o i @s4dZddlmZmZddlZGdddejjZdS)z+VGG2L module definition for custom encoder.)TupleUnionNc seZdZdZddededejjffdd Zdej d ej d e e ej ej fe e ej ej fej fffd d Z d ej d ej fd dZ ZS)VGG2LzVGG2L module for custom encoder. Args: idim: Input dimension. odim: Output dimension. pos_enc: Positional encoding class. Nidimodimpos_enccsttjtjjddddddtjtjjddddddtjtjdtjjddddddtjtjjddddddtjtjd |_|durftjtj d|d d |||_ dStj d|d d ||_ dS) zConstruct a VGG2L object.@)stridepadding)r )r r Nr ) super__init__torchnn SequentialConv2dReLU MaxPool2dvgg2lLinearoutput)selfrrr __class__`/home/ubuntu/.local/lib/python3.10/site-packages/espnet/nets/pytorch_backend/transducer/vgg2l.pyrs$     zVGG2L.__init__feats feats_maskreturnc Csn|d}||}|\}}}}||dd||||}|dur1||}||fS|}||fS)azForward VGG2L bottleneck. Args: feats: Feature sequences. (B, F, D_feats) feats_mask: Mask of feature sequences. (B, 1, F) Returns: vgg_output: VGG output sequences. (B, sub(F), D_out) or ((B, sub(F), D_out), (B, sub(F), D_att)) vgg_mask: Mask of VGG output sequences. (B, 1, sub(F)) rr N) unsqueezersizer transpose contiguousviewcreate_new_mask) rrr vgg_outputbctfvgg_maskrrrforward*s   z VGG2L.forwardcCs|d|dd}|ddddd|fdddddddf}|d|dd}|ddddd|fdddddddf}|S)zCreate a subsampled mask of feature sequences. Args: feats_mask: Mask of feature sequences. (B, 1, F) Returns: vgg_mask: Mask of VGG2L output sequences. (B, 1, sub(F)) r r N)r#)rr vgg1_t_lenr- vgg2_t_lenrrrr'Ls  22zVGG2L.create_new_mask)N)__name__ __module__ __qualname____doc__intrrModulerTensorrrr.r' __classcell__rrrrrs  "r)r4typingrrrrr6rrrrrs