o .wi @srddlmZmZddlZddlmZddlmZddlmZdededefd d Z dededefd d Z d3dededee edfdefddZ d4dedee dedededededefddZdededeeeffddZd5deded!e d"edef d#d$Zd6deded&edefd'd(Zd6deded&edefd)d*Zded+ededefd,d-Zd.ed/ed0defd1d2ZdS)7)OptionalUnionN)Tensor)Literal)rank_zero_warnxyreturncCs8|jtjks |jtjkr||jS||jS)zSafe calculation of matrix multiplication. If input is float16, will cast to float32 for computation and back again. )dtypetorchfloat16floatThalfrrr[/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/torchmetrics/utilities/compute.py _safe_matmuls rcCs|t|}d||dk<|S)zCompute x * log(y). Returns 0 if x=0. Example: >>> import torch >>> x = torch.zeros(1) >>> _safe_xlogy(x, 1/x) tensor([0.]) r)r log)rrresrrr _safe_xlogy"s rrnumdenom zero_division)warnnancCs|r|n|}|r|n|}t|ttfs|dkrS|dkr.t|dkr.td|dkr4dn|}tj||jdj |j |j j dkd}t |dk|||St ||S)a Safe division, by preventing division by zero. Function will cast to float if input is not already to secure backwards compatibility. Args: num: numerator tensor denom: denominator tensor, which may contain zeros zero_division: value to replace elements divided by zero Example: >>> import torch >>> num = torch.tensor([1.0, 2.0, 3.0]) >>> denom = torch.tensor([0.0, 1.0, 2.0]) >>> _safe_divide(num, denom) tensor([0.0000, 2.0000, 1.5000]) rrz:Detected zero division in _safe_divide. Setting 0/0 to 0.0r)r mps) non_blocking)is_floating_pointr isinstanceintr anyrtensorr todevicetypewhere true_divide)rrrzero_division_tensorrrr _safe_divide1s r*scoreaverage multilabeltpfpfntop_kcCsz|dus|dkr |S|dkr||}nt|}|s.d||dkr(|||dkn||dk<t|||jddddS) Nnoneweightedrr+rT)keepdim)r ones_liker*sum)r,r-r.r/r0r1r2weightsrrr_adjust_weights_safe_divideUs  (r:cCs|jdkr |n|}|jdkr|n|}|jdks |jdkr,td|jd|j||krBtd|d|||fS)z Check that auc input is correct.r+zJExpected both `x` and `y` tensor to be 1d, but got tensors with dimension z and zHExpected the same number of elements in `x` and `y` tensor but received )ndimsqueeze ValueErrornumelrrrr_auc_format_inputscsr?r5 directionaxiscCsBttj|||d|}Wd|S1swY|S)zrCompute area under the curve using the trapezoidal rule. Assumes increasing or decreasing order of `x`. dimN)r no_gradtrapz)rrr@rA auc_scorerrr_auc_compute_without_checkss  rGFreordercCst>|rtj|dd\}}||}|dd|dd}|dkr3|dkr/d}ntdd }t|||WdS1sEwYdS) zCompute area under the curve using the trapezoidal rule. Example: >>> import torch >>> x = torch.tensor([1, 2, 3, 4]) >>> y = torch.tensor([1, 2, 3, 4]) >>> _auc_compute(x, y) tensor(7.5000) T)stabler+Nr5rgz_The `x` tensor is neither increasing or decreasing. Try setting the reorder argument to `True`.g?)r rDsortr"allr=rG)rrrHx_idxdxr@rrr _auc_compute~s    $rNcCst||\}}t|||dS)a8Compute Area Under the Curve (AUC) using the trapezoidal rule. Args: x: x-coordinates, must be either increasing or decreasing y: y-coordinates reorder: if True, will reorder the arrays to make it either increasing or decreasing Return: Tensor containing AUC score )rH)r?rN)rrrHrrraucs rOxpc Cst|dd|dd|dd|dd}|dd||dd}tt|dddf|dddfdd}t|dt|d}|||||S)aiOne-dimensional linear interpolation for monotonically increasing sample points. Returns the one-dimensional piecewise linear interpolation to a function with given discrete data points :math:`(xp, fp)`, evaluated at :math:`x`. Adjusted version of this https://github.com/pytorch/pytorch/issues/50334#issuecomment-1000917964 Args: x: the :math:`x`-coordinates at which to evaluate the interpolated values. xp: the :math:`x`-coordinates of the data points, must be increasing. fp: the :math:`y`-coordinates of the data points, same length as `xp`. Returns: the interpolated values, same size as `x`. Example: >>> x = torch.tensor([0.5, 1.5, 2.5]) >>> xp = torch.tensor([1, 2, 3]) >>> fp = torch.tensor([1, 2, 3]) >>> interp(x, xp, fp) tensor([0.5000, 1.5000, 2.5000]) r+Nr5r)r*r r8geclamplen)rrPr0mbindicesrrrinterps 20rWr# normalization)sigmoidsoftmaxcCs|jtdkr$t|dk|dks"|dkr|ntj|dd}|S|dk|dkB}t||dkr>> import torch >>> tensor = torch.tensor([-1.0, 0.0, 1.0]) >>> normalize_logits_if_needed(tensor, normalization="sigmoid") tensor([0.2689, 0.5000, 0.7311]) >>> tensor = torch.tensor([[-1.0, 0.0, 1.0], [1.0, 0.0, -1.0]]) >>> normalize_logits_if_needed(tensor, normalization="softmax") tensor([[0.0900, 0.2447, 0.6652], [0.6652, 0.2447, 0.0900]]) >>> tensor = torch.tensor([0.0, 0.5, 1.0]) >>> normalize_logits_if_needed(tensor, normalization="sigmoid") tensor([0.0000, 0.5000, 1.0000]) cpurr+rYrB)r%r rKrYrZr"r')r#rX conditionrrrnormalize_logits_if_neededs r])r)r+)r5)F)typingrrr rtyping_extensionsrtorchmetrics.utilitiesrrrr r*strboolr!r:tupler?rGrNrOrWr]rrrrsR      %   !