a  d@sddlmZmZmZddlZddlmZddlmZddl m Z ee deddd Z deeee d ed d dZ eee eedddZdeeeeedddZdS))AnyListOptionalN)Tensor) functional)Literal)elementwise_meansumnoneN)x reductionreturncCsD|dkrt|S|dks"|dur&|S|dkr8t|StddS)aBReduces a given tensor by a given reduction method. Args: x: the tensor, which shall be reduced reduction: a string specifying the reduction method ('elementwise_mean', 'none', 'sum') Return: reduced Tensor Raise: ValueError if an invalid reduction parameter was given rr Nr zReduction parameter unknown.)torchmeanr ValueError)r r rk/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/torchmetrics/utilities/distributed.pyreduces  rr micromacroweightedr N)numdenomweightsclass_reductionr cCsd}|dkr t|t|n||}d|||k<|dkr@|S|dkrRt|S|dkrvt||t|S|dks|dur|Std|d |dS) aReduce classification metrics of the form ``num / denom * weights``. For example for calculating standard accuracy the num would be number of true positives per class, denom would be the support per class, and weights would be a tensor of 1s. Args: num: numerator tensor denom: denominator tensor weights: weights for each class class_reduction: reduction method for multiclass problems: - ``'micro'``: calculate metrics globally (default) - ``'macro'``: calculate metrics for each label, and find their unweighted mean. - ``'weighted'``: calculate metrics for each label, and find their weighted mean. - ``'none'`` or ``None``: returns calculated metric per class Raises: ValueError: If ``class_reduction`` is none of ``"micro"``, ``"macro"``, ``"weighted"``, ``"none"`` or ``None``. rrrrrr NzReduction parameter z' unknown. Choose between one of these: )rr rfloatr)rrrrZvalid_reductionfractionrrr class_reduce-s$  r)resultgroup world_sizer cs*fddt|D}tj|||S)Ncsg|]}tqSrrZ zeros_like.0_rrr \z._simple_gather_all_tensors..)ranger distributed all_gather)rr r!gathered_resultrr&r_simple_gather_all_tensors[sr-)rr r c s`|durtjjj}|}tj|}tjj|d|jdkrJt|||Stj |j |j dfddt |D}tjj ||dt|jddjtfdd |D}|rt|||Sg}}t|D]}|d||qt||fd dt |D}tj ||t|D](\} } d d| D} || | || <q2|S) aeGather all tensors from several ddp processes onto a list that is broadcasted to all processes. Works on tensors that have the same number of dimensions, but where each dimension may differ. In this case tensors are padded, gathered and then trimmed to secure equal workload for all processes. Args: result: the value to sync group: the process group to gather results from. Defaults to all processes (world) Return: gathered_result: list with size equal to the process group where ``gathered_result[i]`` corresponds to result tensor from process ``i`` N)r r)devicecsg|]}tqSrr"r#) local_sizerrr'r(z&gather_all_tensors..)Zdimc3s|]}t|kVqdS)N)all)r$Zls)max_sizerr r(z%gather_all_tensors..csg|]}tqSrr"r#) result_paddedrrr'r(cSsg|] }t|qSr)slice)r$Zdim_sizerrrr'r()rr*r ZWORLD contiguousZget_world_sizeZbarrierndimr-Ztensorshaper.r)r+stackmaxvaluesr0detachcpureversedappenditemFpad enumerate) rr r!Z local_sizesZall_sizes_equalZpad_dimsZpad_byvalr,idxZ item_sizeZ slice_paramr)r/r1r3rgather_all_tensorsas4        rE)r )N)typingrrrrrZtorch.nnrr@Ztyping_extensionsrrrintr-rErrrrs    .