o pi@stddlmZddlmZddlZddlmZddlmZm Z ddl m Z ddl m Z ddlmZGdd d e ZdS) ) lru_cache)OptionalN) rearrangereduce)MFCC)Model)TaskcseZdZ   ddededeeffdd Zed ed efd d Zdd ed efddZ dded efddZ e d efddZ de jd e jfddZZS)SimpleEmbeddingModel>N sample_rate num_channelstaskcsPtj|||dt|jjddddd|_tj|jj|jj ddd d d |_ dS) N)r r r(orthoF)r n_mfccdct_typenormlog_mels r T) num_layers batch_first bidirectional) super__init__rhparamsr mfccnnLSTMrr lstm)selfr r r __class__Y/home/ubuntu/.local/lib/python3.10/site-packages/pyannote/audio/models/embedding/debug.pyr%s zSimpleEmbeddingModel.__init__ num_samplesreturncCsD|jjjj}|jjjj}|jjjj}|rd||Sd|||S)a|Compute number of output frames for a given number of input samples Parameters ---------- num_samples : int Number of input samples Returns ------- num_frames : int Number of output frames Source ------ https://pytorch.org/docs/stable/generated/torch.stft.html#torch.stft r rMelSpectrogram spectrogram hop_lengthn_fftcenter)r!r&r+r,r-r$r$r% num_frames=s    zSimpleEmbeddingModel.num_framesr.cCs(|jjjj}|jjjj}||d|S)a Compute size of receptive field Parameters ---------- num_frames : int, optional Number of frames in the output signal Returns ------- receptive_field_size : int Receptive field size. r )rr)r*r+r,)r!r.r+r,r$r$r%receptive_field_sizeZs  z)SimpleEmbeddingModel.receptive_field_sizerframecCs@|jjjj}|jjjj}|jjjj}|r||S|||dS)zCompute center of receptive field Parameters ---------- frame : int, optional Frame index Returns ------- receptive_field_center : int Index of receptive field center. rr()r!r0r+r,r-r$r$r%receptive_field_centerls   z+SimpleEmbeddingModel.receptive_field_centercCsdS)zDimension of output@r$)r!r$r$r% dimensionszSimpleEmbeddingModel.dimension waveformscCs*||}|t|d\}}t|ddS)z Parameters ---------- waveforms : (batch, time, channel) Returns ------- embedding : (batch, dimension) zb c f t -> b t (c f)z b t f -> b fmean)rr rr)r!r4routputhiddenr$r$r%forwards  zSimpleEmbeddingModel.forward)r r N)r )r)__name__ __module__ __qualname__intrrrrr.r/r1propertyr3torchTensorr8 __classcell__r$r$r"r%r $s$r ) functoolsrtypingrr>torch.nnreinopsrrtorchaudio.transformsrpyannote.audio.core.modelrpyannote.audio.core.taskrr r$r$r$r%s