imaginAIry/imaginairy/samplers/plms.py
Bryce 95a8fa31a9 fix: inpainting producing blurry images
while the previous version did produce much better blending it also makes images that lack detail for some reason.

tests: Added more tests to help catch this sort of thing earlies

fix: found that median blur is really slow, so I made sure we only do it on downsampled masks.  Was taking like 3 minutes to run on the large pearl girl picture on M1

- docs: update examples
2022-09-27 17:19:25 -07:00

468 lines
16 KiB
Python

# pylama:ignore=W0613
"""SAMPLING ONLY."""
import logging
import numpy as np
import torch
from tqdm import tqdm
from imaginairy.img_log import log_latent
from imaginairy.modules.diffusion.util import (
extract_into_tensor,
make_ddim_sampling_parameters,
make_ddim_timesteps,
noise_like,
)
from imaginairy.utils import get_device
logger = logging.getLogger(__name__)
class PLMSSampler:
"""probabilistic least-mean-squares"""
def __init__(self, model):
self.model = model
self.ddpm_num_timesteps = model.num_timesteps
self.device_available = get_device()
self.ddim_timesteps = None
def register_buffer(self, name, attr):
if isinstance(attr, torch.Tensor):
if attr.device != torch.device(self.device_available):
attr = attr.to(torch.float32).to(torch.device(self.device_available))
setattr(self, name, attr)
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0.0):
if ddim_eta != 0:
raise ValueError("ddim_eta must be 0 for PLMS")
self.ddim_timesteps = make_ddim_timesteps(
ddim_discr_method=ddim_discretize,
num_ddim_timesteps=ddim_num_steps,
num_ddpm_timesteps=self.ddpm_num_timesteps,
)
alphas_cumprod = self.model.alphas_cumprod
assert (
alphas_cumprod.shape[0] == self.ddpm_num_timesteps
), "alphas have to be defined for each timestep"
def to_torch(x):
return x.clone().detach().to(torch.float32).to(self.model.device)
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
self.register_buffer(
"sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod.cpu()))
)
self.register_buffer(
"sqrt_one_minus_alphas_cumprod",
to_torch(np.sqrt(1.0 - alphas_cumprod.cpu())),
)
self.register_buffer(
"log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod.cpu()))
)
self.register_buffer(
"sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod.cpu()))
)
self.register_buffer(
"sqrt_recipm1_alphas_cumprod",
to_torch(np.sqrt(1.0 / alphas_cumprod.cpu() - 1)),
)
# 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,
)
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", np.sqrt(1.0 - ddim_alphas))
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
)
@torch.no_grad()
def sample(
self,
num_steps,
batch_size,
shape,
conditioning=None,
callback=None,
normals_sequence=None,
img_callback=None,
quantize_x0=False,
eta=0.0,
mask=None,
x0=None,
temperature=1.0,
noise_dropout=0.0,
score_corrector=None,
corrector_kwargs=None,
x_T=None,
unconditional_guidance_scale=1.0,
unconditional_conditioning=None,
# this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
**kwargs,
):
if conditioning is not None:
if isinstance(conditioning, dict):
cbs = conditioning[list(conditioning.keys())[0]].shape[0]
if cbs != batch_size:
logger.warning(
f"Warning: Got {cbs} conditionings but batch-size is {batch_size}"
)
else:
if conditioning.shape[0] != batch_size:
logger.warning(
f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}"
)
self.make_schedule(ddim_num_steps=num_steps, ddim_eta=eta)
samples = self.plms_sampling(
conditioning,
(batch_size, *shape),
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,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning,
)
return samples
@torch.no_grad()
def plms_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,
temperature=1.0,
noise_dropout=0.0,
score_corrector=None,
corrector_kwargs=None,
unconditional_guidance_scale=1.0,
unconditional_conditioning=None,
):
device = self.model.betas.device
b = shape[0]
if x_T is None:
img = torch.randn(shape, device="cpu").to(device)
else:
img = x_T
log_latent(img, "initial img")
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]
time_range = (
list(reversed(range(0, timesteps)))
if ddim_use_original_steps
else np.flip(timesteps)
)
total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
logger.debug(f"Running PLMS Sampling with {total_steps} timesteps")
iterator = tqdm(time_range, desc=" PLMS Sampler", total=total_steps)
old_eps = []
for i, step in enumerate(iterator):
index = total_steps - i - 1
ts = torch.full((b,), step, device=device, dtype=torch.long)
ts_next = torch.full(
(b,),
time_range[min(i + 1, len(time_range) - 1)],
device=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, e_t = self.p_sample_plms(
img,
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,
old_eps=old_eps,
t_next=ts_next,
)
old_eps.append(e_t)
if len(old_eps) >= 4:
old_eps.pop(0)
if callback:
callback(i)
if img_callback:
img_callback(img, "img")
img_callback(pred_x0, "pred_x0")
return img
@torch.no_grad()
def p_sample_plms(
self,
x,
c,
t,
index,
repeat_noise=False,
use_original_steps=False,
quantize_denoised=False,
temperature=1.0,
noise_dropout=0.0,
score_corrector=None,
corrector_kwargs=None,
unconditional_guidance_scale=1.0,
unconditional_conditioning=None,
old_eps=None,
t_next=None,
):
b, *_, device = *x.shape, x.device
def get_model_output(x, t):
if (
unconditional_conditioning is None
or unconditional_guidance_scale == 1.0
):
e_t = self.model.apply_model(x, t, c)
else:
x_in = torch.cat([x] * 2)
t_in = torch.cat([t] * 2)
c_in = torch.cat([unconditional_conditioning, c])
e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
log_latent(e_t_uncond, "noise pred uncond")
log_latent(e_t, "noise pred cond")
e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
log_latent(e_t, "noise pred combined")
if score_corrector is not None:
assert self.model.parameterization == "eps"
e_t = score_corrector.modify_score(
self.model, e_t, x, t, c, **corrector_kwargs
)
return e_t
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.model.ddim_sigmas_for_original_num_steps
if use_original_steps
else self.ddim_sigmas
)
def get_x_prev_and_pred_x0(e_t, index):
# select parameters corresponding to the currently considered timestep
a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
sqrt_one_minus_at = torch.full(
(b, 1, 1, 1), sqrt_one_minus_alphas[index], device=device
)
# current prediction for x_0
pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
if quantize_denoised:
pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
# direction pointing to x_t
dir_xt = (1.0 - a_prev - sigma_t**2).sqrt() * e_t
noise = sigma_t * noise_like(x.shape, 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
e_t = get_model_output(x, t)
if len(old_eps) == 0:
# Pseudo Improved Euler (2nd order)
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
e_t_next = get_model_output(x_prev, t_next)
e_t_prime = (e_t + e_t_next) / 2
elif len(old_eps) == 1:
# 2nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (3 * e_t - old_eps[-1]) / 2
elif len(old_eps) == 2:
# 3nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
elif len(old_eps) >= 3:
# 4nd order Pseudo Linear Multistep (Adams-Bashforth)
e_t_prime = (
55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]
) / 24
x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
log_latent(x_prev, "x_prev")
log_latent(pred_x0, "pred_x0")
return x_prev, pred_x0, e_t
@torch.no_grad()
def stochastic_encode(self, init_latent, t, noise=None):
# fast, but does not allow for exact reconstruction
# t serves as an index to gather the correct alphas
t = t.clamp(0, 1000)
sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
if noise is None:
noise = torch.randn_like(init_latent, device="cpu").to(get_device())
return (
extract_into_tensor(sqrt_alphas_cumprod, t, init_latent.shape) * init_latent
+ extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, init_latent.shape)
* noise
)
@torch.no_grad()
def decode(
self,
x_latent,
cond,
t_start,
unconditional_guidance_scale=1.0,
unconditional_conditioning=None,
img_callback=None,
score_corrector=None,
temperature=1.0,
mask=None,
orig_latent=None,
noise=None,
):
timesteps = self.ddim_timesteps[:t_start]
time_range = np.flip(timesteps)
total_steps = timesteps.shape[0]
iterator = tqdm(time_range, desc="PLMS altering image", total=total_steps)
x_dec = x_latent
old_eps = []
log_latent(x_dec, "x_dec")
# not sure what the downside of using the same noise throughout the process would be...
# seems to work fine. maybe it runs faster?
noise = (
torch.randn_like(x_dec, device="cpu").to(x_dec.device)
if noise is None
else noise
)
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
)
ts_next = torch.full(
(x_latent.shape[0],),
time_range[min(i + 1, len(time_range) - 1)],
device=x_latent.device,
dtype=torch.long,
)
if mask is not None:
assert orig_latent is not None
xdec_orig = self.model.q_sample(orig_latent, ts, noise)
log_latent(xdec_orig, f"xdec_orig i={i} index-{index}")
# this helps prevent the weird disjointed images that can happen with masking
hint_strength = 0.8
if i < 2:
xdec_orig_with_hints = (
xdec_orig * (1 - hint_strength) + orig_latent * hint_strength
)
else:
xdec_orig_with_hints = xdec_orig
x_dec = xdec_orig_with_hints * mask + (1.0 - mask) * x_dec
log_latent(x_dec, f"x_dec {ts}")
x_dec, pred_x0, e_t = self.p_sample_plms(
x_dec,
cond,
ts,
index=index,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=unconditional_conditioning,
temperature=temperature,
old_eps=old_eps,
t_next=ts_next,
)
# original_loss = ((x_dec - x_latent).abs().mean()*70)
# sigma_t = torch.full((1, 1, 1, 1), self.ddim_sigmas[index], device=get_device())
# x_dec = x_dec.detach() + (original_loss * 0.1) ** 2
# cond_grad = -torch.autograd.grad(original_loss, x_dec)[0]
# x_dec = x_dec.detach() + cond_grad * sigma_t ** 2
# x_dec_alt = x_dec + (original_loss * 0.1) ** 2
old_eps.append(e_t)
if len(old_eps) >= 4:
old_eps.pop(0)
if img_callback:
img_callback(x_dec, "x_dec")
img_callback(pred_x0, "pred_x0")
log_latent(x_dec, f"x_dec {i}")
log_latent(x_dec, f"e_t {i}")
log_latent(pred_x0, f"pred_x0 {i}")
return x_dec