o siS/ @slddlZddlmZddlmZGdddejjZdejdejd ed ed ef d d ZGdddejZ dS)N)nn) PITReordercsNeZdZdZ    dfdd Zdd Zd d Zed d ZddZ Z S)LambdaOverlapAddaNOverlap-add with lambda transform on segments (not scriptable). Segment input signal, apply lambda function (a neural network for example) and combine with OLA. `LambdaOverlapAdd` can be used with :mod:`asteroid.separate` and the `asteroid-infer` CLI. Args: nnet (callable): Function to apply to each segment. n_src (Optional[int]): Number of sources in the output of nnet. If None, the number of sources is determined by the network's output, but some correctness checks cannot be performed. window_size (int): Size of segmenting window. hop_size (int): Segmentation hop size. window (str): Name of the window (see scipy.signal.get_window) used for the synthesis. reorder_chunks (bool): Whether to reorder each consecutive segment. This might be useful when `nnet` is permutation invariant, as source assignements might change output channel from one segment to the next (in classic speech separation for example). Reordering is performed based on the correlation between the overlapped part of consecutive segment. Examples >>> from asteroid import ConvTasNet >>> nnet = ConvTasNet(n_src=2) >>> continuous_nnet = LambdaOverlapAdd( >>> nnet=nnet, >>> n_src=2, >>> window_size=64000, >>> hop_size=None, >>> window="hanning", >>> reorder_chunks=True, >>> enable_grad=False, >>> ) >>> # Process wav tensor: >>> wav = torch.randn(1, 1, 500000) >>> out_wavs = continuous_nnet.forward(wav) >>> # asteroid.separate.Separatable support: >>> from asteroid.separate import file_separate >>> file_separate(continuous_nnet, "example.wav") NhannTFc st|ddksJd||_||_|dur|n|d|_||_t|dd|_|rDddlm }|||j d}t |}d|_ nd|_ |d |||_||_dS) NrrzWindow size must be even in_channels) get_windowfloat32TFwindow)super__init__nnet window_sizehop_sizen_srcgetattrr scipy.signalrastypetorch from_numpy use_windowregister_bufferreorder_chunks enable_grad) selfr rrrr rrr __class__L/home/ubuntu/.local/lib/python3.10/site-packages/asteroid/dsp/overlap_add.pyr 4s     zLambdaOverlapAdd.__init__c Cs|jdksJ|\}}}tjjj|d|jdf|jdf|jdfd}g}|j d}t |D]b}| |d|f} |dkr^| jdksHJd|j durY| j d|j ksYJd| j d} | || d} |dkry|jryt| |d| |j|j} |jr| |j| } n| |j|j} || q0t| ||| |j}|dd d}tjjj||df|jdf|jdf|jdfd}|d || dS) zAHeart of the class: segment signal, apply func, combine with OLA.r kernel_sizepaddingstride.z'nnet should return (batch, n_src, time)Nr)ndimsizerr functionalunfold unsqueezerrshaperanger rreshaper_reorder_sourcesrr toappendstackpermutefoldsqueeze) rxbatchchannelsn_framesunfoldedoutn_chunks frame_idxframerrrr ola_forwardTsD     zLambdaOverlapAdd.ola_forwardcCs@tj|j||}|WdS1swYdS)zForward module: segment signal, apply func, combine with OLA. Args: x (:class:`torch.Tensor`): waveform signal of shape (batch, 1, time). Returns: :class:`torch.Tensor`: The output of the lambda OLA. N)rautogradset_grad_enabledrr>)rr5oladrrrforwards $zLambdaOverlapAdd.forwardcCs|jjSN)r sample_rate)rrrrrDszLambdaOverlapAdd.sample_ratecOs|j|g|Ri|SrC)rB)rwavargskwargsrrr _separateszLambdaOverlapAdd._separate)NrTF) __name__ __module__ __qualname____doc__r r>rBpropertyrDrH __classcell__rrrrrs2 1 rcurrentpreviousrrrc sZ|\}}|d||}|d||}||fdd}t|}|||}|||S)a Reorder sources in current chunk to maximize correlation with previous chunk. Used for Continuous Source Separation. Standard dsp correlation is used for reordering. Args: current (:class:`torch.Tensor`): current chunk, tensor of shape (batch, n_src, window_size) previous (:class:`torch.Tensor`): previous chunk, tensor of shape (batch, n_src, window_size) n_src (:class:`int`): number of sources. window_size (:class:`int`): window_size, equal to last dimension of both current and previous. hop_size (:class:`int`): hop_size between current and previous tensors. r csf|ddf}|d df}||jddd}||jddd}tj|d|ddd S)N.r T)keepdimr!r)dim)meanrsumr*)r5y overlap_frr reorder_funcs  z&_reorder_sources..reorder_func)r'r-r) rOrPrrrr6framesrXpitrrVrr.s    r.csJeZdZdZfddZddZd ddZed d Zed d Z Z S)DualPathProcessinga Perform Dual-Path processing via overlap-add as in DPRNN [1]. Args: chunk_size (int): Size of segmenting window. hop_size (int): segmentation hop size. References [1] Yi Luo, Zhuo Chen and Takuya Yoshioka. "Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation" https://arxiv.org/abs/1910.06379 cs$tt|||_||_d|_dSrC)r r[r chunk_sizer n_orig_frames)rr\rrrrr s zDualPathProcessing.__init__cCsh|\}}}|jdksJ|jd|_tjjj|d|j df|j df|j dfd}| |||j dS)ar Unfold the feature tensor from $(batch, channels, time)$ to $(batch, channels, chunksize, nchunks)$. Args: x (:class:`torch.Tensor`): feature tensor of shape $(batch, channels, time)$. Returns: :class:`torch.Tensor`: spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$. rr r!rr") r'r&r+r]rrr(r)r*r\rr-)rr5r6chanrYr9rrrr)s  zDualPathProcessing.unfoldNcCs|dur|n|j}|\}}}}||||j|}tjjj||df|jdf|jdf|jdfd}|t |j|j}||||jS)a Folds back the spliced feature tensor. Input shape $(batch, channels, chunksize, nchunks)$ to original shape $(batch, channels, time)$ using overlap-add. Args: x (:class:`torch.Tensor`): spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$. output_size (int, optional): sequence length of original feature tensor. If None, the original length cached by the previous call of :meth:`unfold` will be used. Returns: :class:`torch.Tensor`: feature tensor of shape $(batch, channels, time)$. .. note:: `fold` caches the original length of the input. Nr!rr") r]r'r-r\rrr(r3rfloat)rr5 output_sizer6r^r\r; to_unfoldrrrr3s zDualPathProcessing.foldcCs^|\}}}}|dd||||dd}||}|||||dddd}|S)a_Performs intra-chunk processing. Args: x (:class:`torch.Tensor`): spliced feature tensor of shape (batch, channels, chunk_size, n_chunks). module (:class:`torch.nn.Module`): module one wish to apply to each chunk of the spliced feature tensor. Returns: :class:`torch.Tensor`: processed spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$. .. note:: the module should have the channel first convention and accept a 3D tensor of shape $(batch, channels, time)$. r!r rr' transposer-r5moduler6r7r\r;rrr intra_processs " z DualPathProcessing.intra_processcCsN|\}}}}|dd||||}||}|||||dd}|S)aiPerforms inter-chunk processing. Args: x (:class:`torch.Tensor`): spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$. module (:class:`torch.nn.Module`): module one wish to apply between each chunk of the spliced feature tensor. Returns: x (:class:`torch.Tensor`): processed spliced feature tensor of shape $(batch, channels, chunksize, nchunks)$. .. note:: the module should have the channel first convention and accept a 3D tensor of shape $(batch, channels, time)$. r!rrbrdrrr inter_process8s z DualPathProcessing.inter_processrC) rIrJrKrLr r)r3 staticmethodrfrgrNrrrrr[s  $ r[) rrlosses.pit_wrapperrModuler FloatTensorintr.r[rrrrs"   -