o pi1*@sddlmZddlmZmZmZddlZddlmZ ddl m Z m Z ddl mZmZmZmZmZejjGdd d ZeGd d d eZGd d d ee ZdS)) dataclass)OptionalTupleUnionN) ConfigMixinregister_to_config)CommonSchedulerStateFlaxKarrasDiffusionSchedulersFlaxSchedulerMixinFlaxSchedulerOutputbroadcast_to_shape_from_leftc @sheZdZUeed<ejed<ejed<ejed<dZee ed<e dedejdejdejfddZ dS) EulerDiscreteSchedulerStatecommoninit_noise_sigma timestepssigmasNnum_inference_stepscCs|||||dS)Nrrrr)clsrrrrrrq/home/ubuntu/SoloSpeech/.venv/lib/python3.10/site-packages/diffusers/schedulers/scheduling_euler_discrete_flax.pycreate)sz"EulerDiscreteSchedulerState.create) __name__ __module__ __qualname__r __annotations__jnpndarrayrrint classmethodrrrrrrs    rc@seZdZUeed<dS) FlaxEulerDiscreteSchedulerOutputstateN)rrrrrrrrrr"0s r"c@s4eZdZUdZddeDZejed<e ddZ e ddd d d d d ej fde dedededeejdededejfddZd1deedefddZdedejde dejfddZ d2ded!e d"edefd#d$Z %d3ded&ejde dejd'edeeeff d(d)Zded*ejd+ejd,ejdejf d-d.Zd/d0Zd S)4FlaxEulerDiscreteSchedulera Euler scheduler (Algorithm 2) from Karras et al. (2022) https://arxiv.org/abs/2206.00364. . Based on the original k-diffusion implementation by Katherine Crowson: https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L51 [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions. Args: num_train_timesteps (`int`): number of diffusion steps used to train the model. beta_start (`float`): the starting `beta` value of inference. beta_end (`float`): the final `beta` value. beta_schedule (`str`): the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from `linear` or `scaled_linear`. trained_betas (`jnp.ndarray`, optional): option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. prediction_type (`str`, default `epsilon`, optional): prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 https://imagen.research.google/video/paper.pdf) dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`): the `dtype` used for params and computation. cCsg|]}|jqSr)name).0errr Rsz%FlaxEulerDiscreteScheduler.dtypecCsdS)NTrselfrrr has_stateVsz$FlaxEulerDiscreteScheduler.has_stateig-C6?g{Gz?linearNepsilonlinspacenum_train_timesteps beta_startbeta_end beta_schedule trained_betasprediction_typetimestep_spacingc Cs ||_dSNr)) r+r0r1r2r3r4r5r6r)rrr__init__Zs z#FlaxEulerDiscreteScheduler.__init__rreturncCs|dur t|}td|jjddd}d|j|jd}t|tdt ||}t |tj dg|j dg}|jj dvrH|}n |ddd}tj||||d S) Nrr ?r8r/trailingrr)r rrarangeconfigr0roundalphas_cumprodinterplen concatenatearrayr)r6maxr)r+rrrrrrr create_statehs   z'FlaxEulerDiscreteScheduler.create_stater#sampletimestepcCs@tj|j|kdd\}|d}|j|}||ddd}|S)a  Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. Args: state (`EulerDiscreteSchedulerState`): the `FlaxEulerDiscreteScheduler` state data class instance. sample (`jnp.ndarray`): current instance of sample being created by diffusion process. timestep (`int`): current discrete timestep in the diffusion chain. Returns: `jnp.ndarray`: scaled input sample r sizerrr<)rwhererr)r+r#rJrK step_indexsigmarrrscale_model_input~s  z,FlaxEulerDiscreteScheduler.scale_model_inputrrshapecCs|jjdkrtj|jjdd||jd}n.|jjdkr:|jj|}td||ddd t }|d7}n t d|jjd|j j |j j d }t|tdt||}t|tjd g|jdg}|jjd vru|}n |d dd }|j||||d S)a Sets the timesteps used for the diffusion chain. Supporting function to be run before inference. Args: state (`EulerDiscreteSchedulerState`): the `FlaxEulerDiscreteScheduler` state data class instance. num_inference_steps (`int`): the number of diffusion steps used when generating samples with a pre-trained model. r/r rr8leadingNr;z=timestep_spacing must be one of ['linspace', 'leading'], got r<r=r>r)rrrr)rAr6rr/r0r)r@rBcopyastypefloat ValueErrorrrCrDrErFrGrHreplace)r+r#rrRr step_ratiorrrrr set_timestepss*   (    z(FlaxEulerDiscreteScheduler.set_timestepsT model_output return_dictc Cs|jdur tdtj|j|kdd\}|d}|j|}|jjdkr*|||}n&|jjdkrF|| |ddd ||dd}n td |jjd |||} |j|d|} || | } |sk| |fSt| |d 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: state (`EulerDiscreteSchedulerState`): the `FlaxEulerDiscreteScheduler` state data class instance. model_output (`jnp.ndarray`): direct output from learned diffusion model. timestep (`int`): current discrete timestep in the diffusion chain. sample (`jnp.ndarray`): current instance of sample being created by diffusion process. order: coefficient for multi-step inference. return_dict (`bool`): option for returning tuple rather than FlaxEulerDiscreteScheduler class Returns: [`FlaxEulerDiscreteScheduler`] or `tuple`: [`FlaxEulerDiscreteScheduler`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. NzaNumber of inference steps is 'None', you need to run 'set_timesteps' after creating the schedulerr rLrr. v_predictionrr<zprediction_type given as z, must be one of `epsilon`, or `v_prediction`) prev_sampler#) rrWrrNrrrAr5r") r+r#r[rKrJr\rOrPpred_original_sample derivativedtr^rrrsteps(    ,   zFlaxEulerDiscreteScheduler.steporiginal_samplesnoisercCs*|j|}t||j}|||}|Sr7)rflattenrrR)r+r#rcrdrrP noisy_samplesrrr add_noises  z$FlaxEulerDiscreteScheduler.add_noisecCs|jjSr7)rAr0r*rrr__len__sz"FlaxEulerDiscreteScheduler.__len__r7)r)T)rrr__doc__r _compatiblesrr)rpropertyr,rfloat32r rVstrrrr9r rrIrQrrZboolrr"rbrgrhrrrrr$5s      /  = r$) dataclassesrtypingrrrflax jax.numpynumpyrconfiguration_utilsrrscheduling_utils_flaxr r r r rstructrr"r$rrrrs