bi)a x d dl Z d dlmZmZmZmZ d dlZd dlZddl m
Z
m Z ddl m
Z
ddlmZmZ G d d ee
Zy)
N)ListOptionalTupleUnion ) ConfigMixinregister_to_config )BrownianTreeNoiseSampler)SchedulerMixinSchedulerOutputc e Zd ZdZg ZdZe d3dedededede de d ed
ed ed e
d
e
de e fd Z e
d Ze
d Ze
d Zd4de fdZd Zd Zd Zdej, deeej, f dej, fdZd5d e d eeej2 f fd Zd5dej, fd Zd5dej, fd!Z d" Z d# Z d6d$ej, dej, dej, fd%Z d5d$ej, dej, d&e ej, dej, fd'Z d5d(e!ej, dej, d&e ej, dej, fd)Z"d6d*Z#d+ Z$ d7d$ej, dee ej, f dej, d,e
dee%e&f f
d-Z'd.ej, d&ej, d/ej, dej, fd0Z(d1 Z)d2 Z*y )8!CosineDPMSolverMultistepSchedulerag
Implements a variant of `DPMSolverMultistepScheduler` with cosine schedule, proposed by Nichol and Dhariwal (2021).
This scheduler was used in Stable Audio Open [1].
[1] Evans, Parker, et al. "Stable Audio Open" https://huggingface.co/papers/2407.14358
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:
sigma_min (`float`, *optional*, defaults to 0.3):
Minimum noise magnitude in the sigma schedule. This was set to 0.3 in Stable Audio Open [1].
sigma_max (`float`, *optional*, defaults to 500):
Maximum noise magnitude in the sigma schedule. This was set to 500 in Stable Audio Open [1].
sigma_data (`float`, *optional*, defaults to 1.0):
The standard deviation of the data distribution. This is set to 1.0 in Stable Audio Open [1].
sigma_schedule (`str`, *optional*, defaults to `exponential`):
Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper
(https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential
schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl.
num_train_timesteps (`int`, defaults to 1000):
The number of diffusion steps to train the model.
solver_order (`int`, defaults to 2):
The DPMSolver order which can be `1` or `2`. It is recommended to use `solver_order=2`.
prediction_type (`str`, defaults to `v_prediction`, *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).
solver_type (`str`, defaults to `midpoint`):
Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
lower_order_final (`bool`, defaults to `True`):
Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
euler_at_final (`bool`, defaults to `False`):
Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail
richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
steps, but sometimes may result in blurring.
final_sigmas_type (`str`, defaults to `"zero"`):
The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final
sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
r
sigma_min sigma_max
sigma_datasigma_schedulenum_train_timesteps solver_orderprediction_typerho solver_typelower_order_finaleuler_at_finalfinal_sigmas_typec
& | dvr1| dv r| j d nt | d| j t j dd| }
|dk( r| j |
}n|d k( r| j
|
}| j | _ t j |t j d|j
g | _ d | _
d g|z | _ d| _ d | _ d | _ | j j% d | _ y ) N)midpointheun)logrhobh1bh2r )r z is not implemented for r r
karras exponential)devicecpu)r NotImplementedError __class__torchlinspace_compute_karras_sigmas_compute_exponential_sigmasprecondition_noise timestepscatzerosr$ sigmasnum_inference_steps
model_outputslower_order_nums _step_index _begin_indexto)selfr r r r r r r r r r r r rampr0 s u/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py__init__z*CosineDPMSolverMultistepScheduler.__init__K s 2 2 66''J'?)[M9QRVR`R`Qa*bcc~~a$78 X %006F
}
,55d;F 008iiQv}})M NO $( "Vl 2 ! kknnU+ c @ | j j dz dz dz S )Nr r
?)configr r7 s r9 init_noise_sigmaz2CosineDPMSolverMultistepScheduler.init_noise_sigmas s# %%q(1,44r; c | j S )zg
The index counter for current timestep. It will increase 1 after each scheduler step.
)r4 r? s r9
step_indexz,CosineDPMSolverMultistepScheduler.step_indexx s
r; c | j S )zq
The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
r5 r? s r9 begin_indexz-CosineDPMSolverMultistepScheduler.begin_index s
r; rE c || _ y)z
Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
Args:
begin_index (`int`):
The begin index for the scheduler.
NrD )r7 rE s r9 set_begin_indexz1CosineDPMSolverMultistepScheduler.set_begin_index s (r; c 2 | j | }||z }|S N)_get_conditioning_c_in)r7 samplesigmac_in
scaled_samples r9 precondition_inputsz5CosineDPMSolverMultistepScheduler.precondition_inputs s! **51
r; c t |t j st j |g }|j t
j z dz S )Nr )
isinstancer( Tensortensoratanmathpi)r7 rL s r9 r, z4CosineDPMSolverMultistepScheduler.precondition_noise s: % .LL%)Ezz|dgg%))r; c j | j j }|dz |dz |dz z z }| j j dk( r||z |dz |dz z dz z }nR| j j dk( r| |z |dz |dz z dz z }n#t d| j j d ||z ||z z }|S )Nr epsilonr= v_predictionzPrediction type z is not supported.)r> r r
ValueError)r7 rK model_outputrL r c_skipc_outdenoiseds r9 precondition_outputsz6CosineDPMSolverMultistepScheduler.precondition_outputs s [[++
Q %(Z]":; ;; & &) 3J&%(Z]*Bs)JJE
[[
(
(N
:FZ'5!8j!m+C*KKE / 0K0K/LL^ _` `F?U\%99r; rK timestepreturnc | j | j | | j | j }| j || }d| _ |S )a
Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm.
Args:
sample (`torch.Tensor`):
The input sample.
timestep (`int`, *optional*):
The current timestep in the diffusion chain.
Returns:
`torch.Tensor`:
A scaled input sample.
T)rB _init_step_indexr0 rO is_scale_input_called)r7 rK r` rL s r9 scale_model_inputz3CosineDPMSolverMultistepScheduler.scale_model_input sM ?? " ! !( + DOO,))&%8%)"
r; Nr1 r$ c || _ t j dd| j }| j j dk( r| j | }n*| j j dk( r| j
| }j t j | }| j | | _
| j j dk( r| j j }n>| j j dk( rd}n"t d| j j t j |t j |gt j | g | _ d g| j j" z | _ d| _ d | _ d | _ | j j 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 r
r" r# )dtyper$ r zerozC`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got Nr% )r1 r( r) r> r r* r+ r6 float32r, r- r r rZ r. rS r0 r r2 r3 r4 r5
noise_sampler)r7 r1 r$ r8 r0
sigma_lasts r9
set_timestepsz/CosineDPMSolverMultistepScheduler.set_timesteps ss $7 ~~aD$<$<= ;; % % 1006F
[[
'
'=
855d;F v> 008 ;; ( (K 7 ..J
[[
*
*f
4J UVZVaVaVsVsUtu
ii zl%--`f)g hi
KK $ $ % !" kknnU+ "r; c |xs | j j }|xs | j j }| j j }|d|z z }|d|z z }||||z
z z |z }|S )z6Constructs the noise schedule of Karras et al. (2022).r
)r> r r r )r7 r8 r r r min_inv_rho max_inv_rhor0 s r9 r* z8CosineDPMSolverMultistepScheduler._compute_karras_sigmas sq 6!6!6 6!6!6 kkoo AG, AG, k(A BBsJ
r; c 4 |xs | j j }|xs | j j }t j t j | t j | t | j j d }|S )zImplementation closely follows k-diffusion.
https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusion/sampling.py#L26
r )
r> r r r( r) rU loglenexpflip)r7 r8 r r r0 s r9 r+ z=CosineDPMSolverMultistepScheduler._compute_exponential_sigmas sn
6!6!6 6!6!6 3TXXi5H#d)TXXZ__`ab
r; c t j t j |d }||d d t j f z
}t j |dk\ d j d j
|j d dz
}|dz }|| }|| }||z
||z
z } t j | dd } d| z
|z | |z z }
|
j |j }
|
S )Ng|=r )axisr )maxr
) nprq maximumnewaxiscumsumargmaxclipshapereshape) r7 rL
log_sigmas log_sigmadistslow_idxhigh_idxlowhighwts r9 _sigma_to_tz-CosineDPMSolverMultistepScheduler._sigma_to_t s FF2::eU34 Jq"**} 55 ))UaZq188a8@EE*JZJZ[\J]`aJaEbQ;!(# 9_t , GGAq! Ug H ,
IIekk "r; c 8 t j d }|}||fS )Nr
)r( rS )r7 rL alpha_tsigma_ts r9 _sigma_to_alpha_sigma_tz9CosineDPMSolverMultistepScheduler._sigma_to_alpha_sigma_t# s ,,q/ r; r[ c ^ | j | j }| j ||| }|S )a0
Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
integral of the data prediction model.
The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise
prediction and data prediction models.
Args:
model_output (`torch.Tensor`):
The direct output from the learned diffusion model.
sample (`torch.Tensor`):
A current instance of a sample created by the diffusion process.
Returns:
`torch.Tensor`:
The converted model output.
)r0 rB r_ )r7 r[ rK rL x0_preds r9 convert_model_outputz6CosineDPMSolverMultistepScheduler.convert_model_output) s/ 6 DOO,++FL%Hr; noisec j | j | j dz | j | j }}| j | \ }}| j | \ }}t j | t j | z
}t j | t j | z
} || z
}
|J ||z t j
|
z |z |dt j
d|
z z
z |z z |t j dt j
d|
z z
z |z z } | S )a
One step for the first-order DPMSolver (equivalent to DDIM).
Args:
model_output (`torch.Tensor`):
The direct output from the learned diffusion model.
sample (`torch.Tensor`):
A current instance of a sample created by the diffusion process.
Returns:
`torch.Tensor`:
The sample tensor at the previous timestep.
r
?)r0 rB r r( rq rs sqrt) r7 r[ rK r r sigma_sr alpha_slambda_tlambda_shx_ts r9 dpm_solver_first_order_updatez?CosineDPMSolverMultistepScheduler.dpm_solver_first_order_updateI s & ;;t':;T[[ =Y 77@ 77@99W% '(::99W% '(:: x
w
A2
.& 8!eiiq 112lB
C
326):#:;;eC
D
r; model_output_listc V | j | j dz | j | j | j | j dz
}}}| j | \ }}| j | \ }}| j | \ } }t j | t j | z
}
t j | t j | z
} t j | t j | z
} |d |d }}
|
| z
| | z
}}||z }|
d|z |
|z
z }}|J | j
j dk( r||z t j | z |z |dt j d|z z
z |z z d|dt j d|z z
z z |z z |t j dt j d|z z
z |z z }|S | j
j dk( r||z t j | z |z |dt j d|z z
z |z z |dt j d|z z
d|z z dz z |z z |t j dt j d|z z
z |z z }S ) a
One step for the second-order multistep DPMSolver.
Args:
model_output_list (`List[torch.Tensor]`):
The direct outputs from learned diffusion model at current and latter timesteps.
sample (`torch.Tensor`):
A current instance of a sample created by the diffusion process.
Returns:
`torch.Tensor`:
The sample tensor at the previous timestep.
r
r r r r r= r ) r0 rB r r( rq r> r rs r )r7 r rK r r sigma_s0sigma_s1r alpha_s0alpha_s1r lambda_s0 lambda_s1m0m1r h_0r0D0D1r s r9 (multistep_dpm_solver_second_order_updatezJCosineDPMSolverMultistepScheduler.multistep_dpm_solver_second_order_updatel s (
KK!+ , KK ( KK!+ , $ 77@ ! 9 9( C( ! 9 9( C(99W% '(::IIh'%))H*== IIh'%))H*== "2&(9"(=BI%y9'<3
1WcBh27+B ;; " "j 08#eiim3v=a%))D1H"5 56"<=A $((;$;<=BC EJJsUYYrAv->'> ??%GH
[[
$
$
.8#eiim3v=a%))D1H"5 56"<=sUYYtax%8 8TAX F LMQSST EJJsUYYrAv->'> ??%GH
r; c | | j }||k( j }t | dk( rt | j dz
}|S t | dkD r|d j }|S |d j }|S )Nr r
)r- nonzerorr item)r7 r` schedule_timestepsindex_candidatesrB s r9 index_for_timestepz4CosineDPMSolverMultistepScheduler.index_for_timestep s %!% .( :CCE A % T^^,q0J !
"Q
&)!,113J *!,113Jr; c | j Vt |t j r%|j | j
j }| j | | _ y| j | _ y)zF
Initialize the step_index counter for the scheduler.
N)
rE rQ r( rR r6 r- r$ r r4 r5 )r7 r` s r9 rc z2CosineDPMSolverMultistepScheduler._init_step_index sW
#(ELL1#;;t~~'<'<= # 6 6x @D # 0 0D r; return_dictc | j t d | j | j | | j t | j
dz
k( xrc | j j xsK | j j xr t | j
dk xs | j j dk( }| j t | j
dz
k( xr0 | j j xr t | j
dk }| j || }t | j j dz
D ]! }| j |dz | j |< # || j d< | j ~d} |>t |t r |D
cg c] }
|
j# c}
n|j# } t% || j j& | j j( | | _ | j | j* | j | j* | j dz j- |j. } | j j dk( s| j0 dk s|r| j3 |||
} nH| j j dk( s| j0 dk s|r | j5 | j ||
} | j0 | j j k r| xj0 dz
c_ | xj6 dz
c_ |s fS t9 S c c}
w ) a
Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
the multistep DPMSolver.
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.
generator (`torch.Generator`, *optional*):
A random number generator.
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.
NzaNumber of inference steps is 'None', you need to run 'set_timesteps' after creating the schedulerr
rh r )rK r )r r seed)rK r ) prev_sample) r1 rZ rB rc rr r- r> r r r r ranger r2 rj rQ list initial_seedr r r r0 r6 r$ r3 r r r4 r
)
r7 r[ r` rK generatorr r lower_order_secondir gr r s
r9 stepz&CosineDPMSolverMultistepScheduler.step s < # # + s
?? " ! !( + "__DNN0Ca0G G
KK & &
7 --J#dnn2E2J
7{{,,6 __DNN 3a 7
7 wT[[=Z=Z w_bcgcqcq_ruw_w 00 f0M t{{//!34 >A$($6$6q1u$=D q ! >!-2 %D$=G SW=Xy9!Q^^%9^g^t^t^v ": (=(=I^I^ei"D ""4;;t#?T___`M`Aabee
;; # #q (D,A,AA,EIZ <<\RX`e<fK
[[
%
%
*d.C.Ca.GK] GGHZHZciqvGwK 4;;#;#; ; ! !Q & !
A > ! ;77/ :s "M original_samplesr- c | j j |j |j }|j j dk( rvt j | ra| j j |j t
j }|j |j t
j }n@| j j |j }|j |j }| j |D cg c] }| j || }}nG| j | j g|j d z }n | j g|j d z }|| j }t |j t |j k r=|j d }t |j t |j k r=|||z z } | S c c}w )N)r$ rg mps)rg r r )r0 r6 r$ rg typer( is_floating_pointr- ri rE r rB r~ flattenrr unsqueeze)
r7 r r r- r0 r r step_indicesrL
noisy_sampless
r9 add_noisez+CosineDPMSolverMultistepScheduler.add_noise s '7'>'>FVF\F\ ] " " ' '5 0U5L5LY5W!% !2!23C3J3JRWR_R_!2!` ! %5%<% r )r7 rL rM s r9 rJ z8CosineDPMSolverMultistepScheduler._get_conditioning_c_in< s. UAX 6 6 99cAB r; c . | j j S rI )r> r r? s r9 __len__z)CosineDPMSolverMultistepScheduler.__len__@ s {{...r; ) g333333?i r r# i r rY g @r TFrh )r )NNrI )NT)+__name__
__module__ __qualname____doc__ _compatiblesorderr floatstrintboolr r: propertyr@ rB rE rG rO r, r_ r( rR r re r$ rl r* r+ r r r r r r r rc r
r r r rJ r r; r9 r r s! )V L
E +#' -%"& $+1%,%, %, %,
%, !
%, %, %, %, %, %, %, $C=%, %,N 5 5 ! !(3 (
*
eU\\FY@Z _d_k_k 0," ,"U3PUP\P\K\E] ,"^ ell SXS_S_ . $ ll
F $(, !ll! ! % !
!L $(, 8 -8 8 % 8
8v(
1"
P8llP8 U\\)*P8 P8
P8
% &P8f ,, || <<
D/r; r )rU typingr r r r numpyrx r( configuration_utilsr r scheduling_dpmsolver_sder scheduling_utilsr r
r r r; r9 r s/ " / / A > =e/ e/r;