U %‰d ã@s:ddlZdgZGdd„dejjƒZGdd„dejjƒZdS)éNÚ DeepSpeechcsDeZdZdZd eeeeddœ‡fdd„ Zejejdœdd „Z ‡Z S) ÚFullyConnectedzh Args: n_feature: Number of input features n_hidden: Internal hidden unit size. éN)Ú n_featureÚn_hiddenÚdropoutÚ relu_max_clipÚreturncs2tt|ƒ ¡tjj||dd|_||_||_dS)NT)Zbias) ÚsuperrÚ__init__ÚtorchÚnnÚLinearÚfcrr)Úselfrrrr©Ú __class__©ú@/tmp/pip-unpacked-wheel-lbdmvq91/torchaudio/models/deepspeech.pyr szFullyConnected.__init__©Úxr cCsL| |¡}tjj |¡}tjj |d|j¡}|jrHtjj ||j|j¡}|S)Nr) rr r Ú functionalÚreluZhardtanhrrZtraining)rrrrrÚforwards  zFullyConnected.forward)r© Ú__name__Ú __module__Ú __qualname__Ú__doc__ÚintÚfloatr r ZTensorrÚ __classcell__rrrrrsrcsDeZdZdZd eeeeddœ‡fdd„ Zejejd œd d „Z ‡Z S) ra DeepSpeech model architecture from *Deep Speech: Scaling up end-to-end speech recognition* [:footcite:`hannun2014deep`]. Args: n_feature: Number of input features n_hidden: Internal hidden unit size. n_class: Number of output classes éé(çN)rrÚn_classrr csxtt|ƒ ¡||_t|||ƒ|_t|||ƒ|_t|||ƒ|_tj j ||dddd|_ t|||ƒ|_ tj   ||¡|_dS)NérT)Z num_layersZ nonlinearityÚ bidirectional)r rr rrÚfc1Úfc2Úfc3r r ZRNNÚbi_rnnÚfc4rÚout)rrrr%rrrrr 'szDeepSpeech.__init__rcCs²| |¡}| |¡}| |¡}| d¡}| dd¡}| |¡\}}|dd…dd…d|j…f|dd…dd…|jd…f}| |¡}| |¡}|  ddd¡}t j j j |dd}|S)zà Args: x (torch.Tensor): Tensor of dimension (batch, channel, time, feature). Returns: Tensor: Predictor tensor of dimension (batch, time, class). r&rNé)Zdim)r(r)r*ZsqueezeZ transposer+rr,r-Zpermuter r rZ log_softmax)rrÚ_rrrr7s     8  zDeepSpeech.forward)r"r#r$rrrrrrs ûú)r Ú__all__r ÚModulerrrrrrÚs