o si@s6ddlZddlmZddlmZGdddejZdS)N)nn)PITLossWrappercsTeZdZdZdfdd ZeddZejddZdd d Ze dd dZ Z S)SinkPITLossWrappera Permutation invariant loss wrapper. Args: loss_func: function with signature (targets, est_targets, **kwargs). n_iter (int): number of the Sinkhorn iteration (default = 200). Supposed to be an even number. hungarian_validation (boolean) : Whether to use the Hungarian algorithm for the validation. (default = True) ``loss_func`` computes pairwise losses and returns a torch.Tensor of shape :math:`(batch, n\_src, n\_src)`. Each element :math:`(batch, i, j)` corresponds to the loss between :math:`targets[:, i]` and :math:`est\_targets[:, j]` It evaluates an approximate value of the PIT loss using Sinkhorn's iterative algorithm. See :meth:`~PITLossWrapper.best_softperm_sinkhorn` and http://arxiv.org/abs/2010.11871 Examples >>> import torch >>> import pytorch_lightning as pl >>> from asteroid.losses import pairwise_neg_sisdr >>> sources = torch.randn(10, 3, 16000) >>> est_sources = torch.randn(10, 3, 16000) >>> # Compute SinkPIT loss based on pairwise losses >>> loss_func = SinkPITLossWrapper(pairwise_neg_sisdr) >>> loss_val = loss_func(est_sources, sources) >>> # A fixed temperature parameter `beta` (=10) is used >>> # unless a cooling callback is set. The value can be >>> # dynamically changed using a cooling callback module as follows. >>> model = NeuralNetworkModel() >>> optimizer = optim.Adam(model.parameters(), lr=1e-3) >>> dataset = YourDataset() >>> loader = data.DataLoader(dataset, batch_size=16) >>> system = System( >>> model, >>> optimizer, >>> loss_func=SinkPITLossWrapper(pairwise_neg_sisdr), >>> train_loader=loader, >>> val_loader=loader, >>> ) >>> >>> trainer = pl.Trainer( >>> max_epochs=100, >>> callbacks=[SinkPITBetaScheduler(lambda epoch : 1.02 ** epoch)], >>> ) >>> >>> trainer.fit(system) Tcs&t||_d|_||_||_dS)N )super__init__ loss_func_betan_iterhungarian_validation)selfr r r  __class__S/home/ubuntu/.local/lib/python3.10/site-packages/asteroid/losses/sinkpit_wrapper.pyr 7s  zSinkPITLossWrapper.__init__cCs|jS)Nr )rrrrbeta>szSinkPITLossWrapper.betacCs|dksJ||_dS)Nrr)rrrrrrBs  Fc Ks|jd}|dksJd||j||fi|}|jdks#Jd|jd|jdks1Jd|sY|js9|jsK|||j|j\}}t |} | St |\}} t |} | St |\}} t |} t || } | | fS)aOEvaluate the loss using Sinkhorn's algorithm. Args: est_targets: torch.Tensor. Expected shape :math:`(batch, nsrc, ...)`. The batch of target estimates. targets: torch.Tensor. Expected shape :math:`(batch, nsrc, ...)`. The batch of training targets return_est: Boolean. Whether to return the reordered targets estimates (To compute metrics or to save example). **kwargs: additional keyword argument that will be passed to the loss function. Returns: - Best permutation loss for each batch sample, average over the batch. torch.Tensor(loss_value) - The reordered targets estimates if return_est is True. torch.Tensor of shape :math:`(batch, nsrc, ...)`. rdz(Expected source axis along dim 1, found zBSomething went wrong with the loss function, please read the docs.rz&PIT loss needs same batch dim as input) shaper ndimtrainingr best_softperm_sinkhornr r torchmeanrfind_best_permreorder_source) r est_targetstargets return_estkwargsn_src pw_lossesmin_loss soft_perm mean_loss batch_indices reorderedrrrforwardGs*        zSinkPITLossWrapper.forwardrcCs|dd}|jd}| |}t|dD]}|tj|ddd}|tj|ddd}qtd|||t|}||}|t|fS)apCompute an approximate PIT loss using Sinkhorn's algorithm. See http://arxiv.org/abs/2010.11871 Args: pair_wise_losses (:class:`torch.Tensor`): Tensor of shape :math:`(batch, n_src, n_src)`. Pairwise losses. beta (float) : Inverse temperature parameter. (default = 10) n_iter (int) : Number of iteration. Even number. (default = 200) Returns: - :class:`torch.Tensor`: The loss corresponding to the best permutation of size (batch,). - :class:`torch.Tensor`: A soft permutation matrix. rT)axiskeepdimz bij,bij->b) transposerranger logsumexpeinsumexp)pair_wise_lossesrr Cr#Zitr%rrrrys   z)SinkPITLossWrapper.best_softperm_sinkhorn)rT)F)rr) __name__ __module__ __qualname____doc__r propertyrsetterr* staticmethodr __classcell__rrrrrs/   2r)rrrModulerrrrrs