o i @s:ddlZdgZGdddejjZGdddejjZdS)N DeepSpeechc sNeZdZdZddededededdf fd d Zd ejdejfd d Z Z S)FullyConnectedzh Args: n_feature: Number of input features n_hidden: Internal hidden unit size.  n_featuren_hiddendropout relu_max_clipreturnNcs2tt|tjj||dd|_||_||_dS)NT)bias) superr__init__torchnnLinearfcrr)selfrrrr __class__Z/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/torchaudio/models/deepspeech.pyr s zFullyConnected.__init__xcCsL||}tjj|}tjj|d|j}|jr$tjj||j|j}|S)Nr) rr r functionalreluhardtanhrrtraining)rrrrrforwards zFullyConnected.forward)r __name__ __module__ __qualname____doc__intfloatr r Tensorr __classcell__rrrrrs$rc sTeZdZdZ   ddedededed d f fd d Zd ejd ejfddZ Z 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 Ncsxtt|||_t||||_t||||_t||||_tj j ||dddd|_ t||||_ tj |||_dS)NrT) num_layers nonlinearity bidirectional)r rr rrfc1fc2fc3r rRNNbi_rnnfc4rout)rrrr(rrrrr &szDeepSpeech.__init__rcCs||}||}||}|d}|dd}||\}}|ddddd|jf|dddd|jdf}||}||}| ddd}t j j j |dd}|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)r-r.r/squeeze transposer1rr2r3permuter rr log_softmax)rr_rrrr6s     8  zDeepSpeech.forward)r%r&r'rrrrrrs" )r __all__rModulerrrrrrs