o si(@s>ddlZddlmZddlZddlmZGdddejZdS)N) combinations)nncs\eZdZdZdfdd ZdddZed d Zed d Zed dZ eddZ Z S)MixITLossWrapperaMixture invariant loss wrapper. Args: loss_func: function with signature (est_targets, targets, **kwargs). generalized (bool): Determines how MixIT is applied. If False , apply MixIT for any number of mixtures as soon as they contain the same number of sources (:meth:`~MixITLossWrapper.best_part_mixit`.) If True (default), apply MixIT for two mixtures, but those mixtures do not necessarly have to contain the same number of sources. See :meth:`~MixITLossWrapper.best_part_mixit_generalized`. reduction (string, optional): Specifies the reduction to apply to the output: ``'none'`` | ``'mean'``. ``'none'``: no reduction will be applied, ``'mean'``: the sum of the output will be divided by the number of elements in the output. For each of these modes, the best partition and reordering will be automatically computed. Examples: >>> import torch >>> from asteroid.losses import multisrc_mse >>> mixtures = torch.randn(10, 2, 16000) >>> est_sources = torch.randn(10, 4, 16000) >>> # Compute MixIT loss based on pairwise losses >>> loss_func = MixITLossWrapper(multisrc_mse) >>> loss_val = loss_func(est_sources, mixtures) References [1] Scott Wisdom et al. "Unsupervised sound separation using mixtures of mixtures." arXiv:2006.12701 (2020) Tmeancs t||_||_||_dS)N)super__init__ loss_func generalized reduction)selfrr r  __class__Q/home/ubuntu/.local/lib/python3.10/site-packages/asteroid/losses/mixit_wrapper.pyr)s  zMixITLossWrapper.__init__Fc Ks|jd|jdks J|jd|jdksJ|js+|j|j||fi|\}}}n|j|j||fi|\}}}|jdkrC|n|}|sI|S|||||} || fS)atFind the best partition and return the loss. Args: est_targets: torch.Tensor. Expected shape :math:`(batch, nsrc, *)`. The batch of target estimates. targets: torch.Tensor. Expected shape :math:`(batch, nmix, ...)`. The batch of training targets return_est: Boolean. Whether to return the estimated mixtures estimates (To compute metrics or to save example). **kwargs: additional keyword argument that will be passed to the loss function. Returns: - Best partition loss for each batch sample, average over the batch. torch.Tensor(loss_value) - The estimated mixtures (estimated sources summed according to the partition) if return_est is True. torch.Tensor of shape :math:`(batch, nmix, ...)`. rr)shaper best_part_mixitrbest_part_mixit_generalizedr rreorder_source) r est_targetstargets return_estkwargsmin_loss min_loss_idxparts returned_loss reorderedrrrforward/s$ zMixITLossWrapper.forwardc  s|jd}|jd}||dkrtd||}fddtt|||}tj|f|||d|}tj|ddd\} } | | |fS) a0Find best partition of the estimated sources that gives the minimum loss for the MixIT training paradigm in [1]. Valid for any number of mixtures as soon as they contain the same number of sources. Args: loss_func: function with signature ``(est_targets, targets, **kwargs)`` The loss function to get batch losses from. est_targets: torch.Tensor. Expected shape :math:`(batch, nsrc, ...)`. The batch of target estimates. targets: torch.Tensor. Expected shape :math:`(batch, nmix, ...)`. The batch of training targets (mixtures). **kwargs: additional keyword argument that will be passed to the loss function. Returns: - :class:`torch.Tensor`: The loss corresponding to the best permutation of size (batch,). - :class:`torch.LongTensor`: The indices of the best partition. - :class:`list`: list of the possible partitions of the sources. rz>The mixtures are assumed to contain the same number of sourcesc3s`|dkr gVdSt||D]fdd|D}|||dD] }tg|Vq"qdS)Nrcg|]}|vr|qSrr.0xcrr zIMixITLossWrapper.best_part_mixit..parts_mixit..r)rlist)lstklrestr parts_mixitr$rr/}s z5MixITLossWrapper.best_part_mixit..parts_mixitrrrTdimkeepdim)r ValueErrorr(rangerloss_set_from_partstorchmin) rrrrnmixnsrcnsrcmixrloss_setrmin_loss_indexesrr.rrXs      z MixITLossWrapper.best_part_mixitc Ksp|jd}|jd}|dkrtddd}|t|}tj|f|||d|}tj|ddd\} } | | |fS) aFind best partition of the estimated sources that gives the minimum loss for the MixIT training paradigm in [1]. Valid only for two mixtures, but those mixtures do not necessarly have to contain the same number of sources e.g the case where one mixture is silent is allowed.. Args: loss_func: function with signature ``(est_targets, targets, **kwargs)`` The loss function to get batch losses from. est_targets: torch.Tensor. Expected shape :math:`(batch, nsrc, ...)`. The batch of target estimates. targets: torch.Tensor. Expected shape :math:`(batch, nmix, ...)`. The batch of training targets (mixtures). **kwargs: additional keyword argument that will be passed to the loss function. Returns: - :class:`torch.Tensor`: The loss corresponding to the best permutation of size (batch,). - :class:`torch.LongTensor`: The indexes of the best permutations. - :class:`list`: list of the possible partitions of the sources. rrzWorks only with two mixturescsRg}tt|dD]}t||D]fdd|D}|t|gqq |S)Nrcr rrr!r$rrr&r'zYMixITLossWrapper.best_part_mixit_generalized..parts_mixit_gen..)r5lenrappendr()r) partitionsr*r,rr$rparts_mixit_genszEMixITLossWrapper.best_part_mixit_generalized..parts_mixit_genr0Tr1)rr4r5rr6r7r8) rrrrr9r:rArr<rr=rrrrs    z,MixITLossWrapper.best_part_mixit_generalizedc  svg}|D]-}tjfdd|Ddd}|||fi|}|jdkr&td||dddfqtj|dd}|S)z(Common loop between both best_part_mixitcs(g|]}dd|ddfdqSNrsumr"idxrrrr&s(z8MixITLossWrapper.loss_set_from_parts..rr2z6Loss function return value should be of size (batch,).N)r7stackndimr4r?cat) rrrrrr< partition est_mixesloss_partitionrrGrr6s z$MixITLossWrapper.loss_set_from_partscsXt|}t|D] \}||}tjfdd|Ddd|ddddf<q |S)aIReorder sources according to the best partition. Args: est_targets: torch.Tensor. Expected shape :math:`(batch, nsrc, ...)`. The batch of target estimates. targets: torch.Tensor. Expected shape :math:`(batch, nmix, ...)`. The batch of training targets. min_loss_idx: torch.LongTensor. The indexes of the best permutations. parts: list of the possible partitions of the sources. Returns: :class:`torch.Tensor`: Reordered sources of shape :math:`(batch, nmix, time)`. cs6g|]}|ddfdddddfdqSrBrCrEbrrrr&s6z3MixITLossWrapper.reorder_source..rrHN)r7 zeros_like enumeraterI)rrrrorderedrFright_partitionrrOrrs zMixITLossWrapper.reorder_source)Tr)F) __name__ __module__ __qualname____doc__rr staticmethodrrr6r __classcell__rrr rrs! ) 7 3 r)warnings itertoolsrr7rModulerrrrrs