o ߥi1$@s<ddlZddgZddZ   d ddZGdddeZdS) NGaussianDiffusion beta_schedulecCs:||j}|dfd|jd}||||S)z?Index tensor using t and format the output according to x. rr)todevicesizendimview)tensortxshaperk/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/multi_modal/video_synthesis/diffusion.py_i s rcCs6|dkrtj|d|d|tjddStd|)N linear_sdg?dtypezUnsupported schedule: )torchlinspacefloat64 ValueError)schedule num_timesteps init_beta last_betarrrrsc@seZdZdZ     dddZid d d fd d Zd d Zeid d d d ddfddZ eid d d d ddfddZ ddZ d S)rz Diffusion Model for DDIM. "Denoising diffusion implicit models." by Song, Jiaming, Chenlin Meng, and Stefano Ermon. See https://arxiv.org/abs/2010.02502 eps learned_rangemse-q=FcCst|tjstj|tjd}t|dkrt|dksJ|dvs"J|dvs(J|dvs.J||_t||_ ||_ ||_ ||_ ||_ ||_d|j}tj|dd|_t|dg|jddg|_t|jdd|dgg|_t|j|_td |j|_td |j|_td |j|_td |jd|_|d |jd |j|_t|jd |_|t|jd |j|_ d |jt|d |j|_!dS) Nrrr)x0zx_{t-1}r)learnedr fixed_large fixed_small)r! rescaled_msekl rescaled_kll1 rescaled_l1 charbonnierdim?g#B ;)" isinstancer DoubleTensorr rminmaxbetaslenr mean_typevar_type loss_typeepsilonrescale_timestepscumprodalphas_cumprodcatnew_onesalphas_cumprod_prev new_zerosalphas_cumprod_nextsqrtsqrt_alphas_cumprodsqrt_one_minus_alphas_cumprodloglog_one_minus_alphas_cumprodsqrt_recip_alphas_cumprodsqrt_recipm1_alphas_cumprodposterior_varianceclampposterior_log_variance_clippedposterior_mean_coef1posterior_mean_coef2)selfr5r7r8r9r:r;alphasrrr__init__$sb            zGaussianDiffusion.__init__NcCs|dur||||fi|}npt|trt|dksJ||||fi|d} ||||fi|d} |jdrE| dn| dd} | ddd| f} || ddd| f| ddd| f} | dd| df}tj| | |gdd}|jdkrt |j ||}t |j ||}|j dkrt |j |||t |j|||}||||\}}}|dur|dkr|dksJtj|d|ddd d ddd}t|t| ||}n |dur|| |}||||fS) z*Distribution of p(x_{t-1} | x_t). Nrrrfixedr-r&rr0r/)_scale_timestepsr1listr6r8 startswithrrr>rrJrLr7rHrIq_posterior_mean_variancequantileflattenabsclamp_r r3r4rK)rOxtr model model_kwargsrK percentile guide_scaleouty_outu_outr.abcvarlog_varr#mu_srrrp_mean_variance]sH  ,   z!GaussianDiffusion.p_mean_variancecCsJt|j|||t|j|||}t|j||}t|j||}|||fS)z/Distribution of q(x_{t-1} | x_t, x_0). )rrMrNrJrL)rOr#r[r rhrfrgrrrrVs z+GaussianDiffusion.q_posterior_mean_variancegc  Cs|j| } ||||||||\} } } } |durWt|j||}t|j|||| t|j||}|d|||||fi|}t|j|||t|j|||} t|j|||| t|j||}t|j||}t|j|| d|}d|d|}d||}| t ||}t |}t d||d|}| d j dgd|jdR}t || ||||}|| fS)zSample from p(x_{t-1} | x_t) using DDIM. - condition_fn: for classifier-based guidance (guided-diffusion). - guide_scale: for classifier-free guidance (glide/dalle-2). Nrrrr/r)rrkrr=rHrIrCrSrKr randn_likenefloatr r )rOr[r r\r]rKr^ condition_fnr_ddim_timestepsetastriderir#alpharrP alphas_prevrcrdsigmasnoise directionmaskxt_1rrr ddim_samples@      &zGaussianDiffusion.ddim_samplec  Cs|d} |} dtd|j|j|d|jdd} | D]} tj| f| tj| jd}| | ||||||||| \} }q | S)Nrr)rr) rrarangerrKflipfulllongrr{)rOrwr\r]rKr^rpr_rqrrrdr[stepsstepr rirrrddim_sample_loops"    z"GaussianDiffusion.ddim_sample_loopcCs|jr |d|jS|S)Ng@@)r;ror)rOr rrrrSsz"GaussianDiffusion._scale_timesteps)rr r!r"F) __name__ __module__ __qualname____doc__rQrkrVrno_gradr{rrSrrrrrsB = 0  0 )rNN)r__all__rrobjectrrrrrs