<Иj,SrSSKJr SSKJrJr SSKrSSKJs J r SSK J r J r Jr SSKJrJrJr SSKJr SSKJr S rS r"S S \5rg) z Disentanglement via Latent Quantization - https://arxiv.org/abs/2305.18378 Code adapted from Jax version in https://github.com/kylehkhsu/latent_quantization ) annotations)CallableListN)pack rearrangeunpack)Tensorint32nn)Module) Optimizerc[U/U5$N)r)tpatterns ڹC:\Users\johnn\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.13_qbz5n2kfra8p0\LocalCache\local-packages\Python313\site-packages\vector_quantize_pytorch/latent_quantization.pypack_oners W c [XU5S$)Nr)r)rpsrs r unpack_oners ! !! $$rc^\rSrSrS SU4SjjjrSSSjjrSSSjjrSSjrSSjrSSjr SSjr SSS jjr SS jr SS jr S rU=r$)LatentQuantizec >[TU]5 X lXl[R "U[ S9n [U[5(aU RU5n URS[R "U[RS9SS9 URS[R "U[RS9SS9 URSU SS9 [R"[R"[R "S/[ S9U SS /S S 9S S 9n URS U SS9 US :aUO [U 5UlUR U-n XPlXlU(aUOUS:nUS:a U(deXplURU :gnU(a![(R*"URU 5O[(R,"5UlU(a [(R*"XR5O[(R,"5UlXlUR4R75R95Ul[=[R>"UR:5S 5nXR@-UR4-nURCU5nURSUSS9 U Vs/sH@nUS-S:Xa[RD"SSU5O[R>"U5U- S- PMB nnU(a\[(RF"UVs/sHn[(RH"U5PM sn5Ul%U bU "URJ5UlggUUl%g![an U eSn A ff=fs snfs snf)aB Initializes the LatentQuantization module. Args: levels (List[int]|init): The number of levels per codebook. If an int is provided, it is used for all codebooks. dim (int): The dimensionality of the input tensor. The input tensor is expected to be of shape [B D ...] num_codebooks (int): The number of codebooks to use. (default is 1) codebook_dim (int): the dimension of the codebook. If levels is a list, codebook_dim is the length of the list. (default to -1) keep_num_codebooks_dim (Optional[bool]): Whether to keep the number of codebooks dimension in the output tensor. If not provided, it is set to True if num_codebooks > 1, otherwise False. optimize_values (Optional[bool]): Whether to optimize the values of the codebook. If not provided, it is set to True. )dtypeNcommitment_loss_weightF) persistentquantization_loss_weight_levelsrdim_basis ... -> ... 1implicit_codebookgg?)&super__init__r$in_place_codebook_optimizertorchtensorr isinstanceintrepeat RuntimeErrorregister_bufferfloat32cumprodconcatlen codebook_dim num_codebookseffective_codebook_dimkeep_num_codebooks_dimr LinearIdentity project_in project_outhas_projectionsr proditem codebook_sizeraranger%_scale_and_shift_inverselinspace ParameterList Parametervalues_per_latent)selflevelsr$rrr8r7r:optimize_valuesr+r er%r9r? all_indicescodes_non_centeredr'levelrHvalues __class__s rr*LatentQuantize.__init__s3< +F(,,vU3 fc " " !..6  $ LL/u}} E   & LL1 G  YEB LL%,,s%8'#2,GQ OUV  Xv%@,81,.distances99QU# #r.Nr"r#)r,stackranger7argminrHrY)rIr[rhiindexquantizes rroLatentQuantize.quantizes  $ t001  2A QsAt|_d.D.DQ.GHb2    ;;t001 1A&&q)%Q-81   1 ,,..%  s 9C!Cc6URS-nUS-U-U-$)z;scale and shift zhat from [-0.5, 0.5] to [0, level_per_dim]r(r )rIzhat_normalized half_widths r_scale_and_shiftLatentQuantize._scale_and_shifts%\\Q& !#j0J>>rc4URS-nX- U- S- $)znormalize zhat to [-0.5, 0.5]r(rr)rIr\rts rrD'LatentQuantize._scale_and_shift_inverses#\\Q& !Z/!33rcURSUR:XdeURU5nXR-R SS9R [ 5$)zSConverts a `code` which contains the number per latent to an index in the codebook.r"r#)shaper7rur%sumtor )rIr\s rcodes_to_indicesLatentQuantize.codes_to_indicessTzz"~!2!2222$$T*{{"''B'/22599rc[US5nXR-UR-nURU5nUR(a [US5nU(aUR U5n[US5nU$)zInverse of `codes_to_indices`.r&z... c d -> ... (c d)b ... d -> b d ...)rr%r rDr:r>)rIindicesr>rNcodess rindices_to_codesLatentQuantize.indices_to_codessrG^4%4 D--.@A  & &e%; b n (c d)b * drb * c ... 1 -> ...)ror}rr>rr:)rIr[is_img_or_videorrrouts rquantize_and_project#LatentQuantize.quantize_and_projects a ''.%!78u%'*12W'2**8G7""rcLUnURSLn[US5n[US5upURSUR:Xd$SURSURS35eUR U5n[USUR S9nURU5nURU5n[US 5nURU5n[XtS5n[US 5n[XdS 5nUR(d [US 5nU(GaYUR(GaGUR(Gd5URS :waURX5O[ R""S5nUUR$S :waUR'X5O[ R""S5- nUR)5 URR+5 URR-5 URU5nURU5n[US 5nURU5n[XtS5n[US 5n[XdS 5nUR(d [US 5nUR(a!URS :waURX'5O[ R""S5n UR(a!UR$S :waUR'X'5O[ R""S5n URU -UR$U --nXvU4$)z} einstein notation b - batch n - sequence (or flattened spatial dimensions) d - feature dimension c - number of codebook dim Nzb d ... -> b ... drr"zexpected dimension of z but found dimension of zb n (c d) -> b n c d)crrrrrg)r+rrrzr$r=r8ror}r>rr:trainingrKrrar,r-rr^backwardstep zero_grad) rIr[original_inputshould_inplace_optimizerrrrlossrar^s rforwardLatentQuantize.forwards"&"B"B$"N a- .G$ GGBK488 # T #DHH:-Eaggbk] S T # OOA  a/43E3E F a ''.%!78u%'*12W'2**8G "t}}}T=Q=Q=Q..!3$$Q,\\#&  00A5&&q.\\#& D MMO  , , 1 1 3  , , 6 6 8MM!$E++E2Ge%;!>!!C  " "> 7c"   ' '/ 9++.?? @ T!!r) r7rBr$r9r?r+r:r8r=r>rH)皙?rr!r"NTN)rJzList[int] | intr$r/r float | Nonerrr8r/r7r/r: bool | NonerKrr+zCallable[..., Optimizer])mean)r[r r\r returnr )r[r rr )rsr rr )r\r rr )T)rr rr )__name__ __module__ __qualname____firstlineno__r*r^rarorurDr}rrr__static_attributes__ __classcell__)rQs@rrrs 0314.2'+ m7m7m7!- m7 #/ m7  m7m7!,m7%m7& m7m7^>><? 4 : "# S"S"rr)__doc__ __future__rtypingrrr,torch.nn.functionalr functionalrWeinopsrrrr r torch.nnr torch.optimr rrrrrrsE #! **##! %Z"VZ"r