o si @s ddlZdddZdddZdS) Nc Cs t|}|j\}}}|durt|||d}|}t|jdkr-||||d}||j}tj |||||jd}| d||||dd||}||}t d||}t d||} t d||} t |t | } | dt | } | tj |ddgdS)a8Compute the deep clustering loss defined in [1]. Args: embedding (torch.Tensor): Estimated embeddings. Expected shape :math:`(batch, frequency * frame, embedding\_dim)`. tgt_index (torch.Tensor): Dominating source index in each TF bin. Expected shape: :math:`(batch, frequency, frame)`. binary_mask (torch.Tensor): VAD in TF plane. Bool or Float. See asteroid.dsp.vad.ebased_vad. Returns: `torch.Tensor`. Deep clustering loss for every batch sample. Examples >>> import torch >>> from asteroid.losses.cluster import deep_clustering_loss >>> spk_cnt = 3 >>> embedding = torch.randn(10, 5*400, 20) >>> targets = torch.LongTensor(10, 400, 5).random_(0, spk_cnt) >>> loss = deep_clustering_loss(embedding, targets) Reference [1] Zhong-Qiu Wang, Jonathan Le Roux, John R. Hershey "ALTERNATIVE OBJECTIVE FUNCTIONS FOR DEEP CLUSTERING" .. note:: Be careful in viewing the embedding tensors. The target indices ``tgt_index`` are of shape :math:`(batch, freq, frames)`. Even if the embedding is of shape :math:`(batch, freq * frames, emb)`, the underlying view should be :math:`(batch, freq, frames, emb)` and not :math:`(batch, frames, freq, emb)`. N)devicez ijk,ijl->ikl)dim)lenuniqueshapetorchonesfloatviewtorzerosscatter_einsumbatch_matrix_normsum) embedding tgt_index binary_maskspk_cntbatchbinsframes tgt_embeddingest_proj true_proj true_est_projcostr K/home/ubuntu/.local/lib/python3.10/site-packages/asteroid/losses/cluster.pydeep_clustering_losss$  r"rcCs$ttd|j}tj|||d|S)zNormalize a matrix according to `norm_order` Args: matrix (torch.Tensor): Expected shape [batch, *] norm_order (int): Norm order. Returns: torch.Tensor, normed matrix of shape [batch] r)pr)listrangendimr norm)matrix norm_order keep_batchr r r!r@s r)N)r)r r"rr r r r!s <