biddlZddlmZddlmZmZmZmZddlZ ddl Z ddl m Z m Z ddlmZddlmZmZddlmZej*eZeGd d eZ dd Zd e j4d e j4fdZGddee Zy)N) dataclass)ListOptionalTupleUnion) ConfigMixinregister_to_config)SchedulerMixin) BaseOutputlogging) randn_tensorcXeZdZUdZej ed<dZeej ed<y)TCDSchedulerOutputa 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_noised_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): The predicted noised sample `(x_{s})` based on the model output from the current timestep. prev_sampleNpred_noised_sample) __name__ __module__ __qualname____doc__torchTensor__annotations__rr^/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/schedulers/scheduling_tcd.pyrr"s' 15.5rrc $|dk(rd}n|dk(rd}ntd|g}t|D]<}||z }|dz|z }|jtd||||z z |>t j |tj S)a Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [0,1]. Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up to that part of the diffusion process. Args: num_diffusion_timesteps (`int`): the number of betas to produce. max_beta (`float`): the maximum beta to use; use values lower than 1 to prevent singularities. alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. Choose from `cosine` or `exp` Returns: betas (`np.ndarray`): the betas used by the scheduler to step the model outputs cosinecftj|dzdz tjzdz dzS)NgMb?gT㥛 ?r)mathcospits r alpha_bar_fnz)betas_for_alpha_bar..alpha_bar_fnMs-88QY%/$''9A=>!C Crexpc2tj|dzS)Ng()r r&r#s rr%z)betas_for_alpha_bar..alpha_bar_fnRs88AI& &rz"Unsupported alpha_transform_type: dtype) ValueErrorrangeappendminrtensorfloat32)num_diffusion_timestepsmax_betaalpha_transform_typer%betasit1t2s rbetas_for_alpha_barr84s.x' D  & '=>R=STUU E * +M ( (!e. . S\"- R0@@@(KLM <<U]] 33rr4returnc(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_snrrIas5[F]]6q1N$))+O(*002'+113((O(,=@Q,QRRO!!#J ^j"o -F YY 1Q0 1F JE Lrc$(eZdZdZdZe d2dedededede e e je efd ed ed ed ed edededededededef"dZd3dZdZedZedZd4defdZd3dej.de edej.fdZd Zdej.dej.fd!Z d5d"e ed#e eej6fd e ed$e e ed%ef d&Z d6d'ej.dedej.d(ed)e ej:d*ede eeffd+Z d,ej.d-ej.d$ejBdej.fd.Z"dej.d-ej.d$ejBdej.fd/Z#d0Z$d1Z%y)7 TCDSchedulera `TCDScheduler` incorporates the `Strategic Stochastic Sampling` introduced by the paper `Trajectory Consistency Distillation`, extending the original Multistep Consistency Sampling to enable unrestricted trajectory traversal. This code is based on the official repo of TCD(https://github.com/jabir-zheng/TCD). This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~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`, 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`. original_inference_steps (`int`, *optional*, defaults to 50): The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule. clip_sample (`bool`, defaults to `True`): Clip the predicted sample for numerical stability. clip_sample_range (`float`, defaults to 1.0): The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. set_alpha_to_one (`bool`, defaults to `True`): 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. steps_offset (`int`, defaults to 0): An offset added to the inference steps, as required by some model families. prediction_type (`str`, defaults to `epsilon`, *optional*): Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen Video](https://imagen.research.google/video/paper.pdf) paper). thresholding (`bool`, defaults to `False`): Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such as Stable Diffusion. dynamic_thresholding_ratio (`float`, defaults to 0.995): The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. sample_max_value (`float`, defaults to 1.0): The threshold value for dynamic thresholding. Valid only when `thresholding=True`. 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. timestep_scaling (`float`, defaults to 10.0): The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation error at the default of `10.0` is already pretty small). 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(Nnum_train_timesteps beta_startbeta_end beta_schedule trained_betasoriginal_inference_steps clip_sampleclip_sample_rangeset_alpha_to_one steps_offsetprediction_type thresholdingdynamic_thresholding_ratiosample_max_valuetimestep_spacingtimestep_scalingrescale_betas_zero_snrc|+tj|tj|_n|dk(r-tj|||tj|_nk|dk(r6tj|dz|dz|tjdz|_n0|dk(rt ||_nt |d|j|rt|j|_d|jz |_ tj|jd |_ | rtjdn|jd |_ d|_ d|_tjt!j"d |ddd j%j't j(|_d |_d|_d|_y) Nr)linear scaled_linear?rsquaredcos_cap_v2z is not implemented for r;rr<r>F)rr/r0r4linspacer8NotImplementedError __class__rIrCr?rDfinal_alpha_cumprodinit_noise_sigmanum_inference_steps from_numpynparangecopyastypeint64 timestepscustom_timesteps _step_index _begin_index)selfrLrMrNrOrPrQrRrSrTrUrVrWrXrYrZr[r\s r__init__zTCDScheduler.__init__s}*  $m5==IDJ h & H>QY^YfYfgDJ o - C3H[chcpcpquvvDJ 1 1,-@ADJ%7OPTP^P^O_&`a a "24::>DJDJJ& #mmDKKQ? 9I5<<#4dNaNabcNd !$$( ))"))A7J*KDbD*Q*V*V*X*_*_`b`h`h*ij % rc| |j}||k(j}t|dkDrdnd}||jS)Nr(r)rnnonzerolenitem)rrtimestepschedule_timestepsindicesposs rindex_for_timestepzTCDScheduler.index_for_timestepsL  %!% %1::< wsi**84 BOSTBTD// >Z^ZrZr,& 00${2q 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*shaperr0float64floatreshaperiprodabsquantileconfigrXclamprY unsqueezer)rrrr* batch_sizechannelsremaining_dims abs_sampless r_threshold_samplezTCDScheduler._threshold_sampleIs 06 - H~  6 6\\^F Hrww~7N,NOZZ\ NN:t{{'M'MST U KK 1$++66  KKNVaR+a/ HF~F5! rrgrrnstrengthc: | | td| | td||n|jj}|||jjkDr=td|d|jjd|jjd|jj|z}t j t tdt||zdz|zdz }nHt j t td t|jj|z}|6t|} g} tdt|D]7} || || dz k\r td || | vs$| j|| 9|d |jjk\r#td |jjd |d k(rU|d |jjdz k7r6tjd|d d|jjdz d| rtjd| d|4t||kDrtjdt|d|dn[t||jjkDr9tjdt|d|jjdt j|tj}t||_d|_t%t|j |z|j } t'|j | z d } || |j(zd}nl||jjkDr=td|d|jjd|jjd|Ft||z}|dkr3td|d|d|dt||zdt+||z d ||_|||kDrQtd|d|d||jjkDr&td|d|jjd |ddd!j-}t j.d t||d"#}t j0|j3tj}||}t5j6|j9|t4j:$|_d|_d|_ y)%a Sets the discrete timesteps used for the diffusion chain (to be run before inference). Args: num_inference_steps (`int`, *optional*): 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. original_inference_steps (`int`, *optional*): The original number of inference steps, which will be used to generate a linearly-spaced timestep schedule (which is different from the standard `diffusers` implementation). We will then take `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute. 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 on the training/distillation timestep schedule is used. If `timesteps` is passed, `num_inference_steps` must be `None`. strength (`float`, *optional*, defaults to 1.0): Used to determine the number of timesteps used for inference when using img2img, inpaint, etc. NzEMust pass exactly one of `num_inference_steps` or `custom_timesteps`.zACan only pass one of `num_inference_steps` or `custom_timesteps`.z`original_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.r(rz/`custom_timesteps` must be in descending order.z=`timesteps` must start before `self.config.train_timesteps`: .r;z6The first timestep on the custom timestep schedule is z-, not `self.config.num_train_timesteps - 1`: zC. You may get unexpected results when using this timestep schedule.zThe custom timestep schedule contains the following timesteps which are not on the original training/distillation timestep schedule: z;The number of timesteps in the custom timestep schedule is zK, which exceeds the the length of the timestep schedule used for training: zH. You may get some unexpected results when using this timestep schedule.r)Tz`num_inference_steps`: z0The combination of `original_steps x strength`: z x z( is smaller than `num_inference_steps`: zK. Make sure to either reduce `num_inference_steps` to a value smaller than z/ or increase `strength` to a value higher than z3 cannot be larger than `original_inference_steps`: zx because the final timestep schedule will be a subset of the `original_inference_steps`-sized initial timestep schedule.z. cannot be larger than `num_train_timesteps`: zs because the final timestep schedule will be a subset of the `num_train_timesteps`-sized initial timestep schedule.r>F)numendpoint)rr*)!r+rrQrLriasarraylistr,intsetrvr-loggerwarningarrayrmrgror.rorderrrkrbfloorrlrrhrlongrnrprq)rrrgrrQrnroriginal_stepsktcd_origin_timestepstrain_timestepsnon_train_timestepsr5 init_timestept_start skipping_stepinference_indicess r set_timestepszTCDScheduler.set_timestepsjs<  &9+<de e  *y/D`a a)A(L $RVR]R]RvRv  $ + ? ?? ((89 7789 $ ? ?@ M //>AA#%::d5CQY@Y>> STXT_T_TsTsSttuv 39Q<4;;3R3RUV3V#VLYWX\N[??C{{?^?^ab?b>cdMN#AAT@UV:;)y>N2NNUVYZcVdUefSSaRbcQR y>DKK$C$CCNNUVYZcVdUefSSWS^S^SrSrRstQR "((;I'*9~D $$(D ! D$<$JZZ]^f]ghPQdPeepqtuCFNuNqOpPP@EFY\jFj@kllmn(;D $)&7$12E1FG*+,WX ')H)HH$12E1FG KK;;<=RS$8"#=#B#B#D " As3G/HNalq r  "): ; B B288 L ,->?I)))477vUZZ7X r model_outputeta generator return_dictc|j td|j|j|d|cxkr dksJdJd|jdz}|t |j kr|j |}nt jd}t jd|z |zjt j} |j|} d| z } |dk\r|j|n |j} |j| } d| z }|jjdk(rQ|| j|zz | jz }|}| j|z|j|zz}n|jjdk(r;|}|| d z|zz | d zz }| j|z|j|zz}n|jjd k(rJ| d z|z| d z|zz }| d z|z| d z|zz}| j|z|j|zz}n#td |jjd |dkDr}|j|jdz k7r^t!|j"||j$|j& }| | z j|zd| | z z j|zz}n|}n|}|xj(dz c_|s||fSt+||S)aq 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). 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. eta (`float`): A stochastic parameter (referred to as `gamma` in the paper) used to control the stochasticity in every step. When eta = 0, it represents deterministic sampling, whereas eta = 1 indicates full stochastic sampling. generator (`torch.Generator`, *optional*): A random number generator. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] or `tuple`. Returns: [`~schedulers.scheduling_utils.TCDSchedulerOutput`] or `tuple`: If return_dict is `True`, [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] is returned, otherwise a tuple is returned where the first element is the sample tensor. zaNumber of inference steps is 'None', you need to run 'set_timesteps' after creating the schedulerrr;z'gamma must be less than or equal to 1.0r(r)epsilonrr` v_predictionzprediction_type given as zI must be one of `epsilon`, `sample` or `v_prediction` for `TCDScheduler`.)rrr*)rr)rgr+rrrvrnrr/rrrrDrerrVr@rrrr*rpr)rrrrxrrrrprev_step_indexr timestep_srrr alpha_prod_s beta_prod_spred_original_sample pred_epsilonrnoisers rstepzTCDScheduler.step sIB  # # +s  ?? "  ! !( +C3I III II//A- S0 0 NN?;M!LLOM[[!c']!:;>>UZZ>P **84 ,& BOSTBTD// >Z^ZrZr**:6 ,&  ;; & &) 3$*[-=-=-?,-N$NR^RcRcRe#e 'L!-!2!2!47K!KkN^N^N`coNo!o  [[ ( (H 4#/ "\c%:=Q%QQU`ehUiiL!-!2!2!47K!KkN^N^N`coNo!o  [[ ( (N :$0#$5#?;PSCSWcBc#c (#-=cAQU[@[!-!2!2!47K!KkN^N^N`coNo!o +DKK,G,G+HI66  7$":":Q">>$ &&)LDWDW_q_w_w 1<?EEGJ\\)L88$&5`!! 1 ,K A!34 4!kN`aaroriginal_samplesrc|jj|j|_|jj|j}|j|j}||dz}|j }t |j t |j kr=|jd}t |j t |j kr=d||z dz}|j }t |j t |j kr=|jd}t |j t |j kr=||z||zz}|SN)rr)r`r>r(rDrrr*flattenrvrr)rrrrrnrDsqrt_alpha_prodsqrt_one_minus_alpha_prod noisy_sampless r add_noisezTCDScheduler.add_noisewsa#1144)rorgrnrurrr/)rrrxindexprev_ts rprevious_timestepzTCDScheduler.previous_timesteps  D$<$<^^x/88$8GJ1ME,,Q/!33b)   2 \F r)ig_QK?g~jt?r_N2Fr;TrrFgףp= ?r;leadingg$@Fr~)r)NNNNr;)g333333?NT)&rrrrrr rrstrrrrindarrayrboolrsr|rpropertyrrrrrrrrrr Generatorrrr IntTensorrrrrrrrrKrKs:x E$(#,BF(*!#&!%(",1"% )"&',%7! 7!7! 7!  7!  bjj$u+&= >? 7!#&7!7!!7!7!7!7!7!%*7! 7! !7!" #7!$!%%7!7!t #1  !!(3(  Y^YeYe$ F.2+/26)- `!%c]`!c5<<'(`!#+3- `! DI& `!  `!N/3 hbllhbhb hb  hb EOO, hbhb !5( )hbV,,||??   65<< QVQ`Q`ejeqeq&/ rrK)g+?r)r dataclassesrtypingrrrrnumpyrirconfiguration_utilsr r schedulers.scheduling_utilsr utilsr r utils.torch_utilsr get_loggerrrrr8rrIrKrrrrs$ !// A8',   H %  6 6  6&!)4Z!U\\!ell!Hm>;mr