2Vh|ddlZddlmZddlmZddlmZddlmZddlm Z ddlm Z ddlm Z dd l m Z d Zd Zd Zd ZdZdZGdde ZeddZy)N) keras_export) KerasTensor)any_symbolic_tensors)shape)prod)reshape) transpose) Operationc i}t||D]\}}tjd|}|r|jdj }|Dcgc] }||vs| } }|Dcgc] }||vs| } }| r)| d} t | Dcgc]}|| c}} || z|| <|j |Dcic]}||| c}|||<|Scc}wcc}wcc}wcc}w)N \(([\w\s]+)\)r)ziprematchgroupsplitrupdate) axes input_shape axes_lengthsaxes_mapaxisdim grouped_axes inner_axesa known_axes inferred_axes inferred_axis known_products D/home/dcms/DCMS/lib/python3.12/site-packages/keras/src/ops/einops.py_create_axes_mapr" sH{+! cxx 0$7 %++A.446J%/E1 3D!EJE(2L1a|6KQLML -a 0 $z%J!l1o%J K .1].B ]+ OODAQ Q/D E HTN!!$ OFL&KEs$ C C $ C.C C . C cg}|D]]}tjd|}|r1|jdj}|j |L|j |g_|S)Nr r )rrrrappend)rgrouped_output_axesrrrs r!_create_grouped_axesr&%sj/xx 0$7 %++A.446J  & &z 2  & &v ./ c@|Dcgc] }|D]}| c}}Scc}}wN)rxsxs r!_flatten_groupr-3s" )"b )A )A )) )scrtt|}|Dcgc]}|j|c}Scc}wr))r-r&index) from_shapeto_shapeflattened_from_shapers r!_get_transpose_orderr37s3)*>z*JK7? @ & &s + @@ @s4cng}|D]$}d}|D] }|||z} |j|&t|S)Nr )r$tuple)rr output_shapersizers r!_compute_output_shaper8=sSL" #D HTN "D #D! "  r'cg}g}|D]v}d|vrKtjd|}|Dcgc]}|| }}|j||j|R|j||j||x|Scc}w)N(z\w+)rfindallextendr$) input_axesrrreshaped_input_axesreshaped_sizesrrrsizess r!_compute_decomposed_shaperAHsN2 $;FD1J.89\!_9E9  & &z 2  ! !% (  & &t ,  ! !(4. 12 :s BceZdZdZdZy) Rearrangec t||fi|Sr)) rearrange)selftensorpatternrs r!callzRearrange.callZs9L99r'c tjd|\}}tjd|}tjd|}t|}t |||} t |} t | | } t| |jS)N\s*->\s* \w+|\(.*?\))rdtype) rrr;rr"r&r8rrM) rFrGrHr input_patternoutput_patternr= output_axesrrr%r6s r!compute_output_speczRearrange.compute_output_spec]sw(*g(F% ~ZZ > jj@ Fm #J \J2;?,X7JK V\\BBr'N)__name__ __module__ __qualname__rIrQr*r'r!rCrCYs : Cr'rCzkeras.ops.rearrangec t|frtj||fi|Stjd|\}}tj d|}tj d|}t |}t|||}t|} t| } t|||} | |j k7r t|| }t|| } t|| }t|| } t|| }|S)azRearranges the axes of a Keras tensor according to a specified pattern, einops-style. Args: tensor: Input Keras tensor. pattern: String describing the rearrangement in einops notation. **axes_lengths: Keyword arguments specifying lengths of axes when axes decomposition is used. Returns: Tensor: A Keras tensor with rearranged axes. Follows the logic of: 1. If decomposition is needed, reshape to match decomposed dimensions. 2. Permute known and inferred axes to match the form of the output. 3. Reshape to match the desired output shape. Example Usage: ``` >>> import numpy as np >>> from keras.ops import rearrange >>> images = np.random.rand(32, 30, 40, 3) # BHWC format # Reordering to BCHW >>> rearrange(images, 'b h w c -> b c h w').shape TensorShape([32, 3, 30, 40]) # "Merge" along first axis - concat images from a batch >>> rearrange(images, 'b h w c -> (b h) w c').shape TensorShape([960, 40, 3]) # "Merge" along second axis - concat images horizontally >>> rearrange(images, 'b h w c -> h (b w) c').shape TensorShape([30, 1280, 3]) # Flatten images into a CHW vector >>> rearrange(images, 'b h w c -> b (c h w)').shape TensorShape([32, 3600]) # Decompose H and W axes into 4 smaller patches >>> rearrange(images, 'b (h1 h) (w1 w) c -> (b h1 w1) h w c', h1=2, w1=2).shape TensorShape([128, 15, 20, 3]) # Space-to-depth decomposition of input axes >>> rearrange(images, 'b (h h1) (w w1) c -> b h w (c h1 w1)', h1=2, w1=2).shape TensorShape([32, 15, 20, 12]) ``` rKrL)rrC symbolic_callrrr;rr"r&r-rArr3r r8)rGrHrrNrOr=rPrrr%flattened_output_axesdecomposed_shapes permute_orderr6s r!rErEjslVI&(y{((ILII%'HH['$B!M>NM:J**^^?2L(FLL(!23)5JKM v} -F)3FGL V\ *F Mr')rkeras.src.api_exportrkeras.src.backendrrkeras.src.ops.corerkeras.src.ops.numpyrrr keras.src.ops.operationr r"r&r-r3r8rArCrEr*r'r!r_si -)2$$')-0 *A "C C"#$R%Rr'