biT|ddlZddlmZmZmZmZddlZddlZddl m Z m Z ddl m Z mZmZ d dZGdd ee Zy) N)ListOptionalTupleUnion) ConfigMixinregister_to_config)KarrasDiffusionSchedulersSchedulerMixinSchedulerOutputc $|dk(rd}n|dk(rd}ntd|g}t|D]<}||z }|dz|z }|jtd||||z z |>t j |tj S)a Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [0,1]. Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up to that part of the diffusion process. Args: num_diffusion_timesteps (`int`): the number of betas to produce. max_beta (`float`): the maximum beta to use; use values lower than 1 to prevent singularities. alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. Choose from `cosine` or `exp` Returns: betas (`np.ndarray`): the betas used by the scheduler to step the model outputs cosinecftj|dzdz tjzdz dzS)NgMb?gT㥛 ?r)mathcospits _/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/schedulers/scheduling_pndm.py alpha_bar_fnz)betas_for_alpha_bar..alpha_bar_fn5s-88QY%/$''9A=>!C Cexpc2tj|dzS)Ng()rrrs rrz)betas_for_alpha_bar..alpha_bar_fn:s88AI& &rz"Unsupported alpha_transform_type: r dtype) ValueErrorrangeappendmintorchtensorfloat32)num_diffusion_timestepsmax_betaalpha_transform_typerbetasit1t2s rbetas_for_alpha_barr+s.x' D  & '=>R=STUU E * +M ( (!e. . S\"- R0@@@(KLM <<U]] 33rceZdZdZeDcgc]}|j c}}ZdZe d!de de de de de e ejee fd ed ed e d e d e fdZd"de de e ej(ffdZ d#dej,de dej,dede eeff dZ d#dej,de dej,dede eeff dZ d#dej,de dej,dede eeff dZdej,dej,fdZdZdej,dej,dej<dej,fdZd Z ycc}}w)$ PNDMSchedulera `PNDMScheduler` uses pseudo numerical methods for diffusion models such as the Runge-Kutta and linear multi-step method. This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic methods the library implements for all schedulers such as loading and saving. Args: num_train_timesteps (`int`, defaults to 1000): The number of diffusion steps to train the model. beta_start (`float`, defaults to 0.0001): The starting `beta` value of inference. beta_end (`float`, defaults to 0.02): The final `beta` value. beta_schedule (`str`, defaults to `"linear"`): The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from `linear`, `scaled_linear`, or `squaredcos_cap_v2`. trained_betas (`np.ndarray`, *optional*): Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. skip_prk_steps (`bool`, defaults to `False`): Allows the scheduler to skip the Runge-Kutta steps defined in the original paper as being required before PLMS steps. set_alpha_to_one (`bool`, defaults to `False`): Each diffusion step uses the alphas product value at that step and at the previous one. For the final step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, otherwise it uses the alpha value at step 0. prediction_type (`str`, defaults to `epsilon`, *optional*): Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process) or `v_prediction` (see section 2.4 of [Imagen Video](https://imagen.research.google/video/paper.pdf) paper). timestep_spacing (`str`, defaults to `"leading"`): The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. steps_offset (`int`, defaults to 0): An offset added to the inference steps, as required by some model families. r Nnum_train_timesteps beta_startbeta_end beta_schedule trained_betasskip_prk_stepsset_alpha_to_oneprediction_typetimestep_spacing steps_offsetc L|+tj|tj|_n|dk(r-tj|||tj|_nk|dk(r6tj|dz|dz|tjdz|_n0|dk(rt ||_nt |d|jd|jz |_tj|jd |_ |rtjdn|jd |_ d|_ d |_ d |_d |_d|_g|_d|_t'j(d |ddd j+|_d|_d|_d|_y) Nrlinear scaled_linear?rsquaredcos_cap_v2z is not implemented for g?r)dim)r!r"r#r'linspacer+NotImplementedError __class__alphascumprodalphas_cumprodfinal_alpha_cumprodinit_noise_sigma pndm_ordercur_model_outputcounter cur_sampleetsnum_inference_stepsnparangecopy _timesteps prk_timestepsplms_timesteps timesteps) selfr.r/r0r1r2r3r4r5r6r7s r__init__zPNDMScheduler.__init__qso  $m5==IDJ h & H>QY^YfYfgDJ o - C3H[chcpcpquvvDJ 1 1,-@ADJ%7OPTP^P^O_&`a aDJJ& #mmDKKQ?8H5<<#4dNaNabcNd !$ !" $( ))A':;DbDAFFH!"rrMdevicec||_|jjdk(r`tjd|jj dz |j jtj|_ nm|jjdk(ry|jj |jz}tjd||zj |_ |xj|jjz c_ n|jjdk(r|jj |jz }tj tj|jj d| dddjtj|_ |xjdzc_ n"t|jjd|jjrstjg|_tj |jdd|jd d|jddgdddj#|_ntj|j|j& dj)d tj*tjd|jj |zd zg|j&z}|ddj)d dddddj#|_|jdd dddj#|_tj |j|j$gjtj}t-j.|j1||_g|_d|_d|_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. device (`str` or `torch.device`, *optional*): The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. r@rr leadingtrailingNr?zY is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'.r)rMconfigr6rNr@r.roundastypeint64rQrOr7rr3arrayrR concatenaterPrSrHrepeattiler! from_numpytorTrLrJrI)rUrMrW step_ratiorRrTs r set_timestepszPNDMScheduler.set_timestepss($7 ;; ' ': 5 At{{>>BDWX^^`gghjhphpq O[[ ) )Y 688Dq"Ddd!K!P!P!RD "&//#2"6"#df  NND$6$68K8K#LMTTUWU]U]^ )))477? !r model_outputtimestepsample return_dictreturnc|jt|jkr+|jjs|j ||||S|j ||||S)aG Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion process from the learned model outputs (most often the predicted noise), and calls [`~PNDMScheduler.step_prk`] or [`~PNDMScheduler.step_plms`] depending on the internal variable `counter`. Args: model_output (`torch.Tensor`): The direct output from learned diffusion model. timestep (`int`): The current discrete timestep in the diffusion chain. sample (`torch.Tensor`): A current instance of a sample created by the diffusion process. return_dict (`bool`): Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. Returns: [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a tuple is returned where the first element is the sample tensor. )rirjrkrl)rJlenrRr]r3step_prk step_plms)rUrirjrkrls rstepzPNDMScheduler.stepsY8 <<#d001 1$++:T:T==lXV\ju=v v>>|hW]kv>w wrc<|j td|jdzrdn%|jj|jzdz}||z }|j |jdzdz}|jdzdk(r;|xj d|zz c_|jj|||_ n|jdz dzdk(r|xj d|zz c_n\|jdz dzdk(r|xj d|zz c_n.|jdz dzdk(r|j d|zz}d|_|j |jn|}|j||||}|xjdz c_|s|fSt| S) a Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with the Runge-Kutta method. It performs four forward passes to approximate the solution to the differential equation. Args: model_output (`torch.Tensor`): The direct output from learned diffusion model. timestep (`int`): The current discrete timestep in the diffusion chain. sample (`torch.Tensor`): A current instance of a sample created by the diffusion process. return_dict (`bool`): Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. Returns: [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a tuple is returned where the first element is the sample tensor. aNumber of inference steps is 'None', you need to run 'set_timesteps' after creating the schedulerrrr>gUUUUUU?r gUUUUUU? prev_sample) rMrrJr]r.rRrIrLrrK_get_prev_sampler ) rUrirjrkrl diff_to_prev prev_timesteprKrws rrpzPNDMScheduler.step_prks8  # # +s !LL1,q$++2Q2QUYUmUm2mqr2r  </ %%dlla&7!&;< < !;77rc|j td|jjs0t |j dkrt|j d||jj|jzz }|jdk7r0|j dd|_|j j|n(|}||jj|jzz}t |j dk(r|jdk(r |}||_ n0t |j dk(r8|jdk(r)||j dzd z }|j}d|_ nt |j d k(r&d|j dz|j d z d z }nt |j dk(r>"#(( !4;;#B#BdF^F^#^^ <<1 xx}DH HHOOL )$M$++"A"ATE]E]"]]H txx=A $,,!"3'L$DO ]a DLLA$5(488B<71 !;77rc|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. Returns: `torch.Tensor`: A scaled input sample. )rUrkargskwargss rscale_model_inputzPNDMScheduler.scale_model_inputs  rc|j|}|dk\r|j|n |j}d|z }d|z }|jjdk(r|dz|z|dz|zz}n<|jjdk7r#t d|jjd||z dz} ||dzz||z|zdzz} | |z||z |z| z z } | S)Nrr v_predictionr;epsilonzprediction_type given as z+ must be one of `epsilon` or `v_prediction`)rErFr]r5r) rUrkrjrzri alpha_prod_talpha_prod_t_prev beta_prod_tbeta_prod_t_prev sample_coeffmodel_output_denom_coeffrws rrxzPNDMScheduler._get_prev_samples!**84 BOSTBTD// >Z^ZrZr,& 00 ;; & &. 8(#-=cAQU[@[ [[ ( (I 5+DKK,G,G+HHst *L8cB $02Bs2K#K ; &): :O$ 6 !%6%E$UXp$p p roriginal_samplesnoiserTc|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}|S)N)rWrr;r?r )rErfrWrflattenroshape unsqueeze)rUrrrTrEsqrt_alpha_prodsqrt_one_minus_alpha_prod noisy_sampless r add_noisezPNDMScheduler.add_noisesa#1144>qAFF>L E$("%BF$!&( )2 22 2  2  bjj$u+&= >? 2222222h:":"eCDU>V:"B! xllxx x  x % & xL! :8ll:8:8 :8  :8 % & :8B! E8llE8E8 E8  E8 % & E8N  %,, )X,,||??   4/[ ?sE r-)g+?r)rtypingrrrrnumpyrNr!configuration_utilsrr scheduling_utilsr r r r+r-rrrrs=" // AXX !)4XT/NKT/r