biAyddlZddlmZddlmZmZmZmZddlZ ddl Z ddl m Z m Z ddlmZddlmZdd lmZmZeGd d eZ dd Zd ZGddee Zy)N) dataclass)ListOptionalTupleUnion) ConfigMixinregister_to_config) BaseOutput) randn_tensor)KarrasDiffusionSchedulersSchedulerMixincXeZdZUdZej ed<dZeej ed<y)DDPMParallelSchedulerOutputaq Output class for the scheduler's `step` function output. Args: prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the denoising loop. pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): The predicted denoised sample `(x_{0})` based on the model output from the current timestep. `pred_original_sample` can be used to preview progress or for guidance. prev_sampleNpred_original_sample) __name__ __module__ __qualname____doc__torchTensor__annotations__rrh/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/schedulers/scheduling_ddpm_parallel.pyrrs' 37(5<<07rrc $|dk(rd}n|dk(rd}ntd|g}t|D]<}||z }|dz|z }|jtd||||z z |>t j |tj S)a Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [0,1]. Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up to that part of the diffusion process. Args: num_diffusion_timesteps (`int`): the number of betas to produce. max_beta (`float`): the maximum beta to use; use values lower than 1 to prevent singularities. alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. Choose from `cosine` or `exp` Returns: betas (`np.ndarray`): the betas used by the scheduler to step the model outputs cosinecftj|dzdz tjzdz dzS)NgMb?gT㥛 ?r)mathcospits r alpha_bar_fnz)betas_for_alpha_bar..alpha_bar_fnKs-88QY%/$''9A=>!C Crexpc2tj|dzS)Ng()r!r'r$s rr&z)betas_for_alpha_bar..alpha_bar_fnPs88AI& &rz"Unsupported alpha_transform_type: r dtype) ValueErrorrangeappendminrtensorfloat32)num_diffusion_timestepsmax_betaalpha_transform_typer&betasit1t2s rbetas_for_alpha_barr82s.x' D  & '=>R=STUU E * +M ( (!e. . S\"- R0@@@(KLM <<U]] 33rc(d|z }tj|d}|j}|dj}|dj}||z}||||z z z}|dz}|dd|ddz }tj|dd|g}d|z }|S)a. Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) Args: betas (`torch.Tensor`): the betas that the scheduler is being initialized with. Returns: `torch.Tensor`: rescaled betas with zero terminal SNR ?rdimrr N)rcumprodsqrtclonecat)r4alphasalphas_cumprodalphas_bar_sqrtalphas_bar_sqrt_0alphas_bar_sqrt_T alphas_bars rrescale_zero_terminal_snrrH_s5[F]]6q1N$))+O(*002'+113((O(,=@Q,QRRO!!#J ^j"o -F YY 1Q0 1F JE Lrc NeZdZdZeDcgc]}|j c}}ZdZdZe d)de de de de d e eej ee fd e d ed e d ede de de de de defdZd*dej*de e dej*fdZ d+de e dee ej.fde ee fdZd,dZdej*dej*fdZ d-dej*de dej*d edeeeff d!Zdej*dee dej*dej*fd"Zd#ej*d$ej*dej>dej*fd%Z dej*d$ej*dej>dej*fd&Z!d'Z"d(Z#ycc}}w).DDPMParallelSchedulera. Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and Langevin dynamics sampling. [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions. For more details, see the original paper: https://huggingface.co/papers/2006.11239 Args: num_train_timesteps (`int`): number of diffusion steps used to train the model. beta_start (`float`): the starting `beta` value of inference. beta_end (`float`): the final `beta` value. beta_schedule (`str`): the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from `linear`, `scaled_linear`, `squaredcos_cap_v2` or `sigmoid`. trained_betas (`np.ndarray`, optional): option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. variance_type (`str`): options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`, `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`. clip_sample (`bool`, default `True`): option to clip predicted sample for numerical stability. clip_sample_range (`float`, default `1.0`): the maximum magnitude for sample clipping. Valid only when `clip_sample=True`. prediction_type (`str`, default `epsilon`, optional): prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 https://imagen.research.google/video/paper.pdf) thresholding (`bool`, default `False`): whether to use the "dynamic thresholding" method (introduced by Imagen, https://huggingface.co/papers/2205.11487). Note that the thresholding method is unsuitable for latent-space diffusion models (such as stable-diffusion). dynamic_thresholding_ratio (`float`, default `0.995`): the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen (https://huggingface.co/papers/2205.11487). Valid only when `thresholding=True`. sample_max_value (`float`, default `1.0`): the threshold value for dynamic thresholding. Valid only when `thresholding=True`. timestep_spacing (`str`, default `"leading"`): The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. steps_offset (`int`, default `0`): An offset added to the inference steps, as required by some model families. rescale_betas_zero_snr (`bool`, defaults to `False`): Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and dark samples instead of limiting it to samples with medium brightness. Loosely related to [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). r FNnum_train_timesteps beta_startbeta_end beta_schedule trained_betas variance_type clip_sampleprediction_type thresholdingdynamic_thresholding_ratioclip_sample_rangesample_max_valuetimestep_spacing steps_offsetrescale_betas_zero_snrc|+tj|tj|_n|dk(r-tj|||tj|_n|dk(r6tj|dz|dz|tjdz|_np|dk(rt ||_nZ|dk(r;tjdd |}tj |||z z|z|_nt|d |j|rt|j|_d |jz |_ tj|jd |_ tjd |_ d |_d|_d|_tj"t%j&d |dddj)|_||_y)Nr)linear scaled_linear?rsquaredcos_cap_v2sigmoidiz is not implemented for r:rr;Fr=)rr/r0r4linspacer8r_NotImplementedError __class__rHrBr>rConeinit_noise_sigmacustom_timestepsnum_inference_steps from_numpynparangecopy timestepsrP)selfrKrLrMrNrOrPrQrRrSrTrUrVrWrXrYr4s r__init__zDDPMParallelScheduler.__init__s(  $m5==IDJ h & H>QY^YfYfgDJ o - C3H[chcpcpquvvDJ 1 1,-@ADJ i 'NN2q*=>Eu-J1FG*TDJ%7OPTP^P^O_&`a a "24::>DJDJJ& #mmDKKQ?<<$!$!&#' ))"))A7J*KDbD*Q*V*V*XY*rsampletimestepreturnc|S)a Ensures interchangeability with schedulers that need to scale the denoising model input depending on the current timestep. Args: sample (`torch.Tensor`): The input sample. timestep (`int`, *optional*): The current timestep in the diffusion chain. Returns: `torch.Tensor`: A scaled input sample. r)rmrorps rscale_model_inputz'DDPMParallelScheduler.scale_model_inputs  rrgdevicerlc l| | td|tdt|D]}||||dz k\std|d|jjk\r#td|jjdt j |t j}d |_nZ||jjkDr=td |d |jjd |jjd ||_ d|_|jjdk(rot jd|jjdz |jdddjjt j}nn|jjdk(r|jj|jz}t jd||zjdddjjt j}||jj z }n|jjdk(r|jj|jz }t jt j|jjd| jt j}|dz}n"t|jjdt#j$|j'||_y)a, Sets the discrete timesteps used for the diffusion chain (to be run before inference). Args: num_inference_steps (`int`): The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` must be `None`. device (`str` or `torch.device`, *optional*): The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. timesteps (`List[int]`, *optional*): Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed, `num_inference_steps` must be `None`. NzACan only pass one of `num_inference_steps` or `custom_timesteps`.r z/`custom_timesteps` must be in descending order.rz=`timesteps` must start before `self.config.train_timesteps`: .r)Tz`num_inference_steps`: z6 cannot be larger than `self.config.train_timesteps`: zG as the unet model trained with this scheduler can only handle maximal z timesteps.Frar=leadingtrailingzY is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'.)r+r,lenconfigrKriarrayint64rfrgrWraroundrkastyperjrXrrhtorl)rmrgrtrlr5 step_ratios r set_timestepsz#DDPMParallelScheduler.set_timestepss*  *y/D`a a  1c)n- XQ<9QU#33$%VWW X|t{{>>> STXT_T_TsTsSttuv"((;I$(D !"T[[%D%DD -.A-BC 7789 $ ? ?@ M (;D $$)D !{{++z9KK4;;#B#BQ#FH[\UWTrT#TVVBHH%  --:![[<<@X@XX  YYq*=>KRRTUYWYUYZ__ahhikiqiqr T[[555 --;![[<?.P;;xU3   KK55M M )H$!/ /yy*Hyyx0Hm +%H/ /yy0Hi '% % o -ii)Gii/G&*a/Dg~TW(< 1, then we threshold xt0 to the range [-s, s] and then divide by s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing pixels from saturation at each step. We find that dynamic thresholding results in significantly better photorealism as well as better image-text alignment, especially when using very large guidance weights." https://huggingface.co/papers/2205.11487 r r;)r.max)r*shaperr0float64floatreshaperiprodabsquantilerzrTrrV unsqueezer)rmror* batch_sizechannelsremaining_dims abs_sampless r_threshold_samplez'DDPMParallelScheduler._threshold_sampleys 06 - H~  6 6\\^F Hrww~7N,NOZZ\ NN:t{{'M'MST U KK 1$++66  KKNVaR+a/ HF~F5! r model_output return_dictcX|}|j|}|jd|jddzk(r7|jdvr)tj||jdd\}}nd}|j |} |dk\r|j |n |j } d| z } d| z } | | z } d| z }|jjdk(r|| dz|zz | dzz }nj|jjd k(r|}nN|jjd k(r| dz|z| dz|zz }n#td |jjd |jjr|j|}nQ|jjr;|j|jj |jj}| dz|z| z }| dz| z| z }||z||zz}d}|dkDr|j}t!|j|||j" }|jdk(r|j%|||z}nW|jdk(r/|j%||}tj&d|z|z}n|j%||dz|z}||z}|s||fSt)||S)a Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion process from the learned model outputs (most often the predicted noise). Args: model_output (`torch.Tensor`): direct output from learned diffusion model. timestep (`int`): current discrete timestep in the diffusion chain. sample (`torch.Tensor`): current instance of sample being created by diffusion process. generator: random number generator. return_dict (`bool`): option for returning tuple rather than DDPMParallelSchedulerOutput class Returns: [`~schedulers.scheduling_utils.DDPMParallelSchedulerOutput`] or `tuple`: [`~schedulers.scheduling_utils.DDPMParallelSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. r rrrr;Nrepsilonr]ro v_predictionprediction_type given as zM must be one of `epsilon`, `sample` or `v_prediction` for the DDPMScheduler.) generatorrtr*r)rr)rr)rrrPrsplitrCrdrzrRr+rSrrQrrUrtr r*rr'r)rmrrprorrr%rrrr beta_prod_tbeta_prod_t_prevcurrent_alpha_trrpred_original_sample_coeffcurrent_sample_coeffpred_prev_samplerrtvariance_noises rstepzDDPMParallelScheduler.steps4 ''*   a FLLOa$7 7D&QU`%`".37:JJ[X68LLOcflOll q5!((F)""ilN`N`N!!%66--aDV-WZhh##6--aDV-W 99S8^4~E ..qEW.X\__cqq+h6 $  +7G^rssrcX|}|jr |jn|jj}||jj|zz }|jdgdg|jdz z}|jdgdg|jdz z}|j d|j ddzk(r7|j dvr)tj||j dd\}}n |jj|j|_ |j|}|jtj|d} tjd| |dk<d|z } d| z } || z } d| z } |jjd k(r|| d z|zz |d zz }nj|jjd k(r|}nN|jjd k(r|d z|z| d z|zz }n#td |jjd|jj r|j#|}nQ|jj$r;|j'|jj( |jj(}| d z| z| z }| d z| z| z }||z||zz}|S)a Batched version of the `step` function, to be able to reverse the SDE for multiple samples/timesteps at once. Also, does not add any noise to the predicted sample, which is necessary for parallel sampling where the noise is pre-sampled by the pipeline. Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion process from the learned model outputs (most often the predicted noise). Args: model_output (`torch.Tensor`): direct output from learned diffusion model. timesteps (`List[int]`): current discrete timesteps in the diffusion chain. This is now a list of integers. sample (`torch.Tensor`): current instance of sample being created by diffusion process. Returns: `torch.Tensor`: sample tensor at previous timestep. r=r rrr;rrr:rr]rorrzU must be one of `epsilon`, `sample` or `v_prediction` for the DDPMParallelScheduler.)rgrzrKviewndimrrPrrrCrrtclipr/rRr+rSrrQrrU)rmrrlror%rgrrrrrrrrrrrrs rbatch_step_no_noisez)DDPMParallelScheduler.batch_step_no_noises0 :>:R:Rd66X\XcXcXwXwT[[448KKK AFF2 8! 1 1A 56 8RBA3,*;*;a*?#@B   a FLLOa$7 7D&QU`%`".37:JJ[X68LLOcflOllroriginal_samplesnoisec|jj|j|_|jj|j}|j|j}||dz}|j }t |j t |j kr=|jd}t |j t |j kr=d||z dz}|j }t |j t |j kr=|jd}t |j t |j kr=||z||zz}|SN)rtr)r]r=r rCrrtr*flattenryrr)rmrrrlrCsqrt_alpha_prodsqrt_one_minus_alpha_prod noisy_sampless r add_noisezDDPMParallelScheduler.add_noiseNsa#1144>qAFF>L E$("%BF* (",1#&"% )',!3+ 3+3+ 3+  3+  bjj$u+&= >? 3+3+3+3+3+%*3+!3+ 3+3+3+ !%!3+3+l  Y^YeYe(.2+/)- H@%c]H@c5<<'(H@DI& H@V&R L atllatat at  at *E1 2atFN llN 9N  N  N b,,||??   65<< QVQ`Q`ejeqeq&/ S?sF!rJ)g+?r)r! dataclassesrtypingrrrrnumpyrirconfiguration_utilsr r utilsr utils.torch_utilsr scheduling_utilsrrrr8rHrJrrrrsd" !// A,G 8*8 8(!)4Z!HFNKFr