i ddlZdgZGddejjZGddejjZy)N DeepSpeechc reZdZdZd dededededdf fd Zd ejdejfd Z xZ S) FullyConnectedzh Args: n_feature: Number of input features n_hidden: Internal hidden unit size. n_featuren_hiddendropout relu_max_clipreturnNctt| tjj ||d|_||_||_y)NT)bias) superr__init__torchnnLinearfcr r)selfrrrr __class__s q/Volumes/fast/ai/experiments/voice-extract-mac/.venv/lib/python3.12/site-packages/torchaudio/models/deepspeech.pyrzFullyConnected.__init__ s9 nd,.((//)XD/A* xcz|j|}tjjj |}tjjj |d|j }|jr?tjjj||j|j}|S)Nr) rrr functionalreluhardtanhr rtraining)rrs rforwardzFullyConnected.forwards GGAJ HH   $ $Q ' HH   ( (At/A/A B <<##++At||T]]KAr) __name__ __module__ __qualname____doc__intfloatrrTensorr __classcell__rs@rrrsM #uUXbf %,,rrc xeZdZdZ d dededededdf fd Zd ejdejfd Z xZ S) ra DeepSpeech architecture introduced in *Deep Speech: Scaling up end-to-end speech recognition* :cite:`hannun2014deep`. Args: n_feature: Number of input features n_hidden: Internal hidden unit size. n_class: Number of output classes rrn_classrr Ncdtt| ||_t ||||_t ||||_t ||||_tjj||ddd|_ t ||||_ tjj|||_y)NrT) num_layers nonlinearity bidirectional)r rrrrfc1fc2fc3rrRNNbi_rnnfc4rout)rrrr*rrs rrzDeepSpeech.__init__&s j$(*  !)Xw?!(Hg>!(Hg>hhll8X!RXhllm !(Hg>88??8W5rrc|j|}|j|}|j|}|jd}|j dd}|j |\}}|ddddd|j f|dddd|j dfz}|j|}|j|}|jddd}tjjj|d}|S)z Args: x (torch.Tensor): Tensor of dimension (batch, channel, time, feature). Returns: Tensor: Predictor tensor of dimension (batch, time, class). r,rN)dim)r0r1r2squeeze transposer4rr5r6permuterrr log_softmax)rr_s rrzDeepSpeech.forward6s HHQK HHQK HHQK IIaL KK1 {{1~1 aOdmmO# $qAt}})>'? ? HHQK HHQK IIaA  HH   + +A1 + 5r)i(grr(s@rrrsc 666 6  6  6 %,,r)r__all__rModulerrrrrCs7 .UXX__,88r