o }oi @sddlmZddlmZmZddlZddlmZddlm Z m Z m Z ddl m Z ddlmZeGdd d ZGd d d Zd ejd edejfddZdd ejdededejfddZd ejdejdejfddZdS)) dataclass)ListOptionalN) DictConfig)BeamSearchSequenceGeneratorGreedySequenceGeneratorTopKSequenceGenerator) TokenizerSpec) NeuralModulec@sVeZdZUdZeed<dZeed<dZeed<dZ e ed<d Z eed <d Z e ed <d S)SequenceGeneratorConfiggreedytypemax_sequence_lengthmax_delta_length? temperature beam_sizelen_penN) __name__ __module__ __qualname__r str__annotations__rintrrfloatrrrr^/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/asr/parts/utils/slu_utils.pyr s     r c seZdZdZdZdZdZeeegZdede de de d e d d f fd d Z  dde j de j dedeedef ddZde j d e j fddZde j d eefddZZS)SequenceGeneratorzF Wrapper class for sequence generators for NeMo transformers. r topkbeamcfg embeddingdecoder log_softmax tokenizerreturnNc s<t|dd|_||_t|dd|_t|dd|_t|dd|_|j|j|j|dd |d d|d d d }|j|j krOt |||fi||_ dS|j|j krp|dd }|dd}t |||||fi||_ dS|j|jkr|dd }|dd} t||||| fi||_ dStd|jd|jd)Nr r pad_idreos_idrbos_idrrr batch_sizer)padboseosrrr-rrrrrz(Sequence Generator only supports one of z , but got z instead.)super__init__get_typer(getattrr*r+r, TYPE_GREEDYr generator TYPE_TOPKr TYPE_BEAMr ValueError SEARCH_TYPES) selfr$r%r&r'r( common_argsrrr __class__rr r24sD              zSequenceGenerator.__init__Fencoder_statesencoder_input_maskreturn_beam_scores pad_max_len return_lengthcCs8|j|||d}|rt|||j}|r|||fS|S)a  Generate sequence tokens given the input encoder states and masks. Params: - encoder_states: a torch Tensor of shape BxTxD - encoder_input_mask: a binary tensor of shape BxTxD - return_beam_scores: whether to return beam scores - pad_max_len: optional int, set it to pad all sequence to the same length - return_length: whether to return the lengths for generated sequences (shape B) Returns: - generated tokens tensor of shape BxT )encoder_hidden_statesrArB)r7 pad_sequencer*get_seq_length)r<r@rArBrCrD predictionsrrr __call__^szSequenceGenerator.__call__seqcCsd|dtj|d|jd}||jk}tj|d|dddf|dddfd}|S)z Get sequence length. Params: - seq: batched sequence tensor of shape BxTxD Returns: - tensor of shape B, where each element is the length of the sequence rrdeviceNdimindexsrc)sizetorchonesrLlongr+nonzeroscatter)r<rJlengthspos seq_lengthsrrr rGs$*z SequenceGenerator.get_seq_length seq_tokenscCsg}|}||}t|jdD])}||}||}|d|}d |j j |}| |q|S)z Decode tokens into strings Rarams: - seq_tokens: integer tensor of shape BxT Returns: - list of strings rN) detachrTcpurGrangeshapenumpytolistitemjoinr( decode_idsappend)r<rZsemantics_listrYindtokenslengthtextrrr decode_semantics_from_tokenss   z.SequenceGenerator.decode_semantics_from_tokens)NFNF)rrr__doc__r6r8r9SEARCHER_TYPESrr r r2rRTensorboolrrrIrGrrrk __classcell__rrr>r r!*sH - ! r!rJr+r)cCsb|dtj|d|jd}||k}tj|d|dddf|dddfd}|S)z Get sequence length. Params: - seq: batched sequence tensor of shape BxTxD - eos_id: integer representing the end of sentence Returns: - tensor of shape B, where each element is the length of the sequence rrrKNrM)rQrRrSrLrTrUrV)rJr+rWrXrYrrr rGs$ *rGmax_len pad_tokencCsR|d}|d}||kr|Stj||||j|jd|}tj||gddS)z Params: - seq: integer token sequences of shape BxT - max_len: integer for max sequence length - pad_token: integer token for padding Returns: - padded sequence of shape B x max_len rr)dtyperL)rN)rQrRzerosrsrLfill_cat)rJrqrrbatchcurr_lenpaddingrrr rFs  rFseq_lenscCsDt|ddddf|j|dddfk}|j|jtdS)z Get the sequence mask based on the actual length of each sequence Params: - seq: tensor of shape [BxLxD] - seq_len: tensor of shape [B] Returns: - binary mask of shape [BxL] rN)rs)rRarangerQtorLro)rJrzmaskrrr get_seq_masks4 r~)r) dataclassesrtypingrrrR omegaconfr(nemo.collections.asr.modules.transformerrrr1nemo.collections.common.tokenizers.tokenizer_specr nemo.core.classes.moduler r r!rnrrGrFr~rrrr s     x