
|
|
if img_callback: img_callback(pred_x0, i)
|
|
|
|
if index % log_every_t == 0 or index == total_steps - 1:
|
|
intermediates['x_inter'].append(img)
|
|
intermediates['pred_x0'].append(pred_x0)
|
|
intermediates['x_inter_action'].append(action)
|
|
intermediates['x_inter_state'].append(state)
|
|
|
|
return img, action, state, intermediates
|
|
|
|
@torch.no_grad()
|
|
def p_sample_ddim(self,
|
|
x,
|
|
x_action,
|
|
x_state,
|
|
c,
|
|
t,
|
|
index,
|
|
repeat_noise=False,
|
|
use_original_steps=False,
|
|
quantize_denoised=False,
|
|
temperature=1.,
|
|
noise_dropout=0.,
|
|
score_corrector=None,
|
|
corrector_kwargs=None,
|
|
unconditional_guidance_scale=1.,
|
|
unconditional_conditioning=None,
|
|
uc_type=None,
|
|
conditional_guidance_scale_temporal=None,
|
|
mask=None,
|
|
x0=None,
|
|
guidance_rescale=0.0,
|
|
**kwargs):
|
|
b, *_, device = *x.shape, x.device
|
|
if x.dim() == 5:
|
|
is_video = True
|
|
else:
|
|
is_video = False
|
|
|
|
if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
|
|
model_output, model_output_action, model_output_state = self.model.apply_model(
|
|
x, x_action, x_state, t, c, **kwargs) # unet denoiser
|
|
else:
|
|
# do_classifier_free_guidance
|
|
if isinstance(c, torch.Tensor) or isinstance(c, dict):
|
|
e_t_cond, e_t_cond_action, e_t_cond_state = self.model.apply_model(
|
|
x, x_action, x_state, t, c, **kwargs)
|
|
e_t_uncond, e_t_uncond_action, e_t_uncond_state = self.model.apply_model(
|
|
x, x_action, x_state, t, unconditional_conditioning,
|
|
**kwargs)
|
|
else:
|
|
raise NotImplementedError
|
|
model_output = torch.lerp(e_t_uncond, e_t_cond,
|
|
unconditional_guidance_scale)
|
|
model_output_action = torch.lerp(e_t_uncond_action,
|
|
e_t_cond_action,
|
|
unconditional_guidance_scale)
|
|
model_output_state = torch.lerp(e_t_uncond_state, e_t_cond_state,
|
|
unconditional_guidance_scale)
|
|
|
|
if guidance_rescale > 0.0:
|
|
model_output = rescale_noise_cfg(
|
|
model_output, e_t_cond, guidance_rescale=guidance_rescale)
|
|
model_output_action = rescale_noise_cfg(
|
|
model_output_action,
|
|
e_t_cond_action,
|
|
guidance_rescale=guidance_rescale)
|
|
model_output_state = rescale_noise_cfg(
|
|
model_output_state,
|
|
e_t_cond_state,
|
|
guidance_rescale=guidance_rescale)
|
|
|
|
if self.model.parameterization == "v":
|
|
e_t = self.model.predict_eps_from_z_and_v(x, t, model_output)
|
|
else:
|
|
e_t = model_output
|
|
|
|
if score_corrector is not None:
|
|
assert self.model.parameterization == "eps", 'not implemented'
|
|
e_t = score_corrector.modify_score(self.model, e_t, x, t, c,
|
|
**corrector_kwargs)
|
|
|
|
alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
|
|
alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
|
|
sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
|
|
sigmas = self.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
|
|
|
|
if use_original_steps:
|
|
sqrt_alphas = alphas.sqrt()
|
|
sqrt_alphas_prev = alphas_prev.sqrt()
|
|
dir_coeffs = (1. - alphas_prev - sigmas**2).sqrt()
|
|
else:
|
|
sqrt_alphas = self.ddim_sqrt_alphas
|
|
sqrt_alphas_prev = self.ddim_sqrt_alphas_prev
|
|
dir_coeffs = self.ddim_dir_coeff
|
|
|
|
if is_video:
|
|
size = (1, 1, 1, 1, 1)
|
|
else:
|
|
size = (1, 1, 1, 1)
|
|
|
|
sqrt_at = sqrt_alphas[index].view(size)
|
|
sqrt_a_prev = sqrt_alphas_prev[index].view(size)
|
|
sigma_t = sigmas[index].view(size)
|
|
sqrt_one_minus_at = sqrt_one_minus_alphas[index].view(size)
|
|
dir_coeff = dir_coeffs[index].view(size)
|
|
|
|
if self.model.parameterization != "v":
|
|
pred_x0 = (x - sqrt_one_minus_at * e_t) / sqrt_at
|
|
else:
|
|
pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
|
|
|
|
if self.model.use_dynamic_rescale:
|
|
scale_t = self.ddim_scale_arr[index].view(size)
|
|
prev_scale_t = self.ddim_scale_arr_prev[index].view(size)
|
|
rescale = (prev_scale_t / scale_t)
|
|
pred_x0 *= rescale
|
|
|
|
if quantize_denoised:
|
|
pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
|
|
|
|
noise = sigma_t * noise_like(x.shape, device,
|
|
repeat_noise) * temperature
|
|
if noise_dropout > 0.:
|
|
noise = torch.nn.functional.dropout(noise, p=noise_dropout)
|
|
x_prev = sqrt_a_prev * pred_x0 + dir_coeff * e_t + noise
|
|
|
|
return x_prev, pred_x0, model_output_action, model_output_state
|
|
|
|
@torch.no_grad()
|
|
def decode(self,
|
|
x_latent,
|
|
cond,
|
|
t_start,
|
|
unconditional_guidance_scale=1.0,
|
|
unconditional_conditioning=None,
|
|
use_original_steps=False,
|
|
callback=None):
|
|
|
|
timesteps = np.arange(self.ddpm_num_timesteps
|
|
) if use_original_steps else self.ddim_timesteps
|
|
timesteps = timesteps[:t_start]
|
|
|
|
time_range = np.flip(timesteps)
|
|
total_steps = timesteps.shape[0]
|
|
print(f"Running DDIM Sampling with {total_steps} timesteps")
|
|
|
|
iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
|
|
x_dec = x_latent
|
|
for i, step in enumerate(iterator):
|
|
index = total_steps - i - 1
|
|
ts = torch.full((x_latent.shape[0], ),
|
|
step,
|
|
device=x_latent.device,
|
|
dtype=torch.long)
|
|
x_dec, _ = self.p_sample_ddim(
|
|
x_dec,
|
|
cond,
|
|
ts,
|
|
index=index,
|
|
use_original_steps=use_original_steps,
|
|
unconditional_guidance_scale=unconditional_guidance_scale,
|
|
unconditional_conditioning=unconditional_conditioning)
|
|
if callback: callback(i)
|
|
return x_dec
|
|
|
|
@torch.no_grad()
|
|
def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
|
|
# fast, but does not allow for exact reconstruction
|
|
if use_original_steps:
|
|
sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
|
|
sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
|
|
else:
|
|
sqrt_alphas_cumprod = self.ddim_sqrt_alphas
|
|
sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
|
|
|
|
if noise is None:
|
|
noise = torch.randn_like(x0)
|
|
return (
|
|
extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
|
|
extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) *
|
|
noise)
|