Skip to content

archival

[ allennlp.models.archival ]

Helper functions for archiving models and restoring archived models.

Archive Objects#

class Archive(NamedTuple)

An archive comprises a Model and its experimental config

model#

model = None

config#

config = None

extract_module#

 | def extract_module(self, path: str, freeze: bool = True) -> Module

This method can be used to load a module from the pretrained model archive.

It is also used implicitly in FromParams based construction. So instead of using standard params to construct a module, you can instead load a pretrained module from the model archive directly. For eg, instead of using params like {"type": "module_type", ...}, you can use the following template::

{
    "_pretrained": {
        "archive_file": "../path/to/model.tar.gz",
        "path": "path.to.module.in.model",
        "freeze": False
    }
}

If you use this feature with FromParams, take care of the following caveat: Call to initializer(self) at end of model initializer can potentially wipe the transferred parameters by reinitializing them. This can happen if you have setup initializer regex that also matches parameters of the transferred module. To safe-guard against this, you can either update your initializer regex to prevent conflicting match or add extra initializer::

[
    [".*transferred_module_name.*", "prevent"]]
]

Parameters

  • path : str
    Path of target module to be loaded from the model. Eg. "_textfield_embedder.token_embedder_tokens"
  • freeze : bool, optional (default = True)
    Whether to freeze the module parameters or not.

CONFIG_NAME#

CONFIG_NAME = "config.json"

archive_model#

def archive_model(
    serialization_dir: str,
    weights: str = _DEFAULT_WEIGHTS,
    archive_path: str = None
) -> None

Archive the model weights, its training configuration, and its vocabulary to model.tar.gz.

Parameters

  • serialization_dir : str
    The directory where the weights and vocabulary are written out.
  • weights : str, optional (default = _DEFAULT_WEIGHTS)
    Which weights file to include in the archive. The default is best.th.
  • archive_path : str, optional (default = None)
    A full path to serialize the model to. The default is "model.tar.gz" inside the serialization_dir. If you pass a directory here, we'll serialize the model to "model.tar.gz" inside the directory.

load_archive#

def load_archive(
    archive_file: str,
    cuda_device: int = -1,
    opt_level: str = None,
    overrides: str = "",
    weights_file: str = None
) -> Archive

Instantiates an Archive from an archived tar.gz file.

Parameters

  • archive_file : str
    The archive file to load the model from.
  • cuda_device : int, optional (default = -1)
    If cuda_device is >= 0, the model will be loaded onto the corresponding GPU. Otherwise it will be loaded onto the CPU.
  • opt_level : str, optional (default = None)
    Each opt_level establishes a set of properties that govern Amp’s implementation of pure or mixed precision training. Must be a choice of "O0", "O1", "O2", or "O3". See the Apex documentation for more details. If None, defaults to the opt_level found in the model params. If cuda_device==-1, Amp is not used and this argument is ignored.
  • overrides : str, optional (default = "")
    JSON overrides to apply to the unarchived Params object.
  • weights_file : str, optional (default = None)
    The weights file to use. If unspecified, weights.th in the archive_file will be used.