From 9ba302a5f4f0ac7e9d33d9c1a7d0989fb9640139 Mon Sep 17 00:00:00 2001 From: Bryce Date: Wed, 5 Oct 2022 21:43:00 -0700 Subject: [PATCH] refactor: begin to standardize samplers --- imaginairy/api.py | 20 +- imaginairy/enhancers/clip_masking.py | 2 +- imaginairy/samplers/base.py | 96 ++++--- imaginairy/samplers/ddim.py | 196 ++++++-------- imaginairy/samplers/kdiff.py | 12 +- imaginairy/samplers/plms.py | 390 +++++++++++---------------- tests/conftest.py | 5 + tests/test_experiments.py | 5 +- tests/test_utils.py | 8 +- 9 files changed, 317 insertions(+), 417 deletions(-) diff --git a/imaginairy/api.py b/imaginairy/api.py index 5e413ca..2a8059f 100755 --- a/imaginairy/api.py +++ b/imaginairy/api.py @@ -26,6 +26,7 @@ from imaginairy.img_log import ( from imaginairy.img_utils import pillow_fit_image_within, pillow_img_to_torch_image from imaginairy.safety import is_nsfw from imaginairy.samplers.base import get_sampler +from imaginairy.samplers.plms import PLMSSchedule from imaginairy.schema import ImaginePrompt, ImagineResult from imaginairy.utils import ( fix_torch_group_norm, @@ -208,6 +209,7 @@ def imagine( log_conditioning(c, "positive conditioning") shape = [ + 1, latent_channels, prompt.height // downsampling_factor, prompt.width // downsampling_factor, @@ -228,9 +230,6 @@ def imagine( if prompt.init_image: generation_strength = 1 - prompt.init_image_strength t_enc = int(prompt.steps * generation_strength) - sampler.make_schedule( - ddim_num_steps=prompt.steps, ddim_eta=ddim_eta - ) try: init_image = pillow_fit_image_within( prompt.init_image, @@ -284,24 +283,35 @@ def imagine( # encode (scaled latent) seed_everything(prompt.seed) noise = torch.randn_like(init_latent, device="cpu").to(get_device()) + schedule = PLMSSchedule( + ddpm_num_timesteps=model.num_timesteps, + ddim_num_steps=prompt.steps, + alphas_cumprod=model.alphas_cumprod, + alphas_cumprod_prev=model.alphas_cumprod_prev, + betas=model.betas, + ddim_discretize="uniform", + ddim_eta=0.0, + ) if generation_strength >= 1: # prompt strength gets converted to time encodings, # which means you can't get to true 0 without this hack # (or setting steps=1000) z_enc = noise else: - z_enc = sampler.stochastic_encode( + z_enc = sampler.noise_an_image( init_latent, torch.tensor([t_enc - 1]).to(get_device()), + schedule=schedule, noise=noise, ) log_latent(z_enc, "z_enc") # decode it samples = sampler.decode( - x_latent=z_enc, + initial_latent=z_enc, cond=c, t_start=t_enc, + schedule=schedule, unconditional_guidance_scale=prompt.prompt_strength, unconditional_conditioning=uc, img_callback=_img_callback, diff --git a/imaginairy/enhancers/clip_masking.py b/imaginairy/enhancers/clip_masking.py index 7259fd2..7da39d7 100644 --- a/imaginairy/enhancers/clip_masking.py +++ b/imaginairy/enhancers/clip_masking.py @@ -56,7 +56,7 @@ def get_img_mask( mask[mask >= 0.5] = 1 log_img(mask, f"mask threshold {0.5}") - mask_np = mask.cpu().numpy() + mask_np = mask.to(torch.float32).cpu().numpy() smoother_strength = 2 # grow the mask area to make sure we've masked the thing we care about for _ in range(smoother_strength): diff --git a/imaginairy/samplers/base.py b/imaginairy/samplers/base.py index 45064da..f5420a2 100644 --- a/imaginairy/samplers/base.py +++ b/imaginairy/samplers/base.py @@ -2,7 +2,7 @@ import torch from torch import nn -from imaginairy.utils import get_device +from imaginairy.img_log import log_latent SAMPLER_TYPE_OPTIONS = [ "plms", @@ -51,11 +51,19 @@ class CFGDenoiser(nn.Module): self.inner_model = model def forward(self, x, sigma, uncond, cond, cond_scale, mask=None, orig_latent=None): - x_in = torch.cat([x] * 2) - sigma_in = torch.cat([sigma] * 2) - cond_in = torch.cat([uncond, cond]) - uncond, cond = self.inner_model(x_in, sigma_in, cond=cond_in).chunk(2) - denoised = uncond + (cond - uncond) * cond_scale + def _wrapper(noisy_latent_in, time_encoding_in, conditioning_in): + return self.inner_model( + noisy_latent_in, time_encoding_in, cond=conditioning_in + ) + + denoised = get_noise_prediction( + denoise_func=_wrapper, + noisy_latent=x, + time_encoding=sigma, + neutral_conditioning=uncond, + positive_conditioning=cond, + signal_amplification=cond_scale, + ) if mask is not None: assert orig_latent is not None @@ -65,51 +73,37 @@ class CFGDenoiser(nn.Module): return denoised -class DiffusionSampler: - """ - wip +def ensure_4_dim(t: torch.Tensor): + if len(t.shape) == 3: + t = t.unsqueeze(dim=0) + return t - hope to enforce an api upon samplers - """ - def __init__(self, noise_prediction_model, sampler_func, device=get_device()): - self.noise_prediction_model = noise_prediction_model - self.cfg_noise_prediction_model = CFGDenoiser(noise_prediction_model) - self.sampler_func = sampler_func - self.device = device - - def zzsample( - self, - num_steps, - text_conditioning, - batch_size, - shape, - unconditional_guidance_scale, - unconditional_conditioning, - eta, - initial_noise_tensor=None, - img_callback=None, - ): - size = (batch_size, *shape) - - initial_noise_tensor = ( - torch.randn(size, device="cpu").to(get_device()) - if initial_noise_tensor is None - else initial_noise_tensor - ) - sigmas = self.noise_prediction_model.get_sigmas(num_steps) - x = initial_noise_tensor * sigmas[0] - - samples = self.sampler_func( - self.cfg_noise_prediction_model, - x, - sigmas, - extra_args={ - "cond": text_conditioning, - "uncond": unconditional_conditioning, - "cond_scale": unconditional_guidance_scale, - }, - disable=False, - ) +def get_noise_prediction( + denoise_func, + noisy_latent, + time_encoding, + neutral_conditioning, + positive_conditioning, + signal_amplification=7.5, +): + noisy_latent = ensure_4_dim(noisy_latent) + + noisy_latent_in = torch.cat([noisy_latent] * 2) + time_encoding_in = torch.cat([time_encoding] * 2) + conditioning_in = torch.cat([neutral_conditioning, positive_conditioning]) + + pred_noise_neutral, pred_noise_positive = denoise_func( + noisy_latent_in, time_encoding_in, conditioning_in + ).chunk(2) + + amplified_noise_pred = signal_amplification * ( + pred_noise_positive - pred_noise_neutral + ) + pred_noise = pred_noise_neutral + amplified_noise_pred + + log_latent(pred_noise_neutral, "neutral noise prediction") + log_latent(pred_noise_positive, "positive noise prediction") + log_latent(pred_noise, "noise prediction") - return samples, None + return pred_noise diff --git a/imaginairy/samplers/ddim.py b/imaginairy/samplers/ddim.py index 24c5a72..4edad7b 100644 --- a/imaginairy/samplers/ddim.py +++ b/imaginairy/samplers/ddim.py @@ -1,5 +1,4 @@ # pylama:ignore=W0613 -"""SAMPLING ONLY.""" import logging import numpy as np @@ -13,41 +12,19 @@ from imaginairy.modules.diffusion.util import ( make_ddim_timesteps, noise_like, ) +from imaginairy.samplers.base import get_noise_prediction from imaginairy.utils import get_device logger = logging.getLogger(__name__) -class DDIMSampler: - """ - Denoising Diffusion Implicit Models - - https://arxiv.org/abs/2010.02502 - """ - - def __init__(self, model): - self.model = model - - def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0.0): - buffers = self._make_schedule( - model_num_timesteps=self.model.num_timesteps, - model_alphas_cumprod=self.model.alphas_cumprod, - model_betas=self.model.betas, - model_alphas_cumprod_prev=self.model.alphas_cumprod_prev, - ddim_num_steps=ddim_num_steps, - ddim_discretize=ddim_discretize, - ddim_eta=ddim_eta, - device=self.model.device, - ) - for k, v in buffers.items(): - setattr(self, k, v) - - @staticmethod - def _make_schedule( +class DDIMSchedule: + def __init__( + self, model_num_timesteps, model_alphas_cumprod, - model_betas, model_alphas_cumprod_prev, + model_betas, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0.0, @@ -71,41 +48,37 @@ class DDIMSampler: ddim_timesteps=ddim_timesteps, eta=ddim_eta, ) - - buffers = { - "ddim_timesteps": ddim_timesteps, - "betas": to_torch(model_betas), - "alphas_cumprod": to_torch(alphas_cumprod), - "alphas_cumprod_prev": to_torch(model_alphas_cumprod_prev), - # calculations for diffusion q(x_t | x_{t-1}) and others - "sqrt_alphas_cumprod": to_torch(np.sqrt(alphas_cumprod.cpu())), - "sqrt_one_minus_alphas_cumprod": to_torch( - np.sqrt(1.0 - alphas_cumprod.cpu()) - ), - "log_one_minus_alphas_cumprod": to_torch( - np.log(1.0 - alphas_cumprod.cpu()) - ), - "sqrt_recip_alphas_cumprod": to_torch(np.sqrt(1.0 / alphas_cumprod.cpu())), - "sqrt_recipm1_alphas_cumprod": to_torch( - np.sqrt(1.0 / alphas_cumprod.cpu() - 1) - ), - "ddim_sigmas": ddim_sigmas.to(torch.float32).to(device), - "ddim_alphas": ddim_alphas.to(torch.float32).to(device), - "ddim_alphas_prev": ddim_alphas_prev, - "ddim_sqrt_one_minus_alphas": np.sqrt(1.0 - ddim_alphas) - .to(torch.float32) - .to(device), - } - - sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( - (1 - buffers["alphas_cumprod_prev"]) - / (1 - buffers["alphas_cumprod"]) - * (1 - buffers["alphas_cumprod"] / buffers["alphas_cumprod_prev"]) + self.ddim_timesteps = ddim_timesteps + self.betas = to_torch(model_betas) + self.alphas_cumprod = to_torch(alphas_cumprod) + self.alphas_cumprod_prev = to_torch(model_alphas_cumprod_prev) + # calculations for diffusion q(x_t | x_{t-1}) and others + self.sqrt_alphas_cumprod = to_torch(np.sqrt(alphas_cumprod.cpu())) + self.sqrt_one_minus_alphas_cumprod = to_torch( + np.sqrt(1.0 - alphas_cumprod.cpu()) + ) + self.log_one_minus_alphas_cumprod = to_torch(np.log(1.0 - alphas_cumprod.cpu())) + self.sqrt_recip_alphas_cumprod = to_torch(np.sqrt(1.0 / alphas_cumprod.cpu())) + self.sqrt_recipm1_alphas_cumprod = to_torch( + np.sqrt(1.0 / alphas_cumprod.cpu() - 1) + ) + self.ddim_sigmas = ddim_sigmas.to(torch.float32).to(device) + self.ddim_alphas = ddim_alphas.to(torch.float32).to(device) + self.ddim_alphas_prev = ddim_alphas_prev + self.ddim_sqrt_one_minus_alphas = ( + np.sqrt(1.0 - ddim_alphas).to(torch.float32).to(device) ) - buffers[ - "ddim_sigmas_for_original_num_steps" - ] = sigmas_for_original_sampling_steps - return buffers + + +class DDIMSampler: + """ + Denoising Diffusion Implicit Models + + https://arxiv.org/abs/2010.02502 + """ + + def __init__(self, model): + self.model = model @torch.no_grad() def sample( @@ -123,31 +96,30 @@ class DDIMSampler: 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, **kwargs, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... ): - 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) + if conditioning.shape[0] != batch_size: + logger.warning( + f"Warning: Got {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, + model_alphas_cumprod_prev=self.model.alphas_cumprod_prev, + model_betas=self.model.betas, + ddim_num_steps=num_steps, + ddim_discretize="uniform", + ddim_eta=0.0, + ) samples = self.ddim_sampling( conditioning, - shape=(batch_size, *shape), + shape=shape, + schedule=schedule, callback=callback, img_callback=img_callback, quantize_denoised=quantize_x0, @@ -155,8 +127,6 @@ class DDIMSampler: x0=x0, 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, @@ -168,6 +138,7 @@ class DDIMSampler: self, cond, shape, + schedule, x_T=None, callback=None, timesteps=None, @@ -177,8 +148,6 @@ class DDIMSampler: img_callback=None, temperature=1.0, noise_dropout=0.0, - score_corrector=None, - corrector_kwargs=None, unconditional_guidance_scale=1.0, unconditional_conditioning=None, ): @@ -192,16 +161,16 @@ class DDIMSampler: log_latent(img, "initial noise") if timesteps is None: - timesteps = self.ddim_timesteps + timesteps = schedule.ddim_timesteps else: subset_end = ( int( - min(timesteps / self.ddim_timesteps.shape[0], 1) - * self.ddim_timesteps.shape[0] + min(timesteps / schedule.ddim_timesteps.shape[0], 1) + * schedule.ddim_timesteps.shape[0] ) - 1 ) - timesteps = self.ddim_timesteps[:subset_end] + timesteps = schedule.ddim_timesteps[:subset_end] time_range = np.flip(timesteps) total_steps = timesteps.shape[0] @@ -225,6 +194,7 @@ class DDIMSampler: cond, ts, index=index, + schedule=schedule, quantize_denoised=quantize_denoised, temperature=temperature, noise_dropout=noise_dropout, @@ -245,6 +215,7 @@ class DDIMSampler: c, t, index, + schedule, repeat_noise=False, quantize_denoised=False, temperature=1.0, @@ -254,26 +225,26 @@ class DDIMSampler: loss_function=None, ): assert unconditional_guidance_scale >= 1 - x_in = torch.cat([x] * 2) - t_in = torch.cat([t] * 2) - c_in = torch.cat([unconditional_conditioning, c]) - # with torch.no_grad(): - noise_pred_uncond, noise_pred = self.model.apply_model(x_in, t_in, c_in).chunk( - 2 - ) - noise_pred = noise_pred_uncond + unconditional_guidance_scale * ( - noise_pred - noise_pred_uncond + 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, ) b = x.shape[0] log_latent(noise_pred, "noise prediction") # select parameters corresponding to the currently considered timestep - a_t = torch.full((b, 1, 1, 1), self.ddim_alphas[index], device=x.device) - a_prev = torch.full((b, 1, 1, 1), self.ddim_alphas_prev[index], device=x.device) - sigma_t = torch.full((b, 1, 1, 1), self.ddim_sigmas[index], device=x.device) + a_t = torch.full((b, 1, 1, 1), schedule.ddim_alphas[index], device=x.device) + a_prev = torch.full( + (b, 1, 1, 1), schedule.ddim_alphas_prev[index], device=x.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), self.ddim_sqrt_one_minus_alphas[index], device=x.device + (b, 1, 1, 1), schedule.ddim_sqrt_one_minus_alphas[index], device=x.device ) return self._p_sample_ddim_formula( x, @@ -310,12 +281,11 @@ class DDIMSampler: return x_prev, pred_x0 @torch.no_grad() - def stochastic_encode(self, init_latent, t, noise=None): - # fast, but does not allow for exact reconstruction + def noise_an_image(self, init_latent, t, schedule, noise=None): # 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 + sqrt_alphas_cumprod = torch.sqrt(schedule.ddim_alphas) + sqrt_one_minus_alphas_cumprod = schedule.ddim_sqrt_one_minus_alphas if noise is None: noise = torch.randn_like(init_latent, device="cpu").to(get_device()) @@ -328,31 +298,34 @@ class DDIMSampler: @torch.no_grad() def decode( self, - x_latent, + initial_latent, cond, t_start, + schedule, unconditional_guidance_scale=1.0, unconditional_conditioning=None, img_callback=None, - score_corrector=None, temperature=1.0, mask=None, orig_latent=None, ): - timesteps = self.ddim_timesteps[:t_start] + timesteps = schedule.ddim_timesteps[:t_start] 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 = x_latent + x_dec = initial_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 + (initial_latent.shape[0],), + step, + device=initial_latent.device, + dtype=torch.long, ) if mask is not None: @@ -374,17 +347,12 @@ class DDIMSampler: x_dec, cond, ts, + schedule=schedule, index=index, unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=unconditional_conditioning, temperature=temperature, ) - # 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 log_latent(x_dec, f"x_dec {i}") log_latent(pred_x0, f"pred_x0 {i}") diff --git a/imaginairy/samplers/kdiff.py b/imaginairy/samplers/kdiff.py index dae309a..ece3a12 100644 --- a/imaginairy/samplers/kdiff.py +++ b/imaginairy/samplers/kdiff.py @@ -8,11 +8,15 @@ from imaginairy.vendored.k_diffusion import sampling as k_sampling from imaginairy.vendored.k_diffusion.external import CompVisDenoiser +class StandardCompVisDenoiser(CompVisDenoiser): + def apply_model(self, *args, **kwargs): + return self.inner_model.apply_model(*args, **kwargs) + + class KDiffusionSampler: def __init__(self, model, sampler_name): self.model = model - self.cv_denoiser = CompVisDenoiser(model) - # self.cfg_denoiser = CompVisDenoiser(self.cv_denoiser) + self.cv_denoiser = StandardCompVisDenoiser(model) self.sampler_name = sampler_name self.sampler_func = getattr(k_sampling, f"sample_{sampler_name}") @@ -28,10 +32,8 @@ class KDiffusionSampler: initial_noise_tensor=None, img_callback=None, ): - size = (batch_size, *shape) - initial_noise_tensor = ( - torch.randn(size, device="cpu").to(get_device()) + torch.randn(shape, device="cpu").to(get_device()) if initial_noise_tensor is None else initial_noise_tensor ) diff --git a/imaginairy/samplers/plms.py b/imaginairy/samplers/plms.py index 944cd10..bca91d8 100644 --- a/imaginairy/samplers/plms.py +++ b/imaginairy/samplers/plms.py @@ -1,5 +1,4 @@ # pylama:ignore=W0613 -"""SAMPLING ONLY.""" import logging import numpy as np @@ -13,65 +12,52 @@ from imaginairy.modules.diffusion.util import ( make_ddim_timesteps, noise_like, ) +from imaginairy.samplers.base import get_noise_prediction 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): +class PLMSSchedule: + def __init__( + self, + ddpm_num_timesteps, # 1000? + ddim_num_steps, # prompt.steps? + alphas_cumprod, + alphas_cumprod_prev, + betas, + 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 + device = get_device() + assert ( - alphas_cumprod.shape[0] == self.ddpm_num_timesteps + alphas_cumprod.shape[0] == 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) - ) + return x.clone().detach().to(torch.float32).to(device) + self.betas = to_torch(betas) + self.alphas_cumprod = to_torch(alphas_cumprod) + self.alphas_cumprod_prev = to_torch(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.sqrt_alphas_cumprod = to_torch(np.sqrt(alphas_cumprod.cpu())) + self.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.log_one_minus_alphas_cumprod = to_torch(np.log(1.0 - alphas_cumprod.cpu())) + self.sqrt_recip_alphas_cumprod = to_torch(np.sqrt(1.0 / alphas_cumprod.cpu())) + self.sqrt_recipm1_alphas_cumprod = to_torch( + np.sqrt(1.0 / alphas_cumprod.cpu() - 1) ) - 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)), + + self.ddim_timesteps = make_ddim_timesteps( + ddim_discr_method=ddim_discretize, + num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=ddpm_num_timesteps, ) # ddim sampling parameters @@ -80,19 +66,21 @@ class PLMSSampler: 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 + self.ddim_sigmas = ddim_sigmas.to(torch.float32).to(torch.device(device)) + self.ddim_alphas = ddim_alphas.to(torch.float32).to(torch.device(device)) + self.ddim_alphas_prev = ddim_alphas_prev + self.ddim_sqrt_one_minus_alphas = ( + np.sqrt(1.0 - ddim_alphas).to(torch.float32).to(torch.device(device)) ) + +class PLMSSampler: + """probabilistic least-mean-squares""" + + def __init__(self, model): + self.model = model + self.device = get_device() + @torch.no_grad() def sample( self, @@ -101,145 +89,90 @@ class PLMSSampler: shape, conditioning=None, callback=None, - normals_sequence=None, img_callback=None, quantize_x0=False, eta=0.0, mask=None, - x0=None, + orig_latent=None, temperature=1.0, noise_dropout=0.0, - score_corrector=None, - corrector_kwargs=None, - x_T=None, + initial_latent=None, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + timesteps=None, + quantize_denoised=False, # 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}" - ) + 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, + schedule = PLMSSchedule( + ddpm_num_timesteps=self.model.num_timesteps, + ddim_num_steps=num_steps, + alphas_cumprod=self.model.alphas_cumprod, + alphas_cumprod_prev=self.model.alphas_cumprod_prev, + betas=self.model.betas, + ddim_discretize="uniform", + ddim_eta=0.0, ) - 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") + device = self.device + # batch_size = shape[0] + if initial_latent is None: + initial_latent = torch.randn(shape, device="cpu").to(device) + log_latent(initial_latent, "initial latent") 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: + timesteps = schedule.ddim_timesteps + elif timesteps is not None: subset_end = ( int( - min(timesteps / self.ddim_timesteps.shape[0], 1) - * self.ddim_timesteps.shape[0] + min(timesteps / schedule.ddim_timesteps.shape[0], 1) + * schedule.ddim_timesteps.shape[0] ) - 1 ) - timesteps = self.ddim_timesteps[:subset_end] + timesteps = schedule.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] + time_range = np.flip(timesteps) + total_steps = 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 = [] + img = initial_latent for i, step in enumerate(iterator): index = total_steps - i - 1 - ts = torch.full((b,), step, device=device, dtype=torch.long) + ts = torch.full((batch_size,), step, device=device, dtype=torch.long) ts_next = torch.full( - (b,), + (batch_size,), 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? + assert orig_latent is not None + img_orig = self.model.q_sample(orig_latent, ts) img = img_orig * mask + (1.0 - mask) * img - img, pred_x0, e_t = self.p_sample_plms( + img, pred_x0, noise_prediction = self.p_sample_plms( img, - cond, + conditioning, ts, + schedule=schedule, 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) + old_eps.append(noise_prediction) if len(old_eps) >= 4: old_eps.pop(0) if callback: @@ -253,119 +186,108 @@ class PLMSSampler: @torch.no_grad() def p_sample_plms( self, - x, - c, - t, + noisy_latent, + positive_conditioning, + time_encoding, + schedule: PLMSSchedule, 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 + batch_size = noisy_latent.shape[0] + noise_prediction = get_noise_prediction( + denoise_func=self.model.apply_model, + noisy_latent=noisy_latent, + time_encoding=time_encoding, + neutral_conditioning=unconditional_conditioning, + positive_conditioning=positive_conditioning, + signal_amplification=unconditional_guidance_scale, ) 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) + alpha_at_t = torch.full( + (batch_size, 1, 1, 1), schedule.ddim_alphas[index], device=self.device + ) + alpha_prev_at_t = torch.full( + (batch_size, 1, 1, 1), + schedule.ddim_alphas_prev[index], + device=self.device, + ) + sigma_t = torch.full( + (batch_size, 1, 1, 1), schedule.ddim_sigmas[index], device=self.device + ) sqrt_one_minus_at = torch.full( - (b, 1, 1, 1), sqrt_one_minus_alphas[index], device=device + (batch_size, 1, 1, 1), + schedule.ddim_sqrt_one_minus_alphas[index], + device=self.device, ) # current prediction for x_0 - pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + pred_x0 = (noisy_latent - sqrt_one_minus_at * e_t) / alpha_at_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 + dir_xt = (1.0 - alpha_prev_at_t - sigma_t**2).sqrt() * e_t + noise = ( + sigma_t + * noise_like(noisy_latent.shape, self.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 + x_prev = alpha_prev_at_t.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 + x_prev, pred_x0 = get_x_prev_and_pred_x0(noise_prediction, index) + e_t_next = get_noise_prediction( + denoise_func=self.model.apply_model, + noisy_latent=x_prev, + time_encoding=t_next, + neutral_conditioning=unconditional_conditioning, + positive_conditioning=positive_conditioning, + signal_amplification=unconditional_guidance_scale, + ) + e_t_prime = (noise_prediction + 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 + e_t_prime = (3 * noise_prediction - 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 + e_t_prime = ( + 23 * noise_prediction - 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] + 55 * noise_prediction + - 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 + return x_prev, pred_x0, noise_prediction @torch.no_grad() - def stochastic_encode(self, init_latent, t, noise=None): + def noise_an_image(self, init_latent, t, schedule, noise=None): + # replace with ddpm.q_sample? # 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 + sqrt_alphas_cumprod = torch.sqrt(schedule.ddim_alphas) + sqrt_one_minus_alphas_cumprod = schedule.ddim_sqrt_one_minus_alphas if noise is None: noise = torch.randn_like(init_latent, device="cpu").to(get_device()) @@ -378,26 +300,26 @@ class PLMSSampler: @torch.no_grad() def decode( self, - x_latent, cond, - t_start, + schedule, + initial_latent=None, + t_start=None, 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] + device = self.device + timesteps = schedule.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 + iterator = tqdm(time_range, desc="PLMS img2img", total=total_steps) + x_dec = initial_latent old_eps = [] log_latent(x_dec, "x_dec") @@ -411,12 +333,15 @@ class PLMSSampler: 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 + (initial_latent.shape[0],), + step, + device=initial_latent.device, + dtype=torch.long, ) ts_next = torch.full( - (x_latent.shape[0],), + (initial_latent.shape[0],), time_range[min(i + 1, len(time_range) - 1)], - device=x_latent.device, + device=device, dtype=torch.long, ) @@ -435,10 +360,11 @@ class PLMSSampler: 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, pred_x0, noise_prediction = self.p_sample_plms( x_dec, cond, ts, + schedule=schedule, index=index, unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=unconditional_conditioning, @@ -446,14 +372,8 @@ class PLMSSampler: 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) + + old_eps.append(noise_prediction) if len(old_eps) >= 4: old_eps.pop(0) diff --git a/tests/conftest.py b/tests/conftest.py index cde2e37..ce5086f 100644 --- a/tests/conftest.py +++ b/tests/conftest.py @@ -47,6 +47,11 @@ def pre_setup(): yield +@pytest.fixture(autouse=True) +def reset_get_device(): + get_device.cache_clear() + + @pytest.fixture() def filename_base_for_outputs(request): filename_base = f"{TESTS_FOLDER}/test_output/{request.node.name}_{get_device()}_" diff --git a/tests/test_experiments.py b/tests/test_experiments.py index 294bed9..95b529f 100644 --- a/tests/test_experiments.py +++ b/tests/test_experiments.py @@ -54,7 +54,6 @@ def experiment_step_repeats(): embedder.to(get_device()) sampler = DDIMSampler(model) - sampler.make_schedule(1000) img = LazyLoadingImage(filepath=f"{TESTS_FOLDER}/data/beach_at_sainte_adresse.jpg") init_image, _, _ = pillow_img_to_torch_image( @@ -89,7 +88,9 @@ def experiment_step_repeats(): # noise_pred = model.apply_model(init_latent, t, neutral_embedding) # log_latent(noise_pred, "noise prediction") for _ in range(100): - x_prev, pred_x0 = sampler.p_sample_ddim(x_prev, neutral_embedding, t, index) + x_prev, pred_x0 = sampler.p_sample_ddim( # noqa + x_prev, neutral_embedding, t, index + ) log_latent(pred_x0, "pred_x0") x_prev = pred_x0 diff --git a/tests/test_utils.py b/tests/test_utils.py index 2355cf7..e861ff5 100644 --- a/tests/test_utils.py +++ b/tests/test_utils.py @@ -12,7 +12,6 @@ from imaginairy.utils import ( get_hardware_description, get_obj_from_str, instantiate_from_config, - platform_appropriate_autocast, ) @@ -79,6 +78,7 @@ def test_instantiate_from_config(): instantiate_from_config(config) -def test_platform_appropriate_autocast(): - with platform_appropriate_autocast("autocast"): - pass +# +# def test_platform_appropriate_autocast(): +# with platform_appropriate_autocast("autocast"): +# pass