o ٷi @sddlmZddlmZmZmZmZmZddlm Z ddl m Z ddl m Z edZeeedfeefZedd d ed ed eeeeffd dZdeed ed eeeeffddZded efddZdedeed ed eefddZdS)) lru_cache)ListSequenceTupleTypeVarUnion) EinopsError) get_backend)ParsedExpressionTensor.)maxsizepatternopnamereturnc Cs|}t|}t|t|krtd|d|dd|vr*td|d|d|D]}|dkrIt|\}}|sItd|d|d|dq,|d}t||d}||} ||| fS) NzDuplicates in axes names in z(..., "z")*z No *-axis in zInvalid axis name z in )splitsetlenrr check_axis_name_return_reasonindex) rraxesaxes_setaxisis_validreason n_axes_before n_axes_aftermin_axesr B/home/ubuntu/.local/lib/python3.10/site-packages/einops/packing.pyanalyze_pattern s   r"tensorsc Cst|d\}}}t|d}g}g}t|D]G\}} || } t| |kr6td|d| d|d|d t| |} || || ||| g| d|d | | dRq|j||d |fS) a Packs several tensors into one. See einops tutorial for introduction into packing (and how it replaces stack and concatenation). Parameters: tensors: tensors to be packed, can be of different dimensionality pattern: pattern that is shared for all inputs and output, e.g. "i j * k" or "batch seq *" Returns: (packed_tensor, packed_shapes aka PS) Example: ```python >>> from numpy import zeros as Z >>> inputs = [Z([2, 3, 5]), Z([2, 3, 7, 5]), Z([2, 3, 7, 9, 5])] >>> packed, ps = pack(inputs, 'i j * k') >>> packed.shape, ps ((2, 3, 71, 5), [(), (7,), (7, 9)]) ``` In this example, axes were matched to: i=2, j=3, k=5 based on order (first, second, and last). All other axes were 'packed' and concatenated. PS (packed shapes) contains information about axes that were matched to '*' in every input. Resulting tensor has as many elements as all inputs in total. Packing can be reversed with unpack, which additionally needs PS (packed shapes) to reconstruct order. ```python >>> inputs_unpacked = unpack(packed, ps, 'i j * k') >>> [x.shape for x in inputs_unpacked] [(2, 3, 5), (2, 3, 7, 5), (2, 3, 7, 9, 5)] ``` Read the tutorial for introduction and application scenarios. packrzpacked tensor #z' (enumeration starts with 0) has shape z, while pattern z assumes at least z axesN)r) r"r enumerateshaperrappendreshapeconcat) r#rrrrbackendreshaped_tensors packed_shapesitensorr'axis_after_packed_axesr r r!r$!s$$    2r$xcCsd}|D]}||9}q|S)Nrr )r1resultr.r r r!prodZs r3r/r-c st|dd\}}}t}t||d|kr&td|d||}dd|D}tdd |D} | dkrGtd|d |d d gt|||g| d krot|d dD]\} } | | | d<q_n5|d} t| D]} | || | d<qxt| dt|d d dD]} | d|| | <q|d |||dd t d d f|zfddt|DWSt y}ztd|ddd||d }~ww)a8 Unpacks a single tensor into several by splitting over a selected axes. See einops tutorial for introduction into packing (and how it replaces stack and concatenation). Parameters: tensor: tensor to be unpacked packed_shapes: packed_shapes (aka PS) is a list of shapes that take place of '*' in each output. output will contain a single tensor for every provided shape pattern: pattern that is shared for input and all outputs, e.g. "i j * k" or "batch seq *", where * designates an axis to be unpacked Returns: list of tensors If framework supports views, results are views to the original tensor. Example: ```python >>> from numpy import zeros as Z >>> inputs = [Z([2, 3, 5]), Z([2, 3, 7, 5]), Z([2, 3, 7, 9, 5])] >>> packed, ps = pack(inputs, 'i j * k') >>> packed.shape, ps ((2, 3, 71, 5), [(), (7,), (7, 9)]) ``` In this example, axes were matched to: i=2, j=3, k=5 based on order (first, second, and last). All other axes were 'packed' and concatenated. PS (packed shapes) contains information about axes that were matched to '*' in every input. Resulting tensor has as many elements as all inputs in total. Packing can be reversed with unpack, which additionally needs PS (packed shapes) to reconstruct order. ```python >>> inputs_unpacked = unpack(packed, ps, 'i j * k') >>> [x.shape for x in inputs_unpacked] [(2, 3, 5), (2, 3, 7, 5), (2, 3, 7, 9, 5)] ``` Read the tutorial for introduction and application scenarios. unpack)rrz unpack(..., z)) received input of wrong dim with shape cSs g|] }d|vr dnt|qS)r%)r3).0p_shaper r r! s zunpack..css|] }t|dkVqdS)r%N)int)r5r1r r r! szunpack..z) received more than one -1 in z and can't infer dimensionsrNr%c sJg|]!\}}gt||dRg|RqS)r)r)slice)r5r. element_shaper+ shape_end shape_start slice_fillersplit_positionsr/r r!r7s"zError during unpack(..., "z!"): could not split axis of size z into requested ) r"r r'rrsumr&rranger: Exception)r/r-rrrr input_shape unpacked_axislengths_of_composed_axesn_unknown_composed_axesr.r1unknown_composed_axisjer r<r!r4asN)      r4N) functoolsrtypingrrrrreinopsreinops._backendsr einops.parsingr r r8Shapestrr"r$r3r4r r r r!s    "&9&