U -e @s:ddlZdgZGdddejjZGdddejjZdS)N DeepSpeechcsDeZdZdZd eeeeddfdd Zejejddd Z Z S) FullyConnectedzh Args: n_feature: Number of input features n_hidden: Internal hidden unit size. N) n_featuren_hiddendropout relu_max_clipreturncs2tt|tjj||dd|_||_||_dS)NT)Zbias) superr__init__torchnnLinearfcrr)selfrrrr __class__]/var/www/html/Darija-Ai-Train/env/lib/python3.8/site-packages/torchaudio/models/deepspeech.pyr szFullyConnected.__init__xr cCsL||}tjj|}tjj|d|j}|jrHtjj||j|j}|S)Nr) rr r functionalreluZhardtanhrrZtraining)rrrrrforwards  zFullyConnected.forward)r __name__ __module__ __qualname____doc__intfloatr r ZTensorr __classcell__rrrrrsrcsDeZdZdZd eeeeddfdd Zejejd d d Z 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 (N)rrn_classrr csxtt|||_t||||_t||||_t||||_tj j ||dddd|_ t||||_ tj |||_dS)NrT)Z num_layersZ nonlinearity bidirectional)r rr rrfc1fc2fc3r r ZRNNbi_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*ZsqueezeZ transposer+rr,r-Zpermuter r rZ log_softmax)rr_rrrr6s     8  zDeepSpeech.forward)r"r#r$rrrrrrs )r __all__r Modulerrrrrrs