o .wij&@sUddlmZddlmZmZddlZddlZddlmZddl m Z ddl m Z ddl mZiZeed<d ed ejd efd d Zdeded eeeffddZdeded eeeffddZ  d"dededede dde dded eeeffddZdeded efd d!ZdS)#) permutations)AnyCallableN)Tensor)Literal)rank_zero_warn)_SCIPY_AVAILABLE_ps_dictspk_numdevicereturncCsJt|t|}|tvrtjttt||d}|t|<|St|}|S)N)r )strr torchtensorlistrrange)r r keypsr^/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/torchmetrics/functional/audio/pit.py_gen_permutationssr metric_mtx eval_funccsvddlm|}ttfdd|D}||j }t |d|dddddf ddg}||fS) aSolves the linear sum assignment problem. This implementation uses scipy and input is therefore transferred to cpu during calculations. Args: metric_mtx: the metric matrix, shape [batch_size, spk_num, spk_num] eval_func: the function to reduce the metric values of different the permutations Returns: best_metric: shape ``[batch]`` best_perm: shape ``[batch, spk]`` r)linear_sum_assignmentcs g|] }|tjkdqS))rmax).0pwmrrrr >s z<_find_best_perm_by_linear_sum_assignment..N) scipy.optimizerdetachcpurrnparraytor gathermean)rrmmtx best_perm best_metricrrr(_find_best_perm_by_linear_sum_assignment*s    *r.c Cs|jdd\}}t||jd}|jd}|jd|||}t|d|}|jdd}||dd\} } | } || ddf} | | fS)aSolves the linear sum assignment problem using exhaustive method. This is done by exhaustively calculating the metric values of all possible permutations, and returns the best metric values and the corresponding permutations. Args: metric_mtx: the metric matrix, shape ``[batch_size, spk_num, spk_num]`` eval_func: the function to reduce the metric values of different the permutations Returns: best_metric: shape ``[batch]`` best_perm: shape ``[batch, spk]`` Nr r r r)N.rdim) shaperr Texpandrr)r*r$) rr batch_sizer rperm_numbpsmetric_of_ps_details metric_of_psr- best_indexesr,rrr$_find_best_perm_by_exhaustive_methodDs  r; speaker-wiserpredstarget metric_funcmoder<permutation-wiserminkwargscKsx|jdd|jddkrtd|dvrtd||dvr(td||jdkr:td|jd |jd |d krAtjntj}|jdd\}}|d krt||jd } | jd} tj |d| ddj || g|jddR} |j | dd} || | fi|} tj | |t | ddd} || dd\}}|}| |ddf}||fS||ddddf|ddddffi|}tj|||f|j|jd}||ddddf<t|D]0}t|D])}|dkr|dkrq||dd|df|dd|dffi||dd||f<qq|dksts1|dkr&ts&td|dt||\}}||fSt||\}}||fS)aUCalculate `Permutation invariant training`_ (PIT). This metric can evaluate models for speaker independent multi-talker speech separation in a permutation invariant way. Args: preds: float tensor with shape ``(batch_size,num_speakers,...)`` target: float tensor with shape ``(batch_size,num_speakers,...)`` metric_func: a metric function accept a batch of target and estimate. if `mode`==`'speaker-wise'`, then ``metric_func(preds[:, i, ...], target[:, j, ...])`` is called and expected to return a batch of metric tensors ``(batch,)``; if `mode`==`'permutation-wise'`, then ``metric_func(preds[:, p, ...], target[:, :, ...])`` is called, where `p` is one possible permutation, e.g. [0,1] or [1,0] for 2-speaker case, and expected to return a batch of metric tensors ``(batch,)``; mode: can be `'speaker-wise'` or `'permutation-wise'`. eval_func: the function to find the best permutation, can be ``'min'`` or ``'max'``, i.e. the smaller the better or the larger the better. kwargs: Additional args for metric_func Returns: Tuple of two float tensors. First tensor with shape ``(batch,)`` contains the best metric value for each sample and second tensor with shape ``(batch,)`` contains the best permutation. Example: >>> from torchmetrics.functional.audio import scale_invariant_signal_distortion_ratio >>> # [batch, spk, time] >>> preds = torch.tensor([[[-0.0579, 0.3560, -0.9604], [-0.1719, 0.3205, 0.2951]]]) >>> target = torch.tensor([[[ 1.0958, -0.1648, 0.5228], [-0.4100, 1.1942, -0.5103]]]) >>> best_metric, best_perm = permutation_invariant_training( ... preds, target, scale_invariant_signal_distortion_ratio, ... mode="speaker-wise", eval_func="max") >>> best_metric tensor([-5.1091]) >>> best_perm tensor([[0, 1]]) >>> pit_permutate(preds, best_perm) tensor([[[-0.0579, 0.3560, -0.9604], [-0.1719, 0.3205, 0.2951]]]) rr z_Predictions and targets are expected to have the same shape at the batch and speaker dimensionsrCz-eval_func can only be "max" or "min" but got rAz>mode can only be "speaker-wise" or "permutation-wise" but got z/Inputs must be of shape [batch, spk, ...], got z and z insteadrrBr/rr!)r1indexN)repeatsr1r0.)dtyper zIn pit metric for speaker-num z8>3, we recommend installing scipy for better performance)r2 RuntimeError ValueErrorndimrrrDrr index_selectreshaperepeat_interleaver*lenr$emptyrHrrrr;r.)r=r>r?r@rrEeval_opr5r permsr6ppredsptargetr9r-r:r, first_eler target_idx preds_idxrrrpermutation_invariant_trainingksb2   .    rYpermcCstddt||DS)a"Permutate estimate according to perm. Args: preds: the estimates you want to permutate, shape [batch, spk, ...] perm: the permutation returned from permutation_invariant_training, shape [batch, spk] Returns: Tensor: the permutated version of estimate cSsg|] \}}t|d|qS)r)rrM)rpredprrrrsz!pit_permutate..)rstackzip)r=rZrrr pit_permutates r_)r<r) itertoolsrtypingrrnumpyr&rrtyping_extensionsrtorchmetrics.utilitiesrtorchmetrics.utilities.importsrr dict__annotations__intr rtupler.r;rYr_rrrrsT           +  m