h SSKrS/r"SS\RR5r"SS\RR5rg)N DeepSpeechc ^\rSrSrSrS S\S\S\S\SS4 U4S jjjrS \RS\R4S jr S r U=r $)FullyConnectedzX Args: n_feature: Number of input features n_hidden: Internal hidden unit size. n_featuren_hiddendropout relu_max_clipreturnNc>[[U] 5 [RR XSS9UlX@lX0lg)NT)bias) superr__init__torchnnLinearfcr r )selfrrr r __class__s T/var/www/auris/envauris/lib/python3.13/site-packages/torchaudio/models/deepspeech.pyrFullyConnected.__init__ s3 nd,.((//)D/A* xcURU5n[RRR U5n[RRR USUR 5nUR(a>[RRRXRUR5nU$)Nr) rrr functionalreluhardtanhr r training)rrs rforwardFullyConnected.forwards} GGAJ HH   $ $Q ' HH   ( (At/A/A B <<##++A||T]]KAr)r rr ) __name__ __module__ __qualname____firstlineno____doc__intfloatrrTensorr__static_attributes__ __classcell__rs@rrrsW #uUXbf %,,rrc ^\rSrSrSrS S\S\S\S\SS4 U4S jjjrS \RS\R4S jr S r U=r $)rzDeepSpeech architecture introduced in *Deep Speech: Scaling up end-to-end speech recognition* :cite:`hannun2014deep`. Args: n_feature: Number of input features n_hidden: Internal hidden unit size. n_class: Number of output classes rrn_classr r NcP>[[U] 5 X l[ XU5Ul[ X"U5Ul[ X"U5Ul[RRX"SSSS9Ul [ X"U5Ul [RRX#5Ulg)NrT) num_layers nonlinearity bidirectional)rrrrrfc1fc2fc3rrRNNbi_rnnfc4rout)rrrr0r rs rrDeepSpeech.__init__&s j$(* !)w?!(g>!(g>hhll8!RXhllm !(g>88??85rrcURU5nURU5nURU5nURS5nUR SS5nUR U5upUSS2SS2SUR 24USS2SS2UR S24-nURU5nURU5nURSSS5n[RRRUSS9nU$)z Args: x (torch.Tensor): Tensor of dimension (batch, channel, time, feature). Returns: Tensor: Predictor tensor of dimension (batch, time, class). r2rN)dim)r6r7r8squeeze transposer:rr;r<permuterrr log_softmax)rr_s rrDeepSpeech.forward6s HHQK HHQK HHQK IIaL KK1 {{1~ aOdmmO# $qAt}})>'? ? HHQK HHQK IIaA  HH   + +A1 + 5r)r:r6r7r8r;rr<)i(gr"r-s@rrrsm 666 6  6  66 %,,r)r__all__rModulerrrrrKs7 .UXX__,88r