o yiL%@sddlZddlZddlmZmZddlZddlmZddlmZddl m Z ddl m Z ej jZe r9ddlmZndZded efd d Zd ed eded eeeffddZ    dd ed edeedededeed efddZdd ed eded efddZdS)N)OptionalTuple)Tensor)norm)_check_same_shape)_FAST_BSS_EVAL_AVAILABLE)toeplitz_conjugate_gradientvectorreturncCsltjtj|dd|dddfgdd}|jd}tj||jdd||f|dddd jddS) aConstruct a symmetric Toeplitz matrix using one vector. Args: vector: shape [..., L] Example: >>> from torchmetrics.functional.audio.sdr import _symmetric_toeplitz >>> import torch >>> v = torch.tensor([0, 1, 2, 3, 4]) >>> _symmetric_toeplitz(v) tensor([[0, 1, 2, 3, 4], [1, 0, 1, 2, 3], [2, 1, 0, 1, 2], [3, 2, 1, 0, 1], [4, 3, 2, 1, 0]]) Returns: a symmetric Toeplitz matrix of shape [..., L, L] ))dims.Nr dim)r r )sizestride)torchcatflipshape as_stridedr)r vec_expv_lenrU/home/ubuntu/.local/lib/python3.10/site-packages/torchmetrics/functional/audio/sdr.py_symmetric_toeplitz$s( (rtargetpredscorr_lencCsdtt|jd|jdd}tjj||dd}tjj|jd|j d|ddd|f}tjj||dd}tjj| ||dddd|f}||fS)a Compute the auto correlation of `target` and the cross correlation of `target` and `preds` using the fast Fourier transform (FFT). Let's denotes the symmetric Toeplitz matric of the auto correlation of `target` as `R`, the cross correlation as 'b', then solving the equation `Rh=b` could have `h` as the coordinate of `preds` in the column space of the `corr_len` shifts of `target`. Args: target: the target (reference) signal of shape [..., time] preds: the preds (estimated) signal of shape [..., time] corr_len: the length of the auto correlation and cross correlation Returns: the auto correlation of `target` of shape [..., corr_len] the cross correlation of `target` and `preds` of shape [..., corr_len] r r )nr)r .N) mathceillog2rrfftrfftirfftrealimagconj)rrrn_fftt_fftr_0p_fftbrrr_compute_autocorr_crosscorr?s (,&r/F use_cg_iter filter_length zero_mean load_diagcCs.t|||j}|}|}|r$||jddd}||jddd}|tjt|ddddd}|tjt|ddddd}t|||d\}}|durU|d|7<|durctrct |||d } n|durotsot d t t |} t| |} td || } | d | } d t| } |tjkr| S| S)ay Calculates Signal to Distortion Ratio (SDR) metric. See `SDR ref1`_ and `SDR ref2`_ for details on the metric. .. note: The metric currently does not seem to work with Pytorch v1.11 and specific GPU hardware. Args: preds: float tensor with shape ``(...,time)`` target: float tensor with shape ``(...,time)`` use_cg_iter: If provided, conjugate gradient descent is used to solve for the distortion filter coefficients instead of direct Gaussian elimination, which requires that ``fast-bss-eval`` is installed and pytorch version >= 1.8. This can speed up the computation of the metrics in case the filters are long. Using a value of 10 here has been shown to provide good accuracy in most cases and is sufficient when using this loss to train neural separation networks. filter_length: The length of the distortion filter allowed zero_mean: When set to True, the mean of all signals is subtracted prior to computation of the metrics load_diag: If provided, this small value is added to the diagonal coefficients of the system metrics when solving for the filter coefficients. This can help stabilize the metric in the case where some reference signals may sometimes be zero Returns: Float tensor with shape ``(...,)`` of SDR values per sample Raises: RuntimeError: If ``preds`` and ``target`` does not have the same shape Example: >>> from torchmetrics.functional.audio import signal_distortion_ratio >>> import torch >>> g = torch.manual_seed(1) >>> preds = torch.randn(8000) >>> target = torch.randn(8000) >>> signal_distortion_ratio(preds, target) tensor(-12.0589) >>> # use with permutation_invariant_training >>> from torchmetrics.functional.audio import permutation_invariant_training >>> preds = torch.randn(4, 2, 8000) # [batch, spk, time] >>> target = torch.randn(4, 2, 8000) >>> best_metric, best_perm = permutation_invariant_training(preds, target, signal_distortion_ratio, 'max') >>> best_metric tensor([-11.6375, -11.4358, -11.7148, -11.6325]) >>> best_perm tensor([[1, 0], [0, 1], [1, 0], [0, 1]]) r Trkeepdimgư>)min)rN).r)n_itera The `use_cg_iter` parameter of `SDR` requires that `fast-bss-eval` is installed. To make this this warning disappear, you could install `fast-bss-eval` using `pip install fast-bss-eval` or set `use_cg_iter=None`. For this time, the solver provided by Pytorch is used.z...l,...l->...r g$@)rdtypedoublemeanrclamprr/rrwarningswarn UserWarningrsolveeinsumlog10float64float)rrr1r2r3r4 preds_dtyper,r.solrcohratiovalrrrsignal_distortion_ratio]s8 <    rKcCst||t|jj}|r"|tj|ddd}|tj|ddd}tj||ddd|tj|dddd|}||}||}tj|ddd|tj|ddd|}dt|}|S)a=`Scale-invariant signal-to-distortion ratio`_ (SI-SDR). The SI-SDR value is in general considered an overall measure of how good a source sound. Args: preds: float tensor with shape ``(...,time)`` target: float tensor with shape ``(...,time)`` zero_mean: If to zero mean target and preds or not Returns: Float tensor with shape ``(...,)`` of SDR values per sample Raises: RuntimeError: If ``preds`` and ``target`` does not have the same shape Example: >>> from torchmetrics.functional.audio import scale_invariant_signal_distortion_ratio >>> target = torch.tensor([3.0, -0.5, 2.0, 7.0]) >>> preds = torch.tensor([2.5, 0.0, 2.0, 8.0]) >>> scale_invariant_signal_distortion_ratio(preds, target) tensor(18.4030) r Tr5rr )rrfinfor9epsr;sumrB)rrr3rNalpha target_scalednoiserJrrr'scale_invariant_signal_distortion_ratios ,rS)Nr0FN)F)r!r=typingrrrr torch.linalgrtorchmetrics.utilities.checksrtorchmetrics.utilities.importsrlinalgr@fast_bss_eval.torch.cgdrrintr/boolrDrKrSrrrrsD    "! q