o i@sbdZddlmZmZmZddlmZddlZGdddejj Z d dd Z Gd d d ejj Z dS) z!Repeat the same layer definition.)DictListOptional) LayerNormNcs.eZdZdZddfdd ZddZZS)MultiSequentialz-Multi-input multi-output torch.nn.Sequential.)layer_drop_ratecstt|j|||_dS)zInitialize MultiSequential with layer_drop. Args: layer_drop_rate (float): Probability of dropping out each fn (layer). N)superr__init__r)selfrargs __class__Z/home/ubuntu/.local/lib/python3.10/site-packages/funasr/models/transformer/utils/repeat.pyr s zMultiSequential.__init__cGsDtt|}t|D]\}}|jr|||jkr||}q |S)zRepeat.)torchemptylenuniform_ enumeratetrainingr)r r _probsidxmrrrforwards zMultiSequential.forward)__name__ __module__ __qualname____doc__r r __classcell__rrr rrs rrcs tfddt|Dd|iS)aRepeat module N times. Args: N (int): Number of repeat time. fn (Callable): Function to generate module. layer_drop_rate (float): Probability of dropping out each fn (layer). Returns: MultiSequential: Repeated model instance. csg|]}|qSrr).0nfnrr 0szrepeat..r)rrange)Nr#rrr"rrepeat$s r'cseZdZdZefdeejjde dejjddffdd Z d e d ej ddfd d Z dd ej dej dej deej dej f ddZ   dd ej dej dej de d e de dej fddZZS) MultiBlockszMultiBlocks definition. Args: block_list: Individual blocks of the encoder architecture. output_size: Architecture output size. norm_class: Normalization module class. norm_args: Normalization module arguments. block_list output_size norm_classreturnNcs0ttj||_|||_t||_dS)zConstruct a MultiBlocks object.N) r r rnn ModuleListblocks norm_blocksr num_blocks)r r)r*r+r rrr <s  zMultiBlocks.__init__ left_contextdevicecCs&t|jD] }|j|||qdS)zInitialize/Reset encoder streaming cache. Args: left_context: Number of left frames during chunk-by-chunk inference. device: Device to use for cache tensor. N)r%r1r/reset_streaming_cache)r r2r3rrrrr4Jsz!MultiBlocks.reset_streaming_cachexpos_encmask chunk_maskcCs8t|jD]\}}|||||d\}}}q||}|S)aRForward each block of the encoder architecture. Args: x: MultiBlocks input sequences. (B, T, D_block_1) pos_enc: Positional embedding sequences. mask: Source mask. (B, T) chunk_mask: Chunk mask. (T_2, T_2) Returns: x: Output sequences. (B, T, D_block_N) )r8)rr/r0)r r5r6r7r8 block_indexblockrrrrSs zMultiBlocks.forwardr chunk_size right_contextc Cs<t|jD]\}}|j||||||d\}}q||}|S)aForward each block of the encoder architecture. Args: x: MultiBlocks input sequences. (B, T, D_block_1) pos_enc: Positional embedding sequences. (B, 2 * (T - 1), D_att) mask: Source mask. (B, T_2) left_context: Number of frames in left context. right_context: Number of frames in right context. Returns: x: MultiBlocks output sequences. (B, T, D_block_N) )r;r2r<)rr/ chunk_forwardr0) r r5r6r7r;r2r< block_idxr:rrrr=js zMultiBlocks.chunk_forward)N)rrr)rrrrrrrr-Moduleintr r3r4Tensorrrr=rrrr rr(3sV   r()r)rtypingrrr$funasr.models.transformer.layer_normrrr- Sequentialrr'r?r(rrrrs