# Copied from VideoCrafter: https://github.com/AILab-CVC/VideoCrafter """ wild mixture of https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py https://github.com/CompVis/taming-transformers -- merci """ from functools import partial from contextlib import contextmanager import numpy as np from tqdm import tqdm from einops import rearrange, repeat import logging mainlogger = logging.getLogger('mainlogger') import torch import torch.nn as nn from torchvision.utils import make_grid import pytorch_lightning as pl from utils.utils import instantiate_from_config from lvdm.ema import LitEma from lvdm.distributions import DiagonalGaussianDistribution from lvdm.models.utils_diffusion import make_beta_schedule from lvdm.modules.encoders.ip_resampler import ImageProjModel, Resampler from lvdm.basics import disabled_train from lvdm.common import ( extract_into_tensor, noise_like, exists, default ) # import ipdb # st = ipdb.set_trace __conditioning_keys__ = {'concat': 'c_concat', 'crossattn': 'c_crossattn', 'adm': 'y'} class DDPM(pl.LightningModule): # classic DDPM with Gaussian diffusion, in image space def __init__(self, unet_config, timesteps=1000, beta_schedule="linear", loss_type="l2", ckpt_path=None, ignore_keys=[], load_only_unet=False, monitor=None, use_ema=True, first_stage_key="image", image_size=256, channels=3, log_every_t=100, clip_denoised=True, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3, given_betas=None, original_elbo_weight=0., v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta l_simple_weight=1., conditioning_key=None, parameterization="eps", # all assuming fixed variance schedules scheduler_config=None, use_positional_encodings=False, learn_logvar=False, logvar_init=0. ): super().__init__() assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"' self.parameterization = parameterization mainlogger.info(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode") self.cond_stage_model = None self.clip_denoised = clip_denoised self.log_every_t = log_every_t self.first_stage_key = first_stage_key self.channels = channels self.temporal_length = unet_config.params.temporal_length self.image_size = image_size if isinstance(self.image_size, int): self.image_size = [self.image_size, self.image_size] self.use_positional_encodings = use_positional_encodings self.model = DiffusionWrapper(unet_config, conditioning_key) self.use_ema = use_ema if self.use_ema: self.model_ema = LitEma(self.model) mainlogger.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") self.use_scheduler = scheduler_config is not None if self.use_scheduler: self.scheduler_config = scheduler_config self.v_posterior = v_posterior self.original_elbo_weight = original_elbo_weight self.l_simple_weight = l_simple_weight if monitor is not None: self.monitor = monitor if ckpt_path is not None: self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet) self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps, linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s) self.loss_type = loss_type self.learn_logvar = learn_logvar self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,)) if self.learn_logvar: self.logvar = nn.Parameter(self.logvar, requires_grad=True) def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): if exists(given_betas): betas = given_betas else: betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s) alphas = 1. - betas alphas_cumprod = np.cumprod(alphas, axis=0) alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) timesteps, = betas.shape self.num_timesteps = int(timesteps) self.linear_start = linear_start self.linear_end = linear_end assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' to_torch = partial(torch.tensor, dtype=torch.float32) self.register_buffer('betas', to_torch(betas)) self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) self.register_buffer('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))) self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) # calculations for posterior q(x_{t-1} | x_t, x_0) posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / ( 1. - alphas_cumprod) + self.v_posterior * betas # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) self.register_buffer('posterior_variance', to_torch(posterior_variance)) # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20)))) self.register_buffer('posterior_mean_coef1', to_torch( betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))) self.register_buffer('posterior_mean_coef2', to_torch( (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod))) if self.parameterization == "eps": lvlb_weights = self.betas ** 2 / ( 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)) elif self.parameterization == "x0": lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod)) else: raise NotImplementedError("mu not supported") # TODO how to choose this term lvlb_weights[0] = lvlb_weights[1] self.register_buffer('lvlb_weights', lvlb_weights, persistent=False) assert not torch.isnan(self.lvlb_weights).all() @contextmanager def ema_scope(self, context=None): if self.use_ema: self.model_ema.store(self.model.parameters()) self.model_ema.copy_to(self.model) if context is not None: mainlogger.info(f"{context}: Switched to EMA weights") try: yield None finally: if self.use_ema: self.model_ema.restore(self.model.parameters()) if context is not None: mainlogger.info(f"{context}: Restored training weights") def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): sd = torch.load(path, map_location="cpu") if "state_dict" in list(sd.keys()): sd = sd["state_dict"] keys = list(sd.keys()) for k in keys: for ik in ignore_keys: if k.startswith(ik): mainlogger.info("Deleting key {} from state_dict.".format(k)) del sd[k] missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( sd, strict=False) mainlogger.info(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") if len(missing) > 0: mainlogger.info(f"Missing Keys: {missing}") if len(unexpected) > 0: mainlogger.info(f"Unexpected Keys: {unexpected}") def q_mean_variance(self, x_start, t): """ Get the distribution q(x_t | x_0). :param x_start: the [N x C x ...] tensor of noiseless inputs. :param t: the number of diffusion steps (minus 1). Here, 0 means one step. :return: A tuple (mean, variance, log_variance), all of x_start's shape. """ mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start) variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape) log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape) return mean, variance, log_variance def predict_start_from_noise(self, x_t, t, noise): return ( extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise ) def q_posterior(self, x_start, x_t, t): posterior_mean = ( extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start + extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t ) posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape) posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape) return posterior_mean, posterior_variance, posterior_log_variance_clipped def p_mean_variance(self, x, t, clip_denoised: bool): model_out = self.model(x, t) if self.parameterization == "eps": x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) elif self.parameterization == "x0": x_recon = model_out if clip_denoised: x_recon.clamp_(-1., 1.) model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) return model_mean, posterior_variance, posterior_log_variance @torch.no_grad() def p_sample(self, x, t, clip_denoised=True, repeat_noise=False): b, *_, device = *x.shape, x.device model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised) noise = noise_like(x.shape, device, repeat_noise) # no noise when t == 0 nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise @torch.no_grad() def p_sample_loop(self, shape, return_intermediates=False): device = self.betas.device b = shape[0] img = torch.randn(shape, device=device) intermediates = [img] for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps): img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long), clip_denoised=self.clip_denoised) if i % self.log_every_t == 0 or i == self.num_timesteps - 1: intermediates.append(img) if return_intermediates: return img, intermediates return img @torch.no_grad() def sample(self, batch_size=16, return_intermediates=False): image_size = self.image_size channels = self.channels return self.p_sample_loop((batch_size, channels, image_size, image_size), return_intermediates=return_intermediates) def q_sample(self, x_start, t, noise=None): noise = default(noise, lambda: torch.randn_like(x_start)) return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start * extract_into_tensor(self.scale_arr, t, x_start.shape) + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) def get_input(self, batch, k): x = batch[k] x = x.to(memory_format=torch.contiguous_format).float() return x def _get_rows_from_list(self, samples): n_imgs_per_row = len(samples) denoise_grid = rearrange(samples, 'n b c h w -> b n c h w') denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) return denoise_grid @torch.no_grad() def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs): log = dict() x = self.get_input(batch, self.first_stage_key) N = min(x.shape[0], N) n_row = min(x.shape[0], n_row) x = x.to(self.device)[:N] log["inputs"] = x # get diffusion row diffusion_row = list() x_start = x[:n_row] for t in range(self.num_timesteps): if t % self.log_every_t == 0 or t == self.num_timesteps - 1: t = repeat(torch.tensor([t]), '1 -> b', b=n_row) t = t.to(self.device).long() noise = torch.randn_like(x_start) x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) diffusion_row.append(x_noisy) log["diffusion_row"] = self._get_rows_from_list(diffusion_row) if sample: # get denoise row with self.ema_scope("Plotting"): samples, denoise_row = self.sample(batch_size=N, return_intermediates=True) log["samples"] = samples log["denoise_row"] = self._get_rows_from_list(denoise_row) if return_keys: if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: return log else: return {key: log[key] for key in return_keys} return log class LatentDiffusion(DDPM): """main class""" def __init__(self, first_stage_config, cond_stage_config, num_timesteps_cond=None, cond_stage_key="caption", cond_stage_trainable=False, cond_stage_forward=None, conditioning_key=None, uncond_prob=0.2, uncond_type="empty_seq", scale_factor=1.0, scale_by_std=False, encoder_type="2d", only_model=False, use_scale=False, scale_a=1, scale_b=0.3, mid_step=400, fix_scale_bug=False, *args, **kwargs): self.num_timesteps_cond = default(num_timesteps_cond, 1) self.scale_by_std = scale_by_std assert self.num_timesteps_cond <= kwargs['timesteps'] # for backwards compatibility after implementation of DiffusionWrapper ckpt_path = kwargs.pop("ckpt_path", None) ignore_keys = kwargs.pop("ignore_keys", []) conditioning_key = default(conditioning_key, 'crossattn') super().__init__(conditioning_key=conditioning_key, *args, **kwargs) self.cond_stage_trainable = cond_stage_trainable self.cond_stage_key = cond_stage_key # scale factor self.use_scale=use_scale if self.use_scale: self.scale_a=scale_a self.scale_b=scale_b if fix_scale_bug: scale_step=self.num_timesteps-mid_step else: #bug scale_step = self.num_timesteps scale_arr1 = np.linspace(scale_a, scale_b, mid_step) scale_arr2 = np.full(scale_step, scale_b) scale_arr = np.concatenate((scale_arr1, scale_arr2)) scale_arr_prev = np.append(scale_a, scale_arr[:-1]) to_torch = partial(torch.tensor, dtype=torch.float32) self.register_buffer('scale_arr', to_torch(scale_arr)) try: self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1 except: self.num_downs = 0 if not scale_by_std: self.scale_factor = scale_factor else: self.register_buffer('scale_factor', torch.tensor(scale_factor)) self.instantiate_first_stage(first_stage_config) self.instantiate_cond_stage(cond_stage_config) self.first_stage_config = first_stage_config self.cond_stage_config = cond_stage_config self.clip_denoised = False self.cond_stage_forward = cond_stage_forward self.encoder_type = encoder_type assert(encoder_type in ["2d", "3d"]) self.uncond_prob = uncond_prob self.classifier_free_guidance = True if uncond_prob > 0 else False assert(uncond_type in ["zero_embed", "empty_seq"]) self.uncond_type = uncond_type self.restarted_from_ckpt = False if ckpt_path is not None: self.init_from_ckpt(ckpt_path, ignore_keys, only_model=only_model) self.restarted_from_ckpt = True def make_cond_schedule(self, ): self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long) ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long() self.cond_ids[:self.num_timesteps_cond] = ids def q_sample(self, x_start, t, noise=None): noise = default(noise, lambda: torch.randn_like(x_start)) if self.use_scale: return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start * extract_into_tensor(self.scale_arr, t, x_start.shape) + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) else: return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) def _freeze_model(self): for name, para in self.model.diffusion_model.named_parameters(): para.requires_grad = False def instantiate_first_stage(self, config): model = instantiate_from_config(config) self.first_stage_model = model.eval() self.first_stage_model.train = disabled_train for param in self.first_stage_model.parameters(): param.requires_grad = False def instantiate_cond_stage(self, config): if not self.cond_stage_trainable: model = instantiate_from_config(config) self.cond_stage_model = model.eval() self.cond_stage_model.train = disabled_train for param in self.cond_stage_model.parameters(): param.requires_grad = False else: model = instantiate_from_config(config) self.cond_stage_model = model def get_learned_conditioning(self, c): if self.cond_stage_forward is None: if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode): c = self.cond_stage_model.encode(c) if isinstance(c, DiagonalGaussianDistribution): c = c.mode() else: c = self.cond_stage_model(c) else: assert hasattr(self.cond_stage_model, self.cond_stage_forward) c = getattr(self.cond_stage_model, self.cond_stage_forward)(c) return c def get_first_stage_encoding(self, encoder_posterior, noise=None): if isinstance(encoder_posterior, DiagonalGaussianDistribution): z = encoder_posterior.sample(noise=noise) elif isinstance(encoder_posterior, torch.Tensor): z = encoder_posterior else: raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented") return self.scale_factor * z @torch.no_grad() def encode_first_stage(self, x): if self.encoder_type == "2d" and x.dim() == 5: b, _, t, _, _ = x.shape x = rearrange(x, 'b c t h w -> (b t) c h w') reshape_back = True else: reshape_back = False encoder_posterior = self.first_stage_model.encode(x) results = self.get_first_stage_encoding(encoder_posterior).detach() if reshape_back: results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t) return results @torch.no_grad() def encode_first_stage_2DAE(self, x): b, _, t, _, _ = x.shape results = torch.cat([self.get_first_stage_encoding(self.first_stage_model.encode(x[:,:,i])).detach().unsqueeze(2) for i in range(t)], dim=2) return results def decode_core(self, z, **kwargs): if self.encoder_type == "2d" and z.dim() == 5: b, _, t, _, _ = z.shape z = rearrange(z, 'b c t h w -> (b t) c h w') reshape_back = True else: reshape_back = False z = 1. / self.scale_factor * z results = self.first_stage_model.decode(z, **kwargs) if reshape_back: results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t) return results @torch.no_grad() def decode_first_stage(self, z, **kwargs): return self.decode_core(z, **kwargs) def apply_model(self, x_noisy, t, cond, **kwargs): if isinstance(cond, dict): # hybrid case, cond is exptected to be a dict pass else: if not isinstance(cond, list): cond = [cond] key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn' cond = {key: cond} x_recon = self.model(x_noisy, t, **cond, **kwargs) if isinstance(x_recon, tuple): return x_recon[0] else: return x_recon def _get_denoise_row_from_list(self, samples, desc=''): denoise_row = [] for zd in tqdm(samples, desc=desc): denoise_row.append(self.decode_first_stage(zd.to(self.device))) n_log_timesteps = len(denoise_row) denoise_row = torch.stack(denoise_row) # n_log_timesteps, b, C, H, W if denoise_row.dim() == 5: # img, num_imgs= n_log_timesteps * bs, grid_size=[bs,n_log_timesteps] denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w') denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') denoise_grid = make_grid(denoise_grid, nrow=n_log_timesteps) elif denoise_row.dim() == 6: # video, grid_size=[n_log_timesteps*bs, t] video_length = denoise_row.shape[3] denoise_grid = rearrange(denoise_row, 'n b c t h w -> b n c t h w') denoise_grid = rearrange(denoise_grid, 'b n c t h w -> (b n) c t h w') denoise_grid = rearrange(denoise_grid, 'n c t h w -> (n t) c h w') denoise_grid = make_grid(denoise_grid, nrow=video_length) else: raise ValueError return denoise_grid # @torch.no_grad() def decode_first_stage_2DAE(self, z, **kwargs): b, _, t, _, _ = z.shape z = 1. / self.scale_factor * z results = torch.cat([self.first_stage_model.decode(z[:,:,i], **kwargs).unsqueeze(2) for i in range(t)], dim=2) return results def p_mean_variance(self, x, c, t, clip_denoised: bool, return_x0=False, score_corrector=None, corrector_kwargs=None, **kwargs): t_in = t model_out = self.apply_model(x, t_in, c, **kwargs) if score_corrector is not None: assert self.parameterization == "eps" model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs) if self.parameterization == "eps": x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) elif self.parameterization == "x0": x_recon = model_out else: raise NotImplementedError() if clip_denoised: x_recon.clamp_(-1., 1.) model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) if return_x0: return model_mean, posterior_variance, posterior_log_variance, x_recon else: return model_mean, posterior_variance, posterior_log_variance @torch.no_grad() def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, return_x0=False, \ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, **kwargs): b, *_, device = *x.shape, x.device outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, return_x0=return_x0, \ score_corrector=score_corrector, corrector_kwargs=corrector_kwargs, **kwargs) if return_x0: model_mean, _, model_log_variance, x0 = outputs else: model_mean, _, model_log_variance = outputs noise = noise_like(x.shape, device, repeat_noise) * temperature if noise_dropout > 0.: noise = torch.nn.functional.dropout(noise, p=noise_dropout) # no noise when t == 0 nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) if return_x0: return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0 else: return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise @torch.no_grad() def p_sample_loop(self, cond, shape, return_intermediates=False, x_T=None, verbose=True, callback=None, \ timesteps=None, mask=None, x0=None, img_callback=None, start_T=None, log_every_t=None, **kwargs): if not log_every_t: log_every_t = self.log_every_t device = self.betas.device b = shape[0] # sample an initial noise if x_T is None: img = torch.randn(shape, device=device) else: img = x_T intermediates = [img] if timesteps is None: timesteps = self.num_timesteps if start_T is not None: timesteps = min(timesteps, start_T) iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(range(0, timesteps)) if mask is not None: assert x0 is not None assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match for i in iterator: ts = torch.full((b,), i, device=device, dtype=torch.long) if self.shorten_cond_schedule: assert self.model.conditioning_key != 'hybrid' tc = self.cond_ids[ts].to(cond.device) cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) img = self.p_sample(img, cond, ts, clip_denoised=self.clip_denoised, **kwargs) if mask is not None: img_orig = self.q_sample(x0, ts) img = img_orig * mask + (1. - mask) * img if i % log_every_t == 0 or i == timesteps - 1: intermediates.append(img) if callback: callback(i) if img_callback: img_callback(img, i) if return_intermediates: return img, intermediates return img class LatentVisualDiffusion(LatentDiffusion): def __init__(self, cond_img_config, finegrained=False, random_cond=False, *args, **kwargs): super().__init__(*args, **kwargs) self.random_cond = random_cond self.instantiate_img_embedder(cond_img_config, freeze=True) num_tokens = 16 if finegrained else 4 self.image_proj_model = self.init_projector(use_finegrained=finegrained, num_tokens=num_tokens, input_dim=1024,\ cross_attention_dim=1024, dim=1280) def instantiate_img_embedder(self, config, freeze=True): embedder = instantiate_from_config(config) if freeze: self.embedder = embedder.eval() self.embedder.train = disabled_train for param in self.embedder.parameters(): param.requires_grad = False def init_projector(self, use_finegrained, num_tokens, input_dim, cross_attention_dim, dim): if not use_finegrained: image_proj_model = ImageProjModel(clip_extra_context_tokens=num_tokens, cross_attention_dim=cross_attention_dim, clip_embeddings_dim=input_dim ) else: image_proj_model = Resampler(dim=input_dim, depth=4, dim_head=64, heads=12, num_queries=num_tokens, embedding_dim=dim, output_dim=cross_attention_dim, ff_mult=4 ) return image_proj_model ## Never delete this func: it is used in log_images() and inference stage def get_image_embeds(self, batch_imgs): ## img: b c h w img_token = self.embedder(batch_imgs) img_emb = self.image_proj_model(img_token) return img_emb class DiffusionWrapper(pl.LightningModule): def __init__(self, diff_model_config, conditioning_key): super().__init__() self.diffusion_model = instantiate_from_config(diff_model_config) self.conditioning_key = conditioning_key def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, c_adm=None, s=None, mask=None, **kwargs): # temporal_context = fps is foNone if self.conditioning_key is None: out = self.diffusion_model(x, t) elif self.conditioning_key == 'concat': xc = torch.cat([x] + c_concat, dim=1) out = self.diffusion_model(xc, t, **kwargs) elif self.conditioning_key == 'crossattn': cc = torch.cat(c_crossattn, 1) out = self.diffusion_model(x, t, context=cc, **kwargs) elif self.conditioning_key == 'hybrid': ## it is just right [b,c,t,h,w]: concatenate in channel dim xc = torch.cat([x] + c_concat, dim=1) cc = torch.cat(c_crossattn, 1) out = self.diffusion_model(xc, t, context=cc) elif self.conditioning_key == 'resblockcond': cc = c_crossattn[0] out = self.diffusion_model(x, t, context=cc) elif self.conditioning_key == 'adm': cc = c_crossattn[0] out = self.diffusion_model(x, t, y=cc) elif self.conditioning_key == 'hybrid-adm': assert c_adm is not None xc = torch.cat([x] + c_concat, dim=1) cc = torch.cat(c_crossattn, 1) out = self.diffusion_model(xc, t, context=cc, y=c_adm) elif self.conditioning_key == 'hybrid-time': assert s is not None xc = torch.cat([x] + c_concat, dim=1) cc = torch.cat(c_crossattn, 1) out = self.diffusion_model(xc, t, context=cc, s=s) elif self.conditioning_key == 'concat-time-mask': # assert s is not None # mainlogger.info('x & mask:',x.shape,c_concat[0].shape) xc = torch.cat([x] + c_concat, dim=1) out = self.diffusion_model(xc, t, context=None, s=s, mask=mask) elif self.conditioning_key == 'concat-adm-mask': # assert s is not None # mainlogger.info('x & mask:',x.shape,c_concat[0].shape) if c_concat is not None: xc = torch.cat([x] + c_concat, dim=1) else: xc = x out = self.diffusion_model(xc, t, context=None, y=s, mask=mask) elif self.conditioning_key == 'hybrid-adm-mask': cc = torch.cat(c_crossattn, 1) if c_concat is not None: xc = torch.cat([x] + c_concat, dim=1) else: xc = x out = self.diffusion_model(xc, t, context=cc, y=s, mask=mask) elif self.conditioning_key == 'hybrid-time-adm': # adm means y, e.g., class index # assert s is not None assert c_adm is not None xc = torch.cat([x] + c_concat, dim=1) cc = torch.cat(c_crossattn, 1) out = self.diffusion_model(xc, t, context=cc, s=s, y=c_adm) else: raise NotImplementedError() return out