o yi @sddlmZmZmZmZmZmZmZmZddl Z ddl m Z ddl m Z m Z dZdee ee fde fdd Zde de fd d Zde de fd d Zde de fddZde de fddZdedefddZdedefddZ d9de deede fddZd:de dedede fddZd;de d ede fd!d"Zdd#d$ed%eeefd&ed'ed(eeeefd)edefd*d+Zde de fd,d-Zd$edefd.d/Zd9de d0eede fd1d2Z de de fd3d4Z!d5e d6e de"fd7d8Z#dS)<)AnyCallableDictListMappingOptionalSequenceUnionN)Tensor)_TORCH_GREATER_EQUAL_1_12_XLA_AVAILABLEgư>xreturncCs@t|ttfr |n|g}dd|D}|stdtj|ddS)z'Concatenation along the zero dimension.cSs0g|]}|dkr|jdkr|dn|qS)r)numelndim unsqueeze).0yrO/home/ubuntu/.local/lib/python3.10/site-packages/torchmetrics/utilities/data.py s0z dim_zero_cat..zNo samples to concatenaterdim) isinstancelisttuple ValueErrortorchcatr rrr dim_zero_cats r!cCtj|ddS)z#Summation along the zero dimension.rr)rsumr rrr dim_zero_sum!r$cCr")z!Average along the zero dimension.rr)rmeanr rrr dim_zero_mean&r%r'cCtj|ddjS)zMax along the zero dimension.rr)rmaxvaluesr rrr dim_zero_max+r+cCr()zMin along the zero dimension.rr)rminr*r rrr dim_zero_min0r,r.cCsdd|DS)z&Flatten list of list into single list.cSsg|] }|D]}|qqSrr)rsublistitemrrrr7sz_flatten..rr rrr_flatten5r%r1cCsHi}|D]\}}t|tr|D]\}}|||<qq|||<q|S)z'Flatten dict of dicts into single dict.)itemsrdict)r new_dictkeyvaluekvrrr _flatten_dict:s   r9 label_tensor num_classescCsr|durt|d}tj|jd|g|jddR|j|jd}| d |}| d|dS)aConverts a dense label tensor to one-hot format. Args: label_tensor: dense label tensor, with shape [N, d1, d2, ...] num_classes: number of classes C Returns: A sparse label tensor with shape [N, C, d1, d2, ...] Example: >>> x = torch.tensor([1, 2, 3]) >>> to_onehot(x) tensor([[0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]) Nrr)dtypedevice?) intr)detachr0rzerosshaper<r=longr expand_asscatter_)r:r; tensor_onehotindexrrr to_onehotFs rHr prob_tensortopkrcCsTt|}|dkr|||j|ddd}|S|||j||djd}|S)avConvert a probability tensor to binary by selecting top-k the highest entries. Args: prob_tensor: dense tensor of shape ``[..., C, ...]``, where ``C`` is in the position defined by the ``dim`` argument topk: number of the highest entries to turn into 1s dim: dimension on which to compare entries Returns: A binary tensor of the same shape as the input tensor of type ``torch.int32`` Example: >>> x = torch.tensor([[1.1, 2.0, 3.0], [2.0, 1.0, 0.5]]) >>> select_topk(x, topk=2) tensor([[0, 1, 1], [1, 1, 0]], dtype=torch.int32) rT)rkeepdimr>)r7r)r zeros_likescatterargmaxrJindicesr?)rIrJrrA topk_tensorrrr select_topkhs rQ argmax_dimcCstj||dS)avConverts a tensor of probabilities to a dense label tensor. Args: x: probabilities to get the categorical label [N, d1, d2, ...] argmax_dim: dimension to apply Return: A tensor with categorical labels [N, d2, ...] Example: >>> x = torch.tensor([[0.2, 0.5], [0.9, 0.1]]) >>> to_categorical(x) tensor([1, 0]) r)rrN)r rRrrrto_categoricalsrS) wrong_dtypedatar<functionargsrTkwargscst|}t|r|dust||s|gRiSt|tr2|fdd|DSt|trJt|drJ|fdd|DSt|trbt|tsb|fdd|DS|S) aRecursively applies a function to all elements of a certain dtype. Args: data: the collection to apply the function to dtype: the given function will be applied to all elements of this dtype function: the function to apply *args: positional arguments (will be forwarded to call of ``function``) wrong_dtype: the given function won't be applied if this type is specified and the given collections is of the :attr:`wrong_type` even if it is of type :attr`dtype` **kwargs: keyword arguments (will be forwarded to call of ``function``) Returns: the resulting collection Example: >>> apply_to_collection(torch.tensor([8, 0, 2, 6, 7]), dtype=Tensor, function=lambda x: x ** 2) tensor([64, 0, 4, 36, 49]) >>> apply_to_collection([8, 0, 2, 6, 7], dtype=int, function=lambda x: x ** 2) [64, 0, 4, 36, 49] >>> apply_to_collection(dict(abc=123), dtype=int, function=lambda x: x ** 2) {'abc': 15129} Ncs,i|]\}}|t|gRiqSrapply_to_collection)rr7r8rWr<rVrXrr s,z'apply_to_collection.._fieldsc3s*|]}t|gRiVqdSNrYrdr[rr s(z&apply_to_collection..cs&g|]}t|gRiqSrrYr_r[rrrs&z'apply_to_collection..)typerrr2rhasattrrstr)rUr<rVrTrWrX elem_typerr[rrZs  rZcCs|dkr |S|S)Nr)rsqueezer rrr_squeeze_scalar_element_tensorsrgcCs t|ttSr^)rZr rg)rUrrr_squeeze_if_scalars rh minlengthcCsp|dur tt|}tststr1|jr1tj||jtj d}t |D] }||k ||<q$|Stj ||dS)agPyTorch currently does not support``torch.bincount`` for: - deterministic mode on GPU. - MPS devices This implementation fallback to a for-loop counting occurrences in that case. Args: x: tensor to count minlength: minimum length to count Returns: Number of occurrences for each unique element in x N)r=r<ri) lenrunique$are_deterministic_algorithms_enabledr r is_mpsrAr=rCranger#bincount)r rioutputirrr _bincounts rscCs4||}t|}t|t|dd}||S)zSimilar to `_bincount`, but works also with tensor that do not contain continuous values. Args: x: tensor to count Returns: Number of occurrences for each unique element in x rrj)r-rrlrsr))r unique_xrqrrr_flexible_bincounts rutensor1tensor2cCs&|j|jkr |j|jd}t||S)zBWrapper of torch.allclose that is robust towards dtype difference.)r<)r<torallclose)rvrwrrrrys  ryr^)rr)r)$typingrrrrrrrr rr torchmetrics.utilities.importsr r METRIC_EPSr!r$r'r+r.rr1r9r?rHrQrSrbrrZrgrhrsruboolryrrrrsV(    "  2