""" This code is adapted from Gong et al.'s 2023 DiffuSeq Model: https://github.com/Shark-NLP/DiffuSeq """ import math import numpy as np import torch as th import sys import os import torch.nn from .utils.nn import mean_flat sys.path.append(".") 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 beta_start = scale * 0.0001 beta_end = scale * 0.02 return np.linspace(beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64) elif schedule_name == "cosine": return betas_for_alpha_bar( num_diffusion_timesteps, lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2, ) elif schedule_name == "sqrt": return betas_for_alpha_bar( num_diffusion_timesteps, lambda t: 1 - np.sqrt(t + 0.0001), ) elif schedule_name == "trunc_cos": return betas_for_alpha_bar_left( num_diffusion_timesteps, lambda t: np.cos((t + 0.1) / 1.1 * np.pi / 2) ** 2, ) elif schedule_name == "trunc_lin": scale = 1000 / num_diffusion_timesteps beta_start = scale * 0.0001 + 0.01 beta_end = scale * 0.02 + 0.01 return np.linspace(beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64) elif schedule_name == "pw_lin": scale = 1000 / num_diffusion_timesteps beta_start = scale * 0.0001 + 0.01 beta_mid = scale * 0.0001 # scale * 0.02 beta_end = scale * 0.02 first_part = np.linspace(beta_start, beta_mid, 10, dtype=np.float64) second_part = np.linspace( beta_mid, beta_end, num_diffusion_timesteps - 10, dtype=np.float64 ) return np.concatenate([first_part, second_part]) else: raise NotImplementedError(f"unknown beta schedule: {schedule_name}") def betas_for_alpha_bar_left(num_diffusion_timesteps, alpha_bar, max_beta=0.999): """ Create a beta schedule that discretizes the given alpha_t_bar function, but shifts towards left interval starting from 0 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 = [] betas.append(min(1 - alpha_bar(0), max_beta)) for i in range(num_diffusion_timesteps - 1): 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) 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) class GaussianDiffusion: """ Utilities for training and sampling diffusion models. """ def __init__( self, *, betas, predict_xstart, rescale_learned_sigmas, learn_sigmas, sigma_small, use_kl, one_noise_step, nll_in_loss, mask_padding, rescale_timesteps=False, ): self.rescale_timesteps = rescale_timesteps self.predict_xstart = predict_xstart self.rescale_learned_sigmas = rescale_learned_sigmas self.learn_sigmas = learn_sigmas self.sigma_small = sigma_small self.use_kl = use_kl self.one_noise_step = one_noise_step self.nll_in_loss = nll_in_loss self.mask_padding = mask_padding # Use float64 for accuracy. betas = np.array(betas, dtype=np.float64) # shape [diffusion_steps] self.betas = betas assert len(betas.shape) == 1, "betas must be 1-D" assert (betas > 0).all() and (betas <= 1).all() self.num_timesteps = int(betas.shape[0]) alphas = 1.0 - betas self.alphas_cumprod = np.cumprod(alphas, axis=0) # will approximate 0 self.alphas_cumprod_prev = np.append( 1.0, self.alphas_cumprod[:-1] ) # shifted one to the right self.alphas_cumprod_next = np.append( self.alphas_cumprod[1:], 0.0 ) # shifted one to the left 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) 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) ) # log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain. self.posterior_log_variance_clipped = np.log( np.append(self.posterior_variance[1], self.posterior_variance[1:]) ) 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) ) self.mapping_func = None # implement in train main() self.add_mask_noise = False # TODO def training_losses(self, model, *args, **kwargs): self.model = model return self.training_losses_seq2seq(model, *args, **kwargs) 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 ) 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) def _scale_timesteps(self, t): if self.rescale_timesteps: return t.float() * (1000.0 / self.num_timesteps) return t 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 q_sample(self, x_start, t, noise=None, mask=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. :param mask: anchoring masked position :return: A noisy version of x_start. """ if noise is None: noise = th.randn_like(x_start) assert noise.shape == x_start.shape x_t = ( _extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start # mu * x_0 + _extract_into_tensor( self.sqrt_one_minus_alphas_cumprod, t, x_start.shape ) # sd * noise * noise ) if mask == None: return x_t else: mask = th.broadcast_to(mask.unsqueeze(dim=-1), x_start.shape) return th.where(mask == 0, x_start, x_t) 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) """ 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 def p_mean_variance( self, model, x, sn_input_ids_emb, pos_enc, mask_sn_padding, mask_transformer_att, t, clip_denoised=True, denoised_fn=None, model_kwargs=None, subwords_list=None, atten_vis=None, atten_vis_fn=None, atten_vis_path=None, batch_idx=None, rank=None, atten_vis_sp=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; the noised input where the embedded condition/sn was not noised and the target/sp was replaced with Gaussian noise; at time step t :param sn_input_ids_emb: the BERT embeddings :param pos_enc: the positional embeddings :param mask_sn_padding: :param mask_transformer_att: the attention mask for the transformer :param t: a 1-D Tensor of timesteps. :param clip_denoised: if True, clip the denoised signal into [-1, 1]. :param denoised_fn: if not None, a function which applies to the x_start prediction before it is used to sample. Applies before clip_denoised. :param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning. :param subwords_list: list of list containing the subwortokens of each instance (for attention visualization) :param atten_vis: bool: if True, attention is visualized (heatmaps) :param atten_vis_fn: the attention visualization function :param atten_vis_path: the path where to save the heatmaps to :param batch_idx: index of the current batch :paran rank: if parallel processing, the GPU index :param atten_vis_sp: bool: save attention scores of the last timestep :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 = {} B, C = x.size(0), x.size(-1) assert t.shape == (B,) if not atten_vis and not atten_vis_sp: model_output = model( x=x, ts=self._scale_timesteps(t), sn_input_ids_emb=sn_input_ids_emb, pos_enc=pos_enc, attention_mask=mask_transformer_att, **model_kwargs, ) else: # visualise the attention (heatmaps) if atten_vis: model_output, attention_scores = model( x=x, ts=self._scale_timesteps(t), sn_input_ids_emb=sn_input_ids_emb, pos_enc=pos_enc, attention_mask=mask_transformer_att, atten_vis=atten_vis, **model_kwargs, ) # visualise attention for last denoising step if t[0] % 200 == 0: atten_vis_fn( attention_scores=attention_scores, subwords_list=subwords_list, batch_idx=batch_idx, path_to_dir=atten_vis_path, denoising_step=t[0].item(), aggregate=True, rank=rank, ) if atten_vis_sp: model_output, attention_scores = model( x=x, ts=self._scale_timesteps(t), sn_input_ids_emb=sn_input_ids_emb, pos_enc=pos_enc, attention_mask=mask_transformer_att, atten_vis=True, **model_kwargs, ) if t[0] == 0: out_path_heatmaps_sp = os.path.join(atten_vis_path, "heatmaps_sps") if not os.path.exists(out_path_heatmaps_sp): os.makedirs(out_path_heatmaps_sp) filename = f"att_scores_rank{rank}_batch{batch_idx}.pt" path_to_file = os.path.join(out_path_heatmaps_sp, filename) torch.save(attention_scores, path_to_file) model_variance = np.append(self.posterior_variance[1], self.betas[1:]) model_log_variance = np.log(np.append(self.posterior_variance[1], self.betas[1:])) model_variance = _extract_into_tensor(model_variance, t, x.shape) model_log_variance = _extract_into_tensor(model_log_variance, t, x.shape) # The denoised_fn is applied to x_start (the model output) before it is used for sampling def process_xstart(x): """here x is the model output""" if denoised_fn is not None: # print(denoised_fn) x = denoised_fn(x, t) if clip_denoised: return x.clamp(-1, 1) return x if self.predict_xstart: # the denoised fn is applied to the model output pred_xstart = process_xstart(model_output) else: ### model is used to predict eps pred_xstart = process_xstart( self._predict_xstart_from_eps(x_t=x, t=t, eps=model_output) ) # this is the mean of the posterior distribution q(x_{t-1} | x_t, x_0), estimated from x_t, which is the noised # input, and pred_xstart, which is what the model predicted to be x_0 from the noised input x_noised/x_t model_mean, _, _ = self.q_posterior_mean_variance(x_start=pred_xstart, x_t=x, t=t) assert model_mean.shape == model_log_variance.shape == pred_xstart.shape == x.shape return { "mean": model_mean, "variance": model_variance, "log_variance": model_log_variance, "pred_xstart": pred_xstart, } def p_sample( self, model, x, sn_input_ids_emb, pos_enc, mask_sn_padding, mask_transformer_att, t, clip_denoised=True, denoised_fn=None, model_kwargs=None, top_p=None, mask=None, x_start=None, subwords_list=None, atten_vis=None, atten_vis_fn=None, atten_vis_path=None, batch_idx=None, rank=None, atten_vis_sp=None, ): """ Sample x_{t-1} from the model at the given timestep. :param model: the model to sample from; the transformer model that learned the denoising :param x: the current tensor at x_{t-1}. :param sn_input_ids_emb: the BERT embeddings :param pos_enc: the positional embeddings :param mask_sn_padding: :param mask_transformer_att: the attention mask for the transformer :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 denoised_fn: if not None, a function which applies to the x_start prediction before it is used to sample. :param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning. :param top_p: :param mask: anchoring masked position to x_start :param x_start: :param subwords_list: list of list containing the subwortokens of each instance (for attention visualization) :param atten_vis: bool: if True, attention is visualized (heatmaps) :param atten_vis_fn: the attention visualization function :param atten_vis_path: the path where to save the heatmaps to :param batch_idx: index of the current batch :paran rank: if parallel processing, the GPU index :param atten_vis_sp: bool: save attention scores of the last timestep :return: a dict containing the following keys: - 'sample': a random sample from the model. - 'pred_xstart': a prediction of x_0. """ out = self.p_mean_variance( model=model, x=x, sn_input_ids_emb=sn_input_ids_emb, pos_enc=pos_enc, mask_sn_padding=mask_sn_padding, mask_transformer_att=mask_transformer_att, t=t, clip_denoised=clip_denoised, denoised_fn=denoised_fn, model_kwargs=model_kwargs, subwords_list=subwords_list, atten_vis=atten_vis, atten_vis_fn=atten_vis_fn, atten_vis_path=atten_vis_path, batch_idx=batch_idx, rank=rank, atten_vis_sp=atten_vis_sp, ) if top_p is not None and top_p > 0: # print('top_p sampling') noise = th.randn_like(x) replace_mask = th.abs(noise) > top_p while replace_mask.any(): noise[replace_mask] = th.randn_like(noise[replace_mask]) replace_mask = th.abs(noise) > top_p assert (th.abs(noise) <= top_p).all() else: noise = th.randn_like(x) nonzero_mask = ( (t != 0).float().view(-1, *([1] * (len(x.shape) - 1))) ) # no noise when t == 0 sample = out["mean"] + nonzero_mask * th.exp(0.5 * out["log_variance"]) * noise if mask == None: pass else: # the original embedding for the sn, and the predicted sample for the sp sample = th.where(mask == 0, x_start, sample) return { "sample": sample, "pred_xstart": out["pred_xstart"], "greedy_mean": out["mean"], "out": out, } def p_sample_loop( self, model, shape, noise=None, sn_input_ids_emb=None, pos_enc=None, mask_sn_padding=None, mask_transformer_att=None, clip_denoised=True, denoised_fn=None, model_kwargs=None, device=None, progress=False, top_p=None, clamp_step=None, clamp_first=None, mask=None, x_start=None, subwords_list=None, atten_vis=None, atten_vis_fn=None, atten_vis_path=None, batch_idx=None, gap=1, rank=None, atten_vis_sp=None, ): """ Generate samples from the model. :param model: the transformer model that was trained to learn the denoising :param shape: the shape of the samples, (N, C, H, W). :param noise: the Gaussian noise that should be denoised at inference (the replaced word ID emb) :param sn_input_ids_emb: the BERT embeddings :param pos_enc: the positional embeddings :param mask_sn_padding: :param mask_transformer_att: the attention mask for the transformer :param clip_denoised: if True, clip x_start predictions to [-1, 1]. :param denoised_fn: if not None, a function which applies to the x_start prediction before it is used to sample. :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. :param top_p: :param clamp_step: in clamp_first mode, choose end clamp step, otherwise starting clamp step :param clamp_first: bool, clamp_first mode :param mask: anchoring masked position to x_start :param x_start: the word ID embedding before replaced by noise :param subwords_list: list of list containing the subwortokens of each instance (for attention visualization) :param atten_vis: bool: if True, attention is visualized (heatmaps) :param atten_vis_fn: the attention visualization function :param atten_vis_path: the path where to save the heatmaps to :param batch_idx: index of the current batch :param gap: :paran rank: if parallel processing, the GPU index :param atten_vis_sp: bool: save attention scores of the last timestep :return: a non-differentiable batch of samples. """ final = [] for sample in self.p_sample_loop_progressive( model, shape, noise=noise, sn_input_ids_emb=sn_input_ids_emb, pos_enc=pos_enc, mask_sn_padding=mask_sn_padding, mask_transformer_att=mask_transformer_att, clip_denoised=clip_denoised, denoised_fn=denoised_fn, model_kwargs=model_kwargs, device=device, progress=progress, top_p=top_p, clamp_step=clamp_step, clamp_first=clamp_first, mask=mask, x_start=x_start, subwords_list=subwords_list, atten_vis=atten_vis, atten_vis_fn=atten_vis_fn, atten_vis_path=atten_vis_path, batch_idx=batch_idx, rank=rank, atten_vis_sp=atten_vis_sp, ): final.append(sample["sample"]) return final def p_sample_loop_progressive( self, model, shape, noise=None, sn_input_ids_emb=None, pos_enc=None, mask_sn_padding=None, mask_transformer_att=None, clip_denoised=True, denoised_fn=None, model_kwargs=None, device=None, progress=False, top_p=None, clamp_step=None, clamp_first=None, mask=None, x_start=None, subwords_list=None, atten_vis=None, atten_vis_fn=None, atten_vis_path=None, batch_idx=None, rank=None, atten_vis_sp=None, ): """ 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)) # noise/sample_x is the input that was noised: the concatenated sn-sp embedding where the sp was completely # replaced with Gaussian noise from the standard normal distribution if noise is not None: sample_x = noise else: sample_x = th.randn(*shape, device=device) # the number of diffusion steps in reverse order indices = list(range(self.num_timesteps))[::-1] if progress: # Lazy import so that we don't depend on tqdm. from tqdm.auto import tqdm indices = tqdm(indices) # denoising from the number of diffusion steps T to t=0 for i in indices: # from T to 0 t = th.tensor([i] * shape[0], device=device) if not clamp_first: if i > clamp_step: denoised_fn_cur = None else: denoised_fn_cur = denoised_fn else: if i >= clamp_step: denoised_fn_cur = denoised_fn else: denoised_fn_cur = None with th.no_grad(): out = self.p_sample( model=model, x=sample_x, sn_input_ids_emb=sn_input_ids_emb, pos_enc=pos_enc, mask_sn_padding=mask_sn_padding, mask_transformer_att=mask_transformer_att, t=t, clip_denoised=clip_denoised, denoised_fn=denoised_fn_cur, model_kwargs=model_kwargs, top_p=top_p, mask=mask, subwords_list=subwords_list, x_start=x_start, atten_vis=atten_vis, atten_vis_fn=atten_vis_fn, atten_vis_path=atten_vis_path, batch_idx=batch_idx, rank=rank, atten_vis_sp=atten_vis_sp, ) yield out sample_x = out["sample"] def _get_x_start(self, x_start_mean, std): """ Word embedding projection from {Emb(w)} to {x_0} :param x_start_mean: word embedding :return: x_0 """ noise = th.randn_like(x_start_mean) assert noise.shape == x_start_mean.shape # print(x_start_mean.device, noise.device) return x_start_mean + std * noise def _token_discrete_loss(self, x_t, get_logits, input_ids, mask=None, truncate=False, t=None): """ the loss of -log p(w|z_0) :param x_start_mean: word embedding :return: x_0 """ reshaped_x_t = x_t logits = get_logits(reshaped_x_t) # shape [microbatch size, seq_len, vocabulary] # print(logits.shape) loss_fct = th.nn.CrossEntropyLoss(reduction="none") decoder_nll = loss_fct(logits.view(-1, logits.size(-1)), input_ids.view(-1)).view( input_ids.shape ) if mask != None: decoder_nll *= mask # print(decoder_nll.shape) if mask != None: decoder_nll = decoder_nll.sum(dim=-1) / mask.sum(dim=-1) else: decoder_nll = decoder_nll.mean(dim=-1) return decoder_nll def _x0_helper(self, model_output, x, t): if self.predict_xstart: pred_xstart = model_output pred_prev, _, _ = self.q_posterior_mean_variance(x_start=pred_xstart, x_t=x, t=t) else: # predict eps pred_xstart = self._predict_xstart_from_eps(x_t=x, t=t, eps=model_output) pred_prev, _, _ = self.q_posterior_mean_variance(x_start=pred_xstart, x_t=x, t=t) return {"pred_xprev": pred_prev, "pred_xstart": pred_xstart} def training_losses_seq2seq( self, model, # the transformer model t, # the number of noise adding steps for each instance in the microbatch sn_sp_repr, mask, sn_input_ids, indices_pos_enc, mask_sn_padding, mask_transformer_att, noise=None, ): """ Compute training losses for a single timestep. :param model: the transformer model :param t: a batch of timestep indices. :param sn_sp_repr: the word IDs of sn and sp :param mask: masking the sn :param sn_input_ids: the tokenizer input IDs :param indices_pos_enc: the indices for pos enc :param_mask_sn_padding: :param mask_transformer_att: the transformer att :param model_kwargs: if not None, a dict of extra keyword arguments to pass to the model. This can be used for conditioning. :param noise: if specified, the specific Gaussian noise to try to remove. :return: a dict with the key "loss" containing a tensor of shape [N]. Some mean or variance settings may also have other keys. """ microbatch_size, seq_len = sn_sp_repr.shape # get the word ID embedding, BERT embedding, positional embedding sn_sp_emb, pos_enc, sn_input_ids_emb = model.model.module.get_embeds( sn_sp_repr=sn_sp_repr, sn_input_ids=sn_input_ids, indices_pos_enc=indices_pos_enc, ) # get the standard deviation, shape [microbatch, args.seq_len, hidden_size=768] std = _extract_into_tensor( self.sqrt_one_minus_alphas_cumprod, th.tensor([0]).to(sn_sp_emb.device), sn_sp_emb.shape ) # map sn_sp_emb to x_start, which is a one-step noised sn_sp_emb (in paper it's z_0) if ( self.one_noise_step ): # this should always be true actually (without it performance is bad) x_start = self._get_x_start(sn_sp_emb, std) else: x_start = sn_sp_emb # sample noise in the same shape as our input if noise is None: noise = th.randn_like(x_start) # get the noised sample x_t, which is still of shape [microbatch, args.seq_len, hidden_size=768] # the condition/sn is not noised (hence the input mask) # each instance in the microbatch receives a different amount of noise (t noising steps, as given in vector t) x_t = self.q_sample( x_start=x_start, t=t, noise=noise, mask=mask, ) terms = {} target = x_start # model_output is of shape [microbatch, args.seq_len, emb_dim=768] model_output = model( x=x_t, ts=self._scale_timesteps(t), sn_input_ids_emb=sn_input_ids_emb, pos_enc=pos_enc, attention_mask=mask_transformer_att, ) assert model_output.shape == target.shape == x_start.shape # Loss 1: Mean Squared Error (MSE) (L_{VLB}) terms["mse"] = mean_flat((target - model_output) ** 2) model_out_x_start = self._x0_helper(model_output, x_t, t)["pred_xstart"] t0_mask = ( t == 0 ) # mask that says true for every instance where no noise was received, i.e. t=0 # MSE between the model output and the embedded input before the one noise step t0_loss = mean_flat((sn_sp_emb - model_out_x_start) ** 2) # update the MSE between the model output and the one-step noised input embeddings with the MSE between the # model output and the embeddings before the one noise step wherever there was no noise received in the noising # process (i.e., wherever t was 0) terms["mse"] = th.where(t0_mask, t0_loss, terms["mse"]) # Loss 2: L_{round} out_mean, _, _ = self.q_mean_variance( x_start, th.LongTensor([self.num_timesteps - 1]).to(x_start.device) ) tT_loss = mean_flat(out_mean**2) # for the NLL losses, we need to convert the model output into logits get_logits = model.model.module.get_logits # Loss 3: L_{EMB} # compute the NLL between the one-noised embeddings and the initial representation (word IDs) # embedding regularisation decoder_nll = self._token_discrete_loss(x_start, get_logits, sn_sp_repr) # unused Loss terms["nll"] = self._token_discrete_loss(model_output, get_logits, sn_sp_repr, mask=mask) # combined loss if self.nll_in_loss: # should be False; model performance drops if nll included terms["loss"] = terms["mse"] + tT_loss + decoder_nll + terms["nll"] else: terms["loss"] = terms["mse"] + tT_loss + decoder_nll return terms def ddim_sample( self, model, x, t, clip_denoised=True, denoised_fn=None, model_kwargs=None, eta=0.0, langevin_fn=None, mask=None, x_start=None, ): """ Sample x_{t-1} from the model using DDIM. Same usage as p_sample(). """ out = self.p_mean_variance( model, x, t, clip_denoised=clip_denoised, denoised_fn=denoised_fn, model_kwargs=model_kwargs, ) # Usually our model outputs epsilon, but we re-derive it # in case we used x_start or x_prev prediction. eps = self._predict_eps_from_xstart(x, t, out["pred_xstart"]) alpha_bar = _extract_into_tensor(self.alphas_cumprod, t, x.shape) alpha_bar_prev = _extract_into_tensor(self.alphas_cumprod_prev, t, x.shape) sigma = ( eta * th.sqrt((1 - alpha_bar_prev) / (1 - alpha_bar)) * th.sqrt(1 - alpha_bar / alpha_bar_prev) ) # Equation 12. noise = th.randn_like(x) mean_pred = ( out["pred_xstart"] * th.sqrt(alpha_bar_prev) + th.sqrt(1 - alpha_bar_prev - sigma**2) * eps ) nonzero_mask = ( (t != 0).float().view(-1, *([1] * (len(x.shape) - 1))) ) # no noise when t == 0 # print(sigma.mean()) sample = mean_pred + nonzero_mask * sigma * noise if langevin_fn: print(t.shape) sample = langevin_fn(sample, mean_pred, sigma, self.alphas_cumprod_prev[t[0]], t, x) if mask == None: pass else: sample = th.where(mask == 0, x_start, sample) return {"sample": sample, "pred_xstart": out["pred_xstart"]} def ddim_reverse_sample( self, model, x, t, clip_denoised=True, denoised_fn=None, model_kwargs=None, eta=0.0, ): """ Sample x_{t+1} from the model using DDIM reverse ODE. """ assert eta == 0.0, "Reverse ODE only for deterministic path" out = self.p_mean_variance( model, x, t, clip_denoised=clip_denoised, denoised_fn=denoised_fn, model_kwargs=model_kwargs, ) # Usually our model outputs epsilon, but we re-derive it # in case we used x_start or x_prev prediction. eps = ( _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x.shape) * x - out["pred_xstart"] ) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x.shape) alpha_bar_next = _extract_into_tensor(self.alphas_cumprod_next, t, x.shape) # Equation 12. reversed mean_pred = out["pred_xstart"] * th.sqrt(alpha_bar_next) + th.sqrt(1 - alpha_bar_next) * eps return {"sample": mean_pred, "pred_xstart": out["pred_xstart"]} def ddim_sample_loop( self, model, shape, noise=None, clip_denoised=True, denoised_fn=None, model_kwargs=None, device=None, progress=False, top_p=None, clamp_step=None, clamp_first=None, mask=None, x_start=None, gap=1, ): """ Generate samples from the model using DDIM. :param gap: compute ddim sampling for each {gap} step Same usage as p_sample_loop(). """ final = [] for sample in self.ddim_sample_loop_progressive( model, shape, noise=noise, clip_denoised=clip_denoised, denoised_fn=denoised_fn, model_kwargs=model_kwargs, device=device, progress=progress, mask=mask, x_start=x_start, gap=gap, ): final.append(sample["sample"]) return final def ddim_sample_loop_progressive( self, model, shape, noise=None, clip_denoised=True, denoised_fn=None, model_kwargs=None, device=None, progress=False, eta=0.0, langevin_fn=None, mask=None, x_start=None, gap=1, ): """ Use DDIM to sample from the model and yield intermediate samples from each timestep of DDIM. Same usage as p_sample_loop_progressive(). """ if device is None: device = next(model.parameters()).device assert isinstance(shape, (tuple, list)) if noise is not None: sample_x = noise else: sample_x = th.randn(*shape, device=device) indices = list(range(self.num_timesteps))[::-1][::gap] if progress: # Lazy import so that we don't depend on tqdm. from tqdm.auto import tqdm indices = tqdm(indices) for i in indices: t = th.tensor([i] * shape[0], device=device) with th.no_grad(): out = self.ddim_sample( model, sample_x, t, clip_denoised=clip_denoised, denoised_fn=denoised_fn, model_kwargs=model_kwargs, mask=mask, x_start=x_start, ) yield out sample_x = out["sample"] 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.expand(broadcast_shape) def space_timesteps(num_timesteps, section_counts): """ Create a list of timesteps to use from an original diffusion process, given the number of timesteps we want to take from equally-sized portions of the original process. For example, if there's 300 timesteps and the section counts are [10,15,20] then the first 100 timesteps are strided to be 10 timesteps, the second 100 are strided to be 15 timesteps, and the final 100 are strided to be 20. If the stride is a string starting with "ddim", then the fixed striding from the DDIM paper is used, and only one section is allowed. :param num_timesteps: the number of diffusion steps in the original process to divide up. :param section_counts: either a list of numbers, or a string containing comma-separated numbers, indicating the step count per section. As a special case, use "ddimN" where N is a number of steps to use the striding from the DDIM paper. :return: a set of diffusion steps from the original process to use. """ if isinstance(section_counts, str): if section_counts.startswith("ddim"): desired_count = int(section_counts[len("ddim") :]) for i in range(1, num_timesteps): if len(range(0, num_timesteps, i)) == desired_count: return set(range(0, num_timesteps, i)) raise ValueError(f"cannot create exactly {num_timesteps} steps with an integer stride") section_counts = [int(x) for x in section_counts.split(",")] size_per = num_timesteps // len(section_counts) extra = num_timesteps % len(section_counts) start_idx = 0 all_steps = [] for i, section_count in enumerate(section_counts): size = size_per + (1 if i < extra else 0) if size < section_count: raise ValueError(f"cannot divide section of {size} steps into {section_count}") if section_count <= 1: frac_stride = 1 else: frac_stride = (size - 1) / (section_count - 1) cur_idx = 0.0 taken_steps = [] for _ in range(section_count): taken_steps.append(start_idx + round(cur_idx)) cur_idx += frac_stride all_steps += taken_steps start_idx += size return set(all_steps) class SpacedDiffusion(GaussianDiffusion): """ A diffusion process which can skip steps in a base diffusion process. :param use_timesteps: a collection (sequence or set) of timesteps from the original diffusion process to retain. :param kwargs: the kwargs to create the base diffusion process. """ def __init__(self, use_timesteps, **kwargs): self.use_timesteps = set(use_timesteps) self.timestep_map = [] self.original_num_steps = len(kwargs["betas"]) # print(kwargs.keys()) base_diffusion = GaussianDiffusion(**kwargs) # pylint: disable=missing-kwoa last_alpha_cumprod = 1.0 new_betas = [] for i, alpha_cumprod in enumerate(base_diffusion.alphas_cumprod): if i in self.use_timesteps: new_betas.append(1 - alpha_cumprod / last_alpha_cumprod) last_alpha_cumprod = alpha_cumprod self.timestep_map.append(i) kwargs["betas"] = np.array(new_betas) super().__init__(**kwargs) def p_mean_variance(self, model, *args, **kwargs): # pylint: disable=signature-differs # print('called p_mean_var') return super().p_mean_variance(self._wrap_model(model), *args, **kwargs) def training_losses(self, model, *args, **kwargs): # pylint: disable=signature-differs # print('called training_losses') return super().training_losses(self._wrap_model(model), *args, **kwargs) def _wrap_model(self, model): if isinstance(model, _WrappedModel): return model return _WrappedModel( model, self.timestep_map, self.rescale_timesteps, self.original_num_steps ) def _scale_timesteps(self, t): # Scaling is done by the wrapped model. return t class _WrappedModel: def __init__(self, model, timestep_map, rescale_timesteps, original_num_steps): self.model = model self.timestep_map = timestep_map self.rescale_timesteps = rescale_timesteps self.original_num_steps = original_num_steps def __call__(self, x, ts, **kwargs): # print(ts) map_tensor = th.tensor(self.timestep_map, device=ts.device, dtype=ts.dtype) new_ts = map_tensor[ts] # print(new_ts) if self.rescale_timesteps: new_ts = new_ts.float() * (1000.0 / self.original_num_steps) # temp = self.model(x, new_ts, **kwargs) # print(temp.shape) # return temp # print(new_ts) return self.model(x, new_ts, **kwargs)