o }oi0 @sddlZddlZddlmZddlmZmZmZddlm Z dZ z ddl m Z m Z mZWn ey7dZ Ynw ddd d d d efd dZdd d d ded efddZdd d d defddZ  ddddddeeedeeeeffddZdS)N) is_dataclass)DictListOptional)loggingT) DictConfig OmegaConf open_dictF model_clsz#nemo.core.config.modelPT.NemoConfig update_cfgrdrop_missing_subconfigscCsHts tdtdt|st|tstdt|ts"t |}t|r+t |}t ||d|d}t ||d|d}t ||d|d}t ||d|d}t ||dd }d |j vrwd |j vrwt|j |j d Wd n1srwYd |j vrd |j vrt|j |j d Wd n1swYt||}|S) a\ Helper class that updates the default values of a ModelPT config class with the values in a DictConfig that mirrors the structure of the config class. Assumes the `update_cfg` is a DictConfig (either generated manually, via hydra or instantiated via yaml/model.cfg). This update_cfg is then used to override the default values preset inside the ModelPT config class. If `drop_missing_subconfigs` is set, the certain sub-configs of the ModelPT config class will be removed, if they are not found in the mirrored `update_cfg`. The following sub-configs are subject to potential removal: - `train_ds` - `validation_ds` - `test_ds` - `optim` + nested `sched`. Args: model_cls: A subclass of NemoConfig, that details in entirety all of the parameters that constitute the NeMo Model. update_cfg: A DictConfig that mirrors the structure of the NemoConfig data class. Used to update the default values of the config class. drop_missing_subconfigs: Bool which determins whether to drop certain sub-configs from the NemoConfig class, if the corresponding sub-config is missing from `update_cfg`. Returns: A DictConfig with updated values that can be used to instantiate the NeMo Model along with supporting infrastructure. @This function requires Hydra/Omegaconf and it was not installed.z@`model_cfg` must be a dataclass or a structured OmegaConf objecttrain_ds) subconfig_keyr validation_dstest_dsoptim)rtargetN nemo_version) _HAS_HYDRArerrorexitr isinstancer ValueErrorrcreate structured_update_subconfig_add_subconfig_keysmodelr popmerge)r r r model_cfgr#K/home/ubuntu/.local/lib/python3.10/site-packages/nemo/utils/config_utils.pyupdate_model_configsD           r%r"rcCsts tdtdt|jE||jvr#||jvr#|j||j|<||jvr=||jvrE|rM|j|Wd|SWd|SWd|SWd|S1sXwY|S)a Updates the NemoConfig DictConfig such that: 1) If the sub-config key exists in the `update_cfg`, but does not exist in ModelPT config: - Add the sub-config from update_cfg to ModelPT config 2) If the sub-config key does not exist in `update_cfg`, but exists in ModelPT config: - Remove the sub-config from the ModelPT config; iff the `drop_missing_subconfigs` flag is set. Args: model_cfg: A DictConfig instantiated from the NemoConfig subclass. update_cfg: A DictConfig that mirrors the structure of `model_cfg`, used to update its default values. subconfig_key: A str key used to check and update the sub-config. drop_missing_subconfigs: A bool flag, whether to allow deletion of the NemoConfig sub-config, if its mirror sub-config does not exist in the `update_cfg`. Returns: The updated DictConfig for the NemoConfig r rN)rrrrr rr )r"r rr r#r#r$rns,           rcCsts tdtdt|j;||jvrC||jvr d|j|<t|j|}t|j|}t ||}||j|<Wd|SWd|S1sNwY|S)a? For certain sub-configs, the default values specified by the NemoConfig class is insufficient. In order to support every potential value in the merge between the `update_cfg`, it would require explicit definition of all possible cases. An example of such a case is Optimizers, and their equivalent Schedulers. All optimizers share a few basic details - such as name and lr, but almost all require additional parameters - such as weight decay. It is impractical to create a config for every single optimizer + every single scheduler combination. In such a case, we perform a dual merge. The Optim and Sched Dataclass contain the bare minimum essential components. The extra values are provided via update_cfg. In order to enable the merge, we first need to update the update sub-config to incorporate the keys, with dummy temporary values (merge update config with model config). This is done on a copy of the update sub-config, as the actual override values might be overriden by the NemoConfig defaults. Then we perform a merge of this temporary sub-config with the actual override config in a later step (merge model_cfg with original update_cfg, done outside this function). Args: model_cfg: A DictConfig instantiated from the NemoConfig subclass. update_cfg: A DictConfig that mirrors the structure of `model_cfg`, used to update its default values. subconfig_key: A str key used to check and update the sub-config. Returns: A ModelPT DictConfig with additional keys added to the sub-config. r rN) rrrrr rcopydeepcopyrr!)r"r r subconfigupdate_subconfigr#r#r$rs$         rcls class_typedatacls dataclass ignore_args remap_argsc Cst|j}tdi|j}|dt|}tdi|j}|ddt|}t|}|dury|D]>\}} ||vr[| || | t d|d| d|j ||vrx| || | t d|d| d|j q:|durt|}||}||}t d|t||} || } || } t|t|kst| dkst| dkrt d |j d |j d t| dkrt d |j d | t| rt d |j d | d| | fSdS)aG Analyses the signature of a provided class and its respective data class, asserting that the dataclass signature matches the class __init__ signature. Note: This is not a value based check. This function only checks if all argument names exist on both class and dataclass and logs mismatches. Args: cls: Any class type - but not an instance of a class. Pass type(x) where x is an instance if class type is not easily available. datacls: A corresponding dataclass for the above class. ignore_args: (Optional) A list of string argument names which are forcibly ignored, even if mismatched in the signature. Useful when a dataclass is a superset of the arguments of a class. remap_args: (Optional) A dictionary, mapping an argument name that exists (in either the class or its dataclass), to another name. Useful when argument names are mismatched between a class and its dataclass due to indirect instantiation via a helper method. Returns: A tuple containing information about the analysis: 1) A bool value which is True if the signatures matched exactly / after ignoring values. False otherwise. 2) A set of arguments names that exist in the class, but *do not* exist in the dataclass. If exact signature match occurs, this will be None instead. 3) A set of argument names that exist in the data class, but *do not* exist in the class itself. If exact signature match occurs, this will be None instead. self_target_Nz Remapped z -> z in zRemoving ignored arguments - rzClass z" arguments do not match Dataclass !z has additional arguments : z Dataclass F)TNNr#)inspect signature__init__dict parametersr setkeysitemsremoveaddrinfo__name__ intersectionlenr) r*r,r.r/ class_sig class_params dataclass_sigdataclass_params original_argnew_argr? subset_clssubset_dataclsr#r#r$ assert_dataclass_signature_matchsF "          (  rI)T)NN)r&r3 dataclassesrtypingrrr nemo.utilsrr omegaconfrrr ModuleNotFoundErrorboolr%strrrrIr#r#r#r$sP    P '4