o }oiO@sddlmZmZmZddlZddlmZddlmZddl m Z ddl m Z m Z mZmZddlmZdd Zd d Zd d ZGddde ZGdddeZGddde ZGddde ZdS))DictListOptionalN)ExternalFeatureLoader) collections)Dataset)AcousticEncodedRepresentation LabelsType LengthsType NeuralType)loggingc Cstt|}t|dkr|\}}}}}nt|dkr#d}|\}}}}ntdgg}}|D]}|d|d} } || || q.t|}t|}t|}t|}|durb||||fStj|tjd}|||||fS)acollate batch of feat sig, feat len, labels, labels len, assuming all features have the same shape. Args: batch (FloatTensor, LongTensor, LongTensor, LongTensor): A tuple of tuples of feature, feature lengths, encoded labels, and encoded labels length. N$Expects 4 or 5 tensors in the batch!rdtype) listziplen ValueErrorappendtorchstacktensorint32) batch packed_batch_ feat_lengthslabels_lengths sample_idsfeatureslabelsbfeat_ilabels_ir'^/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/asr/data/feature_to_label.py_feature_collate_fns(           r)cCstt|}t|dkr|\}}}}}nt|dkr#d}|\}}}}ntdd}|ddu} | r7t|}t|} gg} } |D]P} | d| d| d| df\}}}}| rw|}||krrd||f}tjjj |||d }| ||}|| krd| |f}tjjj |||d }| |qD| rt | } t |}nd \} }t | } t |}|dur| || |fStj |tj d }| || ||fS) a}collate batch of audio feature, audio len, labels, labels len Args: batch (Optional[FloatTensor], Optional[LongTensor], LongTensor, LongTensor): A tuple of tuples of feature, feature lengths, labels, and label lengths. This collate func assumes the features are torch tensors of Log-Melspectrogram (i.e. [N_MEL, T]). r rNrrr)valueNNr)rrrrmaxitemrnn functionalpadrrrr)r feat_pad_val label_pad_idrrrr r! max_feat_lenhas_featmax_labels_lenr"r#r$r% feat_i_lenlabel_i label_i_lenr2feature_lengthsr'r'r(_audio_feature_collate_fn;sH       $         r<cCst||}t|\}}}}tt|t|}t||} |ddu} |djd} ggggf\}} } }tj|ddd}|tj|ddd}|D]b\}}}}t| |}t| |}tj|||fdd}||7}| rtdtj||| dd}t |D]}|| }||}|dd||f}| |q{| || |g|| |g|qH| rt |}t |}nd\}}t | } t | }||| |fS) acollate batch of audio features, features len, tokens, tokens len Args: batch (Optional[FloatTensor], Optional[LongTensor], LongTensor, LongTensor): A tuple of tuples of signal, signal lengths, encoded tokens, and encoded tokens length. This collate func assumes the signals are 1d torch tensors (i.e. mono audio). batch size equals to 1. rNrtrunc) rounding_moder*)dimr-)intrminr.shaperdivzeroscatrangerextendrr)rwindow_length_in_secshift_length_in_secframe_unit_in_sec slice_lengthaudio_featuresrrtokens_lengthsshift has_audiof_dim num_slicestokensappend_len_startappend_len_endr%r8tokens_istartendslicesslice_id start_idxend_idx feat_slicer'r'r(_vad_feature_segment_collate_fnqs@             r]csfeZdZdZedeeeeffddZ dddede ed e ffd d Z d d Z ddZZS)_FeatureSeqSpeakerLabelDataseta. Dataset that loads tensors via a json file containing paths to feature files, sequences of labels. Each new line is a different sample. Example below: and their target labels. JSON files should be of the following format: {"feature_filepath": "/path/to/feature_0.p", "seq_label": speakerA speakerB SpeakerA ....} ... {"feature_filepath": "/path/to/feature_n.p", "seq_label": target_seq_label_n} target_seq_label_n is the string of sequence of speaker label, separated by space. Args: manifest_filepath (str): Dataset parameter. Path to JSON containing data. labels (Optional[list]): Dataset parameter. List of unique labels collected from all samples. feature_loader : Dataset parameter. Feature loader to load (external) feature. returncCstdtttdtd}|jr2|tdtttdttdtttdtd|S|tdtttdtd|S)4Returns definitions of module output ports. BDTrb) external_feat feat_lengthrbrd)embs embs_lengthlabel label_length)rjrk)r rtupler is_speaker_embupdater self output_typesr'r'r(rqs    z+_FeatureSeqSpeakerLabelDataset.output_typesF)rmmanifest_filepathr#rmcsttj|dd|_||_|r|n|jj|_||_ ii|_ |_ t |jD]\}}||j |<||j |<q*t t|jddD]}td||j |qDdS)N,)manifests_filesr $ label id {} and its mapped label {})super__init__rASRFeatureSequenceLabelsplit collectionfeature_loader uniq_labelsr#rmlabel2idid2label enumeraterFrr debugformat)rprrr#r{rmlabel_idrjidx __class__r'r(rws   z'_FeatureSeqSpeakerLabelDataset.__init__cC t|jSNrrzrpr'r'r(__len__ z&_FeatureSeqSpeakerLabelDataset.__len__cCsb|j|}|j|j}|t|jd}}t|j }tt |j}||||fSNr) rzr{process feature_filerrrBlong seq_labelfloatrrpindexsampler"fflttlr'r'r( __getitem__s  z*_FeatureSeqSpeakerLabelDataset.__getitem__)__name__ __module__ __qualname____doc__propertyrrstrr rqrboolrwrr __classcell__r'r'rr(r^sr^c@seZdZdZddZdS)FeatureToSeqSpeakerLabelDataseta# Dataset that loads tensors via a json file containing paths to feature files and sequence of speakers. Each new line is a different sample. Example below: {"feature_filepath": "/path/to/feature_0.p", "seq_label": speakerA speakerB SpeakerA ....} ... {"feature_filepath": "/path/to/feature_n.p", "seq_label": target_seq_label_n} target_seq_label_n is the string of sequence of speaker label, separated by space. Args: manifest_filepath (str): Path to manifest json as described above. Canbe comma-separated paths. labels (Optional[list]): String containing all the possible labels to map to if None then automatically picks from ASRFeatureSequenceLabel collection. feature_loader, Feature load to loader (external) feature. cCst|Sr)r)rprr'r'r( _collate_fnsz+FeatureToSeqSpeakerLabelDataset._collate_fnN)rrrrrr'r'r'r(rs rcseZdZdZdZdZedeee e ffddZ ddddd d dddd d e d e e d dde de dedeede dee dee ffddZddZddZddZddZZS) FeatureToLabelDataseta Dataset that loads tensors via a json file containing paths to feature files and their labels. Each new line is a different sample. Example below: and their target labels. JSON files should be of the following format: {"feature_filepath": "/path/to/audio_feature.pt", "label": "1"} ... {"feature_filepath": "/path/to/audio_feature.pt", "label": "0"} Args: manifest_filepath (str): Path to JSON containing data. labels (Optional[list]): List of unique labels collected from all samples. augmentor (Optional): feature augmentation window_length_in_sec (float): Window length in seconds. shift_length_in_sec (float): Shift length in seconds. is_regression_task (bool): if True, the labels are treated as for a regression task. cal_labels_occurrence (bool): if True, the labels occurrence will be calculated. zero_spec_db_val (float): Value to replace non-speech signals in log-melspectrogram. min_duration (float): Minimum duration of the audio file in seconds. max_duration (float): Maximum duration of the audio file in seconds. (\0g{Gz?r_cCs:tdtttdttdtttdtd}|S)r`rarb audio_featrfr# labels_lengthr rrlr r ror'r'r(rq  z"FeatureToLabelDataset.output_typesNg)\(?F) r# augmentorrHrIis_regression_taskcal_labels_occurrencezero_spec_db_val min_duration max_durationrrr#r1nemo.collections.asr.parts.perturb.AudioAugmentorrHrIrrrrrc sZt|_|_|_t|tr|d}tj |||| | d_ t |d_ |r-|nj j _|_|s|r;|nj j _jdurJtjnd_ii__ig__tjD]\} } | j| <| j| <|ryj j| j| <q`|rfddtjD_ttjddD]} td| j| qdSg_d_dS) Nrs)rtrrrrrr*cg|]}j|qSr') id2occurrence).0krr'r( Yz2FeatureToLabelDataset.__init__..r ru)rvrwrHrIr isinstancerryrASRFeatureLabelrzrr{r|r#rr num_classesr}r~rlabels_occurrencersortedrFr rr)rprrr#rrHrIrrrrrrrjrrrr(rw)sD       zFeatureToLabelDataset.__init__cCrrrrr'r'r(rarzFeatureToLabelDataset.__len__cCs^|j|}|j|j}|t|jd}}t|j|j }td}||||fS)Nr*) rzr{rrrrrBrr}rjrr'r'r(rds  z!FeatureToLabelDataset.__getitem__cCt||jdSrr<rrr'r'r(roz!FeatureToLabelDataset._collate_fncCst||j|j|jSr)r]rHrIFRAME_UNIT_TIME_SECSrr'r'r(_vad_segment_collate_fnrsz-FeatureToLabelDataset._vad_segment_collate_fn)rrrrZERO_LEVEL_SPEC_DB_VALrrrrrr rqrrrrwrrrrrr'r'rr(rsP    8 rcseZdZdZdZedeeee ffddZ dddddddddd ed e ed d d eede dee de dee dee ffddZddZdefddZddZddZddZZS) FeatureToMultiLabelDatasetaK Dataset that loads tensors via a json file containing paths to feature files and their labels. Each new line is a different sample. Example below: and their target labels. JSON files should be of the following format: {"feature_filepath": "/path/to/audio_feature.pt", "label": "1 1 0 0 1"} ... {"feature_filepath": "/path/to/audio_feature.pt", "label": "0 1 0 0"} Args: manifest_filepath (str): Path to JSON containing data. labels (Optional[list]): List of unique labels collected from all samples. augmentor (Optional): feature augmentation delimiter (str): delimiter to split the labels. is_regression_task (bool): if True, the labels are treated as for a regression task. cal_labels_occurrence (bool): if True, the labels occurrence will be calculated. zero_spec_db_val (float): Value to replace non-speech signals in log-melspectrogram. min_duration (float): Minimum duration of the audio file in seconds. max_duration (float): Maximum duration of the audio file in seconds. rr_cCs:tdtttdttdtttdtd}|S)r`rarbrgrrror'r'r(rqrz'FeatureToMultiLabelDataset.output_typesNF)r#r delimiterrrrrrrrr#rrrrrrrrc sJt||_||_t|tr|d}tj|||||| d|_ ||_ t |d|_ |r.|n|j j |_ii|_|_|s|r@|n||_|jdurOt|jnd|_ii|_|_t|jD]"\} } | |j| <| |j| <|r|j j| |j| <|j|j| q^tt|jddD]} td| |j| qdSg|_d|_dS)Nrs)rtrrrrrrr*r ru)rvrwrrrrryrrrzrrr{r|r#r}r~_get_label_setrrrrrrrFr rr) rprrr#rrrrrrrrrjrrr'r(rwsB       z#FeatureToMultiLabelDataset.__init__cCsJg}|jD]}|j}|r|jr||jn|}||qtt|Sr)rzrjrryrGrset)rpr#r label_strlabel_str_listr'r'r(rs   z)FeatureToMultiLabelDataset._get_label_setrcsdjr |jn|}jr dd|D}t|}|Sfdd|D}t|}|S)NcSsg|]}t|qSr')rrsr'r'r(rszCFeatureToMultiLabelDataset._label_str_to_tensor..crr')r}rrr'r(rr)rryrrrrr)rprr#r'rr(_label_str_to_tensorsz/FeatureToMultiLabelDataset._label_str_to_tensorcCrrrrr'r'r(rrz"FeatureToMultiLabelDataset.__len__cCs^|j|}|j|j}|t|jd}}||j }t| d}||||fS)Nr*r) rzr{rrrrrBrrrjsizerr'r'r(rs   z&FeatureToMultiLabelDataset.__getitem__cCrrrrr'r'r(rrz&FeatureToMultiLabelDataset._collate_fn)rrrrrrrrrr rqrrrrwrrrrrrr'r'rr(rxsJ   3   r)typingrrrr7nemo.collections.asr.parts.preprocessing.feature_loaderr+nemo.collections.common.parts.preprocessingrnemo.core.classesrnemo.core.neural_typesrr r r nemo.utilsr r)r<r]r^rrrr'r'r'r(s     "62Kt