o ॵi/@sddlmZmZmZmZddlZddlmZddl m Z ddl m Z ddl mZmZddlmZddlmZdd lmZmZd gZejejejd ejejejd Gd d d eZdS) )AnyDictOptionalUnionN) Pipelines)Model) OutputKeys)PipelineTensor) PIPELINES) Preprocessor) ModelFileTasksFillMaskPipeline) module_namec seZdZ      ddeeefdeeded effd d Zd e ee fd e ee ffddZ d e ee fd e ee ffddZ ZS)rNgpuTsentencemodel preprocessor config_filedevicec stj||||||dd|didt|jts#Jdtj|dur6tj |jj f||d||_ |j t |j dsEJd dS) ajThe inference pipeline for all the fill mask sub-tasks. Args: model (`str` or `Model` or module instance): A model instance or a model local dir or a model id in the model hub. preprocessor (`Preprocessor`, `optional`): A Preprocessor instance. kwargs (dict, `optional`): Extra kwargs passed into the preprocessor's constructor. Examples: >>> from modelscope.pipelines import pipeline >>> pipeline_ins = pipeline('fill-mask', model='damo/nlp_structbert_fill-mask_english-large') >>> input = 'Everything in [MASK] you call reality is really [MASK] a reflection of your [MASK].' >>> print(pipeline_ins(input)) Examples: >>> from modelscope.pipelines import pipeline >>> pipeline_ins = pipeline('fill-mask', model='damo/nlp_ponet_fill-mask_english-base') >>> input = 'Everything in [MASK] you call reality is really [MASK] a reflection of your [MASK].' >>> print(pipeline_ins(input)) NOTE2: Please pay attention to the model's special tokens. If bert based model(bert, structbert, etc.) is used, the mask token is '[MASK]'. If the xlm-roberta(xlm-roberta, veco, etc.) based model is used, the mask token is ''. To view other examples plese check tests/pipelines/test_fill_mask.py. compileFcompile_options)rrrr auto_collaterrz,please check whether model config exists in N)first_sequencesequence_lengthmask_idz9The input preprocessor should have the mask_id attribute.)super__init__pop isinstancerrr CONFIGURATIONr from_pretrained model_dirrevalhasattr) selfrrrrrrrkwargs __class___/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/pipelines/nlp/fill_mask_pipeline.pyrs6%    zFillMaskPipeline.__init__inputsreturncKs|jdi||S)Nr+)r)r'r-forward_paramsr+r+r,forwardRszFillMaskPipeline.forwardc Cs|tj}|tj}tj|dd}t||j j k||}g}|D]}|j j |ddd}| |q,tj |iS)aprocess the prediction results Args: inputs (Dict[str, Any]): The model outputs. The output should follow some rules: 1. Values can be retrieved by keys(dict-like, or the __getitem__ method is overriden) 2. 'logits' and 'input_ids' key exists. Models in modelscope will return the output dataclass `modelscope.outputs.FillMaskModelOutput`. Returns: Dict[str, str]: the prediction results )axisT)skip_special_tokensclean_up_tokenization_spaces)rLOGITSdetachcpunumpy INPUT_IDSnpargmaxwhererrdecodeappendTEXT) r'r-logits input_idspred_idsrst_ids pred_stringsids pred_stringr+r+r, postprocessVs   zFillMaskPipeline.postprocess)NNrTrr)__name__ __module__ __qualname__rrstrrr rrrr0r rG __classcell__r+r+r)r,rs( ;  *)typingrrrrr8r:modelscope.metainformodelscope.modelsrmodelscope.outputsrmodelscope.pipelines.baser r modelscope.pipelines.builderr modelscope.preprocessorsr modelscope.utils.constantr r__all__register_module fill_maskfill_mask_ponetrr+r+r+r,s