# Modified from OpenAI's diffusion repos # GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py # ADM: https://github.com/openai/guided-diffusion/blob/main/guided_diffusion # IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py import math import numpy as np import torch as th import enum from .diffusion_utils import discretized_gaussian_log_likelihood, normal_kl from pdb import set_trace as stx import copy from tqdm import tqdm import os from einops import rearrange from PIL import Image def mean_flat(tensor): """ Take the mean over all non-batch dimensions. """ return tensor.mean(dim=list(range(1, len(tensor.shape)))) ''' ModelMeanType 是一个枚举类, 里面有三个枚举值, 可以通过属性的方式来访问, 如 ModelMeanType.PREVIOUS_X enum.auto() 会自动分配一个值, 从 1 开始, 依次递增 ''' class ModelMeanType(enum.Enum): """ Which type of output the model predicts. """ PREVIOUS_X = enum.auto() # the model predicts x_{t-1}, 从 x_t 预测 x_{t-1} START_X = enum.auto() # the model predicts x_0 EPSILON = enum.auto() # the model predicts epsilon ''' Gaussian 的 variance 不应该是 schedule 的吗? ''' class ModelVarType(enum.Enum): """ What is used as the model's output variance. The LEARNED_RANGE option has been added to allow the model to predict values between FIXED_SMALL and FIXED_LARGE, making its job easier. """ LEARNED = enum.auto() FIXED_SMALL = enum.auto() FIXED_LARGE = enum.auto() LEARNED_RANGE = enum.auto() class LossType(enum.Enum): MSE = enum.auto() # use raw MSE loss (and KL when learning variances) RESCALED_MSE = ( enum.auto() ) # use raw MSE loss (with RESCALED_KL when learning variances) KL = enum.auto() # use the variational lower-bound RESCALED_KL = enum.auto() # like KL, but rescale to estimate the full VLB def is_vb(self): return self == LossType.KL or self == LossType.RESCALED_KL # num_diffusion_timesteps = T # warmup_frac - 指的是 warmup 的比例 # beta 逐步变化，最后到一个 beta_end 后保持不变 def _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, warmup_frac): betas = beta_end * np.ones(num_diffusion_timesteps, dtype=np.float64) warmup_time = int(num_diffusion_timesteps * warmup_frac) betas[:warmup_time] = np.linspace( beta_start, beta_end, warmup_time, dtype=np.float64 ) return betas # * 之前位置参数, * 之后必须是关键字参数 # 不同的 beta_schedule 策略 def get_beta_schedule(beta_schedule, *, beta_start, beta_end, num_diffusion_timesteps): """ This is the deprecated API for creating beta schedules. See get_named_beta_schedule() for the new library of schedules. """ if beta_schedule == "quad": betas = ( np.linspace( beta_start**0.5, beta_end**0.5, num_diffusion_timesteps, dtype=np.float64, ) ** 2 ) elif beta_schedule == "linear": betas = np.linspace( beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64 ) elif beta_schedule == "warmup10": betas = _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, 0.1) elif beta_schedule == "warmup50": betas = _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, 0.5) elif beta_schedule == "const": betas = beta_end * np.ones(num_diffusion_timesteps, dtype=np.float64) elif beta_schedule == "jsd": # 1/T, 1/(T-1), 1/(T-2), ..., 1 betas = 1.0 / np.linspace( num_diffusion_timesteps, 1, num_diffusion_timesteps, dtype=np.float64 ) else: raise NotImplementedError(beta_schedule) assert betas.shape == (num_diffusion_timesteps,) return betas # 另一些之前被命名过的 beta scheduler def get_named_beta_schedule(schedule_name, num_diffusion_timesteps): """ Get a pre-defined beta schedule for the given name. The beta schedule library consists of beta schedules which remain similar in the limit of num_diffusion_timesteps. Beta schedules may be added, but should not be removed or changed once they are committed to maintain backwards compatibility. """ if schedule_name == "linear": # Linear schedule from Ho et al, extended to work for any number of # diffusion steps. scale = 1000 / num_diffusion_timesteps return get_beta_schedule( "linear", beta_start=scale * 0.0001, beta_end=scale * 0.02, num_diffusion_timesteps=num_diffusion_timesteps, ) elif schedule_name == "squaredcos_cap_v2": return betas_for_alpha_bar( num_diffusion_timesteps, lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2, ) else: raise NotImplementedError(f"unknown beta schedule: {schedule_name}") # 还是一个 scheduler def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): """ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [0,1]. :param num_diffusion_timesteps: the number of betas to produce. :param alpha_bar: a lambda that takes an argument t from 0 to 1 and produces the cumulative product of (1-beta) up to that part of the diffusion process. :param max_beta: the maximum beta to use; use values lower than 1 to prevent singularities. """ betas = [] for i in range(num_diffusion_timesteps): t1 = i / num_diffusion_timesteps t2 = (i + 1) / num_diffusion_timesteps betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) return np.array(betas) # 注意，这只是个 general 的框架, model 是灵活的 ''' 如果想要 x_0 output 的话，好像就只用在调用 GaussianDiffusion 时的时候，设置 model_mean_type = ModelMeanType.START_X 即可 因为 x_{t-1} 一直是拿 x_t 和 x_0 来算的, q(x_{t-1} | x_t, x_0) ''' class GaussianDiffusion: """ Utilities for training and sampling diffusion models. Original ported from this codebase: https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/diffusion_utils_2.py#L42 :param betas: a 1-D numpy array of betas for each diffusion timestep, starting at T and going to 1. """ def __init__(self, *, betas, model_mean_type, model_var_type, loss_type): self.model_mean_type = model_mean_type self.model_var_type = model_var_type self.loss_type = loss_type # Use float64 for accuracy. ''' beta 控制的是前向传播过程中逐步所加的噪声的方差 q(x_t | x_{t-1}) = N(x_t; sqrt(1-beta_t) x_{t-1}, beta_t) alpha_t = 1 - beta_t ''' betas = np.array(betas, dtype=np.float64) self.betas = betas assert len(betas.shape) == 1, "betas must be 1-D, [B,]" assert (betas > 0).all() and (betas <= 1).all() self.num_timesteps = int(betas.shape[0]) alphas = 1.0 - betas # cumulative product of alphas over time self.alphas_cumprod = np.cumprod(alphas, axis=0) # 在 self.alphas_cumprod 数组的开始处添加了一个元素 1.0，并且移除了最后一个元素 self.alphas_cumprod_prev = np.append(1.0, self.alphas_cumprod[:-1]) # 在 self.alphas_cumprod 数组的末尾添加了一个元素 0.0，并且移除了第一个元素 self.alphas_cumprod_next = np.append(self.alphas_cumprod[1:], 0.0) assert self.alphas_cumprod_prev.shape == (self.num_timesteps,) # calculations for diffusion q(x_t | x_{t-1}) and others self.sqrt_alphas_cumprod = np.sqrt(self.alphas_cumprod) self.sqrt_one_minus_alphas_cumprod = np.sqrt(1.0 - self.alphas_cumprod) self.log_one_minus_alphas_cumprod = np.log(1.0 - self.alphas_cumprod) self.sqrt_recip_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod) # reciprocal - 倒数 self.sqrt_recipm1_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod - 1) # calculations for posterior q(x_{t-1} | x_t, x_0) self.posterior_variance = ( betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod) ) # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain # 起始值为 0 不可计算, 只可计算后续的值 self.posterior_log_variance_clipped = ( np.log(np.append(self.posterior_variance[1], self.posterior_variance[1:])) if len(self.posterior_variance) > 1 else np.array([]) ) ''' 下边两个是用来计算 q(x_{t-1} | x_t, x_0) 均值的两个系数 coef1 是 x_0 的系数, coef2 是 x_t 的系数 然后 x_0 是由 x_t 和 predict 的 噪声 eps 线性组合得到的, 是模型这个阶段猜出的 x_0, 并不是真正的 x_0 ''' self.posterior_mean_coef1 = ( betas * np.sqrt(self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod) ) self.posterior_mean_coef2 = ( (1.0 - self.alphas_cumprod_prev) * np.sqrt(alphas) / (1.0 - self.alphas_cumprod) ) ''' q(x_t | x_0) = N(x_t; sqrt(alpha_t_bar) x_0, sqrt(1 - alpha_t_bar)) ''' 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. 真实的 x_0, 非模型所猜 :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 # sqrt(alpha_t_bar) x_0 ) variance = _extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape) # sqrt(1 - alpha_t_bar) log_variance = _extract_into_tensor( self.log_one_minus_alphas_cumprod, t, x_start.shape ) return mean, variance, log_variance # x_0, t -> x_t # q(x_t | x_0) = N(x_t; sqrt(alpha_t_bar) x_0, sqrt(1 - alpha_t_bar)) # 注意：此处不是重参数化采样 ！！ 因为此处不需要求导了, x_t 是U-Net 的输入 # 重参数化采样发生在 q(x_{t-1} | x_t, x_0), 因为此时就需要求导，loss在 eps 上计算，直接关联 x_{t-1} 的 μ 和 σ, 需要梯度回传 def q_sample(self, x_start, t, noise=None): """ Diffuse the data for a given number of diffusion steps. In other words, sample from q(x_t | x_0). :param x_start: the initial data batch. :param t: the number of diffusion steps (minus 1). Here, 0 means one step. :param noise: if specified, the split-out normal noise. :return: A noisy version of x_start. """ if noise is None: noise = th.randn_like(x_start) assert noise.shape == x_start.shape 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 ) # prior - 啥也没观测到时，事件A的概率 p(A) # posterior - 观测到一些数据或者结果 B 后, A事件发生的概率 p(A|B) # 此处的 x_start 是模型猜的 x_0 # 由于此时 U-Net 已经完成了 eps 的预测，所以这里没有 model def q_posterior_mean_variance(self, x_start, x_t, t): """ Compute the mean and variance of the diffusion posterior: q(x_{t-1} | x_t, x_0) """ # stx() assert x_start.shape == x_t.shape 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 ) assert ( posterior_mean.shape[0] == posterior_variance.shape[0] == posterior_log_variance_clipped.shape[0] == x_start.shape[0] ) return posterior_mean, posterior_variance, posterior_log_variance_clipped # 此处调用 model def p_mean_variance(self, model, input_batch, t, clip_denoised=True, model_kwargs=None): """ Apply the model to get p(x_{t-1} | x_t), as well as a prediction of the initial x, x_0. :param model: the model, which takes a signal and a batch of timesteps as input. :param x: the [N x C x ...] tensor at time t. :param t: a 1-D Tensor of timesteps. :param clip_denoised: if True, clip the denoised signal into [-1, 1]. :param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning. :return: a dict with the following keys: - 'mean': the model mean output. - 'variance': the model variance output. - 'log_variance': the log of 'variance'. - 'pred_xstart': the prediction for x_0. """ if model_kwargs is None: model_kwargs = {} x = input_batch["image_noisy"] B, V, C = x.shape[:3] assert t.shape == (B,) ''' input_batch: dict_keys(['image', 'c2w', 'fxfycxcy', 'index', 'depth', 'normal', 'image_noisy']) image: [8, 8, 3, 256, 256] image_noisy: [8, 3, 3, 256, 256] model_output: [8, 3, 3, 256, 256] 这里存一组图片来看一下 ''' # stx() if t[0] > 0: input_batch['image'] = th.cat([input_batch['image'][:,0:1],input_batch['image_noisy']],dim=1) # select first images and update input render_imgs, pred_guassians = model(input_batch, t) #render_imgs, pred_guassians = model(**input_batch) model_output = render_imgs[:,1:] output_dict = { 'render_images':render_imgs, 'pred_gaussians':pred_guassians, } else: input_batch['image'] = th.cat([input_batch['image'][:,0:1],input_batch['image_noisy']],dim=1) # select first images and update input render_imgs, pred_guassians = model(input_batch, t) model_output = render_imgs[:,1:] output_dict = { 'render_images':render_imgs, 'pred_gaussians':pred_guassians, } # variance 的两种方式，一是模型直接输出 # 二是预先设定好的 variance (beta) 直接按照 timestep 取值 if self.model_var_type in [ModelVarType.LEARNED, ModelVarType.LEARNED_RANGE]: assert model_output.shape == (B, C * 2, *x.shape[2:]) model_output, model_var_values = th.split(model_output, C, dim=1) min_log = _extract_into_tensor( self.posterior_log_variance_clipped, t, x.shape ) max_log = _extract_into_tensor(np.log(self.betas), t, x.shape) # The model_var_values is [-1, 1] for [min_var, max_var]. frac = (model_var_values + 1) / 2 model_log_variance = frac * max_log + (1 - frac) * min_log model_variance = th.exp(model_log_variance) else: model_variance, model_log_variance = { # for fixedlarge, we set the initial (log-)variance like so # to get a better decoder log likelihood. ModelVarType.FIXED_LARGE: ( np.append(self.posterior_variance[1], self.betas[1:]), np.log(np.append(self.posterior_variance[1], self.betas[1:])), ), ModelVarType.FIXED_SMALL: ( self.posterior_variance, self.posterior_log_variance_clipped, ), }[self.model_var_type] model_variance = _extract_into_tensor(model_variance, t, x.shape) model_log_variance = _extract_into_tensor(model_log_variance, t, x.shape) # 为什么要 clamp 到 [-1, 1]? def process_xstart(x): if clip_denoised: return x.clamp(-1, 1) return x if self.model_mean_type == ModelMeanType.START_X: # stx() pred_xstart = process_xstart(model_output) else: pred_xstart = process_xstart( self._predict_xstart_from_eps(x_t=x, t=t, eps=model_output) ) # q(x_{t-1} | x_t, x_0), 这里的 x_0 是模型猜的 model_mean, _, _ = self.q_posterior_mean_variance( x_start=pred_xstart, x_t=x, t=t ) ###################################################################################################################### ########################################### sampling 存图 debug ############################################### ###################################################################################################################### # sampling_debug_path = os.path.join("/sensei-fs/users/yuanhaoc/gslrm_dit/minLRM/debug_sampling", "gaussian_diffusion") # os.makedirs(sampling_debug_path, exist_ok=True) # im = model_output # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(sampling_debug_path, "render_input.png")) # im = input_batch["image_noisy"] # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(sampling_debug_path, "xt_noisy.png")) # im = model_mean # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(sampling_debug_path, "model_mean_next_t.png")) # im = model_variance # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(sampling_debug_path, "model_variance_next_t.png")) # im = pred_xstart # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(sampling_debug_path, "pred_xstart_after_clipping.png")) # stx() assert ( model_mean.shape == model_log_variance.shape == pred_xstart.shape == x.shape ) return { "mean": model_mean, # μθ of x_{t-1} "variance": model_variance, # σ^2 "log_variance": model_log_variance, # log(σ^2) "pred_xstart": pred_xstart, # x_0 "denoiser_output_dict": output_dict, } # 从 eps 预测 x_0, 正常走的是这一步 # x_0 = 1/sqrt(alpha_t_bar) x_t - sqrt((1-alpha_t_bar) / alpha_t_bar) eps def _predict_xstart_from_eps(self, x_t, t, eps): assert x_t.shape == eps.shape 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) * eps ) # 由 x_0 预测 eps, 这里假设模型输出 x_0 # eps = (x_t - sqrt(alpha_t_bar) x_0) / sqrt(1-alpha_t_bar) def _predict_eps_from_xstart(self, x_t, t, pred_xstart): return ( _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart ) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) # predict x_{t-1} def p_sample( self, model, input_batch, t, clip_denoised=True, model_kwargs=None, ): """ Sample x_{t-1} from the model at the given timestep. :param model: the model to sample from. :param x: the current tensor at x_t. :param t: the value of t, starting at 0 for the first diffusion step. :param clip_denoised: if True, clip the x_start prediction to [-1, 1]. :param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning. :return: a dict containing the following keys: - 'sample': a random sample from the model. - 'pred_xstart': a prediction of x_0. """ # stx() x = input_batch["image_noisy"] out = self.p_mean_variance( model, input_batch, t, clip_denoised=clip_denoised, model_kwargs=model_kwargs, ) noise = th.randn_like(x) nonzero_mask = ( (t != 0).float().view(-1, *([1] * (len(x.shape) - 1))) ) # no noise when t == 0 # 此处是重参数化采样 x_{t-1} = μ + σ * eps # [1] 为何要 nonzero_mask, 因为底 0 步可以直接输出 x_0, 不需要加噪声 # [2] 为何要有一个 torch.exp(0.5), 给方差开根号就是协方差 sample = out["mean"] + nonzero_mask * th.exp(0.5 * out["log_variance"]) * noise input_batch["image_noisy"] = sample return {"sample": sample, "pred_xstart": out["pred_xstart"], "input_batch": input_batch, "denoiser_output_dict": out["denoiser_output_dict"]} # 返回 x_{t-1}, x_0, 和 input_batch def p_sample_loop( self, model, shape, input_batch=None, clip_denoised=True, model_kwargs=None, device=None, progress=True, ): """ Generate samples from the model. :param model: the model module. :param shape: the shape of the samples, (N, C, H, W). :param noise: if specified, the noise from the encoder to sample. Should be of the same shape as `shape`. :param clip_denoised: if True, clip x_start predictions to [-1, 1]. :param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning. :param device: if specified, the device to create the samples on. If not specified, use a model parameter's device. :param progress: if True, show a tqdm progress bar. :return: a non-differentiable batch of samples. """ final = None # 对 self.p_sample_loop_progressive 函数返回的对象进行 forloop # stx() for sample in self.p_sample_loop_progressive( model, shape, input_batch=input_batch, clip_denoised=clip_denoised, model_kwargs=model_kwargs, device=device, progress=progress, ): final = sample # 此处对应的 sample 应该是 p_sample 函数的 return 值，但这个函数好像就只需要里面的 "sample" return final # 这个 "sample" 对应 x_{t-1}, 即每一步的状态, 到最后一步会把 x_0 输出 # 生成一系列的 x_{t-1}，从 t = T 逐渐到 t = 0 def p_sample_loop_progressive( self, model, shape, input_batch=None, clip_denoised=True, model_kwargs=None, device=None, progress=False, ): """ Generate samples from the model and yield intermediate samples from each timestep of diffusion. Arguments are the same as p_sample_loop(). Returns a generator over dicts, where each dict is the return value of p_sample(). """ if device is None: device = next(model.parameters()).device assert isinstance(shape, (tuple, list)) indices = list(range(self.num_timesteps))[::-1] # 表示的是 t 的序号，indices_t 不一定等于 t if progress: # Lazy import so that we don't depend on tqdm. from tqdm.auto import tqdm # 自动根据环境选择进度条 indices = tqdm(indices) # stx() for i in indices: t = th.tensor([i] * shape[0], device=device) with th.no_grad(): # stx() output = self.p_sample( model, input_batch, # img = x_t, p_sample 会返回 x_{t-1}. img 的初始值是 noise t, clip_denoised=clip_denoised, model_kwargs=model_kwargs, ) yield output # 修改 input_batch 里面的迭代字典 # 应该在 p_sample 里面修改 input_batch 的关键字, 然后这边就直接传就好 input_batch = output["input_batch"] def _vb_terms_bpd( self, model, x_start, x_t, t, clip_denoised=True, model_kwargs=None ): """ Get a term for the variational lower-bound. The resulting units are bits (rather than nats, as one might expect). This allows for comparison to other papers. :return: a dict with the following keys: - 'output': a shape [N] tensor of NLLs or KLs. - 'pred_xstart': the x_0 predictions. """ # 计算 q(x_{t-1} | x_t, x_0) - 此处的 x_0 是真实值 true_mean, _, true_log_variance_clipped = self.q_posterior_mean_variance( x_start=x_start, x_t=x_t, t=t ) # 调用 model 计算 p(x_{t-1} | x_t) - 此处的 x_0 是模型猜的 out = self.p_mean_variance( model, x_t, t, clip_denoised=clip_denoised, model_kwargs=model_kwargs ) # 计算 KL(q(x_{t-1} | x_t, x_0) || p(x_{t-1} | x_t)), true 和 out 的 KL 散度 kl = normal_kl( true_mean, true_log_variance_clipped, out["mean"], out["log_variance"] ) kl = mean_flat(kl) / np.log(2.0) # 第一步输出对数似然，后面输出 KL 散度 decoder_nll = -discretized_gaussian_log_likelihood( x_start, means=out["mean"], log_scales=0.5 * out["log_variance"] ) assert decoder_nll.shape == x_start.shape decoder_nll = mean_flat(decoder_nll) / np.log(2.0) # At the first timestep return the decoder NLL, # otherwise return KL(q(x_{t-1}|x_t,x_0) || p(x_{t-1}|x_t)) output = th.where((t == 0), decoder_nll, kl) return {"output": output, "pred_xstart": out["pred_xstart"]} def training_losses(self, model, batch, t, model_kwargs=None, noise=None, mode='training', create_visual = False, x0_loss_weight = 1.0): """ 被调用形式: loss_metrics = self.diffusion.training_losses(self.denoiser, batch, t) Compute training losses for a single timestep. :param model: the model to evaluate loss on. :param batch: data batch - 一个 dict: dict_keys(['image', 'c2w', 'fxfycxcy', 'index', 'depth', 'normal']) :param t: a batch of timestep indices. :return: a dict with the key "loss" containing a tensor of shape [N]. Some mean or variance settings may also have other keys. 返回一个 loss 的 dict """ # 取 batch 的第二个 view 开始作为 x_0 # 这里的 batch 还是一整个的 batch # (b, v, c, h, w) img_condition = batch["image"][:,0:1, ...] # [8, 1, 3, 256, 256] x_start = batch["image"][:,1:4, ...].clone() # [8, 3, 3, 256, 256] if model_kwargs is None: model_kwargs = {} if noise is None: noise = th.randn_like(x_start) # [8, 3, 3, 256, 256] # # x_0, t -> x_t # q(x_t | x_0) = N(x_t; sqrt(alpha_t_bar) x_0, sqrt(1 - alpha_t_bar)) # x_t: (b, v-1, c, h, w) x_t = self.q_sample(x_start, t, noise=noise) batch['image_noisy'] = x_t # 这个值不会引起别的改变, 因为是新开辟的 key # stx() terms = {} # 计算 KL(q(x_{t-1} | x_t, x_0) || p(x_{t-1} | x_t)) 散度 Loss, KL 散度 Loss 用得其实不多 # 这里的 if 逻辑有点奇怪，就是说如果是 KL 散度 Loss 的话，就直接返回 KL 散度 Loss # 如果是 MSE Loss 的话，不仅要算 MSE Loss，还要算 KL 散度 Loss，返回的 loss 是两者之和 # 此处是 False, 不执行 if self.loss_type == LossType.KL or self.loss_type == LossType.RESCALED_KL: terms["loss"] = self._vb_terms_bpd( model=model, x_start=x_start, x_t=x_t, t=t, clip_denoised=False, model_kwargs=model_kwargs, )["output"] if self.loss_type == LossType.RESCALED_KL: terms["loss"] *= self.num_timesteps elif self.loss_type == LossType.MSE or self.loss_type == LossType.RESCALED_MSE: denoiser_output = model(batch, t, **model_kwargs, mode = mode, create_visual=create_visual) # 过 denoiser model_output = denoiser_output['render_input'] ####################################################################################################### ############################# 可视化 debug - 计算 loss 前把所有的图片存下来进行调试 ####################### ####################################################################################################### # debug_path = os.path.join("/sensei-fs/users/yuanhaoc/gslrm_dit/minLRM/debug", "gaussian_diffusion") # os.makedirs(debug_path, exist_ok=True) # im = model_output # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(debug_path, "model_render_input.png")) # im = denoiser_output['render'] # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(debug_path, "model_render_target.png")) # im = denoiser_output['input']['image'] # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(debug_path, "input_image.png")) # im = denoiser_output['target']['image'] # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(debug_path, "target_image.png")) # im_x_start = x_start # im_x_start = rearrange(im_x_start, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im_x_start = (im_x_start * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im_x_start).save(os.path.join(debug_path, "x_start.png")) # im = x_t # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(debug_path, "x_t.png")) # combo_image_condition = th.cat([img_condition, x_t], dim=1) # im = combo_image_condition # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(debug_path, "combo_input_image.png")) # stx() ################################################################################################################ ############################################ 可视化 debug 结束 ############################################# ################################################################################################################ ''' model_output 理论上也是一个字典 dict_keys(['input', 'target', 'gaussians', 'pixelalign_xyz', 'img_tokens', 'loss_metrics', 'render']) ——> dict_keys(['input', 'target', 'gaussians', 'pixelalign_xyz', 'img_tokens', 'loss_metrics', 'render', 'render_input']) 要从中抽出 multi-view 的 x0 然后变成 x_{t-1} ''' # learn_sigma = True, 所以模型不仅要 predict mean, 还要 predict variance # 我已经设置成 False 了 if self.model_var_type in [ ModelVarType.LEARNED, ModelVarType.LEARNED_RANGE, ]: B, C = x_t.shape[:2] assert model_output.shape == (B, C * 2, *x_t.shape[2:]) model_output, model_var_values = th.split(model_output, C, dim=1) # Learn the variance using the variational bound, but don't let # it affect our mean prediction. frozen_out = th.cat([model_output.detach(), model_var_values], dim=1) terms["vb_loss"] = self._vb_terms_bpd( model=lambda *args, r=frozen_out: r, x_start=x_start, x_t=x_t, t=t, clip_denoised=False, )["output"] if self.loss_type == LossType.RESCALED_MSE: # Divide by 1000 for equivalence with initial implementation. # Without a factor of 1/1000, the VB term hurts the MSE term. terms["vb_loss"] *= self.num_timesteps / 1000.0 target = { ModelMeanType.PREVIOUS_X: self.q_posterior_mean_variance( x_start=x_start, x_t=x_t, t=t )[0], # x_{t-1} ModelMeanType.START_X: x_start, # x_0 ModelMeanType.EPSILON: noise, # eps }[self.model_mean_type] # mean type 决定了在什么上边算 MSE loss # stx() assert model_output.shape == target.shape == x_start.shape ####################################################################################################### ################################### x0_loss 会导致掉点 - 再多渲染一次看看 ############################### ####################################################################################################### # debug_path = os.path.join("/sensei-fs/users/yuanhaoc/gslrm_dit/minLRM/debug_2object", "gaussian_diffusion") # os.makedirs(debug_path, exist_ok=True) # im = target # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(debug_path, "input_gt.png")) # im = model_output # im = rearrange(im, "b v c h w -> (b h) (v w) c").detach().cpu().numpy() # im = (im * 255).clip(0.0, 255.0).astype(np.uint8) # Image.fromarray(im).save(os.path.join(debug_path, "render_input.png")) # stx() ####################################################################################################### ################################################## 调试结束 ########################################## ####################################################################################################### eps_loss = False x0_loss = False if self.model_mean_type == ModelMeanType.START_X and eps_loss: predict_eps = self._predict_eps_from_xstart(x_t=x_t, t=t, pred_xstart=model_output) terms["mse_x0_loss"] = mean_flat((noise - predict_eps) ** 2) elif self.model_mean_type == ModelMeanType.EPSILON and x0_loss: predict_x0 = self._predict_xstart_from_eps(x_t=x_t, t=t, eps=model_output) terms["mse_x0_loss"] = mean_flat((x_start - predict_x0) ** 2) else: terms["mse_x0_loss"] = mean_flat((target - model_output) ** 2) * x0_loss_weight # 前面的 predict variance 已经被设置成 False # terms 字典里没有 vb output_loss_metrics = denoiser_output['loss_metrics'] if "vb_loss" in terms: terms["loss"] = terms["mse_x0_loss"] + terms["vb_loss"] else: terms["loss"] = terms["mse_x0_loss"] # 遍历 output_loss_metrics 的键值对 for key, value in output_loss_metrics.items(): if key == 'loss': terms[key] = terms[key] + value # term 里面的 loss 是全部的 loss 之和 else: terms[key] = value else: raise NotImplementedError(self.loss_type) return terms, denoiser_output # 从一个一维的 numpy 数组中提取出一批索引对应的值，并将这些值转换为 PyTorch 的 Tensor 格式 # 按照选取的 timestep 去取预先计算好的数值 def _extract_into_tensor(arr, timesteps, broadcast_shape): """ Extract values from a 1-D numpy array for a batch of indices. :param arr: the 1-D numpy array. :param timesteps: a tensor of indices into the array to extract. :param broadcast_shape: a larger shape of K dimensions with the batch dimension equal to the length of timesteps. :return: a tensor of shape [batch_size, 1, ...] where the shape has K dims. """ res = th.from_numpy(arr).to(device=timesteps.device)[timesteps].float() while len(res.shape) < len(broadcast_shape): res = res[..., None] return res + th.zeros(broadcast_shape, device=timesteps.device)