o ߥi#@sddlZddlmZddlmZmZddlmZm Z m Z m Z m Z ddl mZddlmZddlmZmZddlmZmZddlmZmZdd lmZmZmZdd lmZdd l m!Z!dd l"m#Z#m$Z$e!Z%e d Z&GdddeZ'dS)N)ABCabstractmethod)AnyDictListOptionalUnion)snapshot_download)Tasks)build_backbone build_model)can_load_by_mstry_to_load_hf_model)Config ConfigDict)DEFAULT_MODEL_REVISIONInvoke ModelFile) verify_device) get_logger)register_modelhub_reporegister_plugins_repo)z torch.Tensorz tf.Tensorc @seZdZdZddZdeeeffddZe deeeffddZ d eeefdeeeffd d Z e d d Z e eddfdedeededefddZ  ddeeejfdeeeefdeefddZdS)ModelzBase model interface. cOs$||_|dd}t|||_dS)Ndevicegpu) model_dirgetr _device_name)selfrargskwargs device_namer"U/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/base/base_model.py__init__s  zModel.__init__returncOs||j|i|S)N) postprocessforwardrrr r"r"r#__call__"szModel.__call__cOsdS)z Run the forward pass for a model. Returns: Dict[str, Any]: output from the model forward pass Nr"r(r"r"r#r'%sz Model.forwardinputscKs|S)a$ Model specific postprocess and convert model output to standard model outputs. Args: inputs: input data Return: dict of results: a dict containing outputs of model, each output should have the standard output name. r")rr*r r"r"r#r&/s zModel.postprocesscKs|di|S)a Define the instantiation method of a model,default method is by calling the constructor. Note that in the case of no loading model process in constructor of a task model, a load_model method is added, and thus this method is overloaded Nr"r")clsr r"r"r# _instantiate<szModel._instantiateNmodel_name_or_pathrevisioncfg_dictrcKs|d}|dur|d|tj}|dur|tjntj}|dd}t|r0|} n|dur8tddtj|f}t||||d} t d| t | t j } d} |dur_|} n tj| rjt| } t| d d} d |vry|d } t| d t} t| d rt| d s| j| _t| d d}t|tr|d rd|dd}|dd}|durt| | |rd}d}|dvrt| | |fi|}|dur|dd}|dur|dur||}|S| durtdt j d| | _t| dt!| | dd|"D] \}}|| |<q|dur|| _#| t$j%ur%d| _&t'| }nt(| | d}t| dr5| j)|_)t|ds>| |_*| dd||_+| |_|S)asInstantiate a model from local directory or remote model repo. Note that when loading from remote, the model revision can be specified. Args: model_name_or_path(str): A model dir or a model id to be loaded revision(str, `optional`): The revision used when the model_name_or_path is a model id of the remote hub. default `master`. cfg_dict(Config, `optional`): An optional model config. If provided, it will replace the config read out of the `model_name_or_path` device(str, `optional`): The device to load the model. **kwargs: task(str, `optional`): The `Tasks` enumeration value to replace the task value read out of config in the `model_name_or_path`. This is useful when the model to be loaded is not equal to the model saved. For example, load a `backbone` into a `text-classification` model. Other kwargs will be directly fed into the `model` key, to replace the default configs. use_hf(bool, `optional`): If set to True, it will initialize the model using AutoModel or AutoModelFor* from hf. If set to False, the model is loaded using the modelscope mode. If set to None, the loading mode will be automatically selected. ignore_file_pattern(List[str], `optional`): This parameter is passed to snapshot_download device_map(str | Dict[str, str], `optional`): This parameter is passed to AutoModel or AutoModelFor* torch_dtype(torch.dtype, `optional`): This parameter is passed to AutoModel or AutoModelFor* config(PretrainedConfig, `optional`): This parameter is passed to AutoModel or AutoModelFor* Returns: A model instance. Examples: >>> from modelscope.models import Model >>> Model.from_pretrained('damo/nlp_structbert_backbone_base_std', task='text-classification') model_prefetchedNignore_file_patternTz9Expecting model is pre-fetched locally, but is not found.z%s/%s) user_agentr1zinitialize model from taskmodel model_typetypercudause_hfF>NT device_map`z` file not found.plugins allow_remote) task_namepipelinecfgr),rpoprKEY PRETRAINEDospexists RuntimeErrorr loggerinfojoinr CONFIGURATIONospathr from_filegetattrrhasattrr5r6 isinstancestr startswithr rtoFileNotFoundErrorrrsafe_getritemsrr backbone init_backboner r r?r@name)r+r-r.r/rr prefetched invoked_byr1local_model_dirconfiguration_pathr@r> model_cfgr5r9r4r:kvr"r"r#from_pretrainedEs *                      zModel.from_pretrained target_foldersave_checkpoint_namesconfigcKstd)a save the pretrained model, its configuration and other related files to a directory, so that it can be re-loaded Args: target_folder (Union[str, os.PathLike]): Directory to which to save. Will be created if it doesn't exist. save_checkpoint_names (Union[str, List[str]]): The checkpoint names to be saved in the target_folder config (Optional[dict], optional): The config for the configuration.json, might not be identical with model.config z>save_pretrained method need to be implemented by the subclass.)NotImplementedError)rrbrcrdr r"r"r#save_pretrainedszModel.save_pretrained)NN)__name__ __module__ __qualname____doc__r$rrQrr)rr'r& classmethodr,rrrrarrKPathLikerdictrfr"r"r"r#rs@"    r)(rKos.pathrLrDabcrrtypingrrrrr modelscope.hub.snapshot_downloadr modelscope.metainfor modelscope.models.builderr r modelscope.utils.automodel_utilsr rmodelscope.utils.configrrmodelscope.utils.constantrrrmodelscope.utils.devicermodelscope.utils.loggerrmodelscope.utils.pluginsrrrGTensorrr"r"r"r#s