convert_ndarray_dtype(np_array: ndarray, target_dtype: str) ndarray[source]#

convert ndarray

  • np_array (np.ndarray) – numpy ndarray instance

  • target_dtype (str) – the target dtype


converted numpy ndarray instance

Return type:


get_scale_by_dtype(dtype: str | None = None, return_positive: bool = True) float[source]#

get scale value by dtype


dtype (str) – the string dtype value


the scale value

Return type:


fn_args_to_dict(func, *args, **kwargs)[source]#

Inspect function func and its arguments for running, and extract a dict mapping between argument names and keys.

adapt_stale_fwd_patch(self, name, value)[source]#

Since there are some monkey patches for forward of PretrainedModel, such as model compression, we make these patches compatible with the latest forward method.

class InitTrackerMeta(name, bases, attrs)[source]#

Bases: type

This metaclass wraps the __init__ method of a class to add init_config attribute for instances of that class, and init_config use a dict to track the initial configuration. If the class has _pre_init or _post_init method, it would be hooked before or after __init__ and called as _pre_init(self, init_fn, init_args) or _post_init(self, init_fn, init_args). Since InitTrackerMeta would be used as metaclass for pretrained model classes, which always are Layer and type(Layer) is not type, thus use type(Layer) rather than type as base class for it to avoid inheritance metaclass conflicts.

static init_and_track_conf(init_func, pre_init_func=None, post_init_func=None)[source]#

wraps init_func which is __init__ method of a class to add init_config attribute for instances of that class. :param init_func: It should be the __init__ method of a class.

warning: self always is the class type of down-stream model, eg: BertForTokenClassification

  • pre_init_func (callable, optional) – If provided, it would be hooked after init_func and called as pre_init_func(self, init_func, *init_args, **init_args). Default None.

  • post_init_func (callable, optional) – If provided, it would be hooked after init_func and called as post_init_func(self, init_func, *init_args, **init_args). Default None.


the wrapped function

Return type:


param_in_func(func, param_field: str) bool[source]#

check if the param_field is in func method, eg: if the bert param is in __init__ method

  • cls (type) – the class of PretrainedModel

  • param_field (str) – the name of field


the result of existence

Return type:


resolve_cache_dir(from_hf_hub: bool, from_aistudio: bool, cache_dir: str | None = None) str[source]#

resolve cache dir for PretrainedModel and PretrainedConfig

  • from_hf_hub (bool) – if load from huggingface hub

  • cache_dir (str) – cache_dir for models

find_transformer_model_type(model_class: Type) str[source]#
get the model type from module name,

BertModel -> bert, RobertaForTokenClassification -> roberta


model_class (Type) – the class of model


the type string

Return type:


find_transformer_model_class_by_name(model_name: str) Type[PretrainedModel] | None[source]#

find transformer model_class by name


model_name (str) – the string of class name


optional pretrained-model class

Return type:


convert_file_size_to_int(size: int | str)[source]#

Converts a size expressed as a string with digits an unit (like "5MB") to an integer (in bytes). :param size: The size to convert. Will be directly returned if an int. :type size: int or str

Example: `py >>> convert_file_size_to_int("1MiB") 1048576 `

cached_file(path_or_repo_id: str | PathLike, filename: str, cache_dir: str | PathLike | None = None, subfolder: str = '', from_aistudio: bool = False, _raise_exceptions_for_missing_entries: bool = True, _raise_exceptions_for_connection_errors: bool = True, pretrained_model_name_or_path=None) str[source]#

Tries to locate a file in a local folder and repo, downloads and cache it if necessary. :param path_or_repo_id: This can be either:

  • a string, the model id of a model repo on huggingface.co.

  • a path to a directory potentially containing the file.

  • filename (str) – The name of the file to locate in path_or_repo.

  • cache_dir (str or os.PathLike, optional) – Path to a directory in which a downloaded pretrained model configuration should be cached if the standard cache should not be used.

  • subfolder (str, optional, defaults to "") – In case the relevant files are located inside a subfolder of the model repo on huggingface.co, you can specify the folder name here.


Returns the resolved file (to the cache folder if downloaded from a repo).

Return type:


Examples: `python # Download a model weight from the Hub and cache it. model_weights_file = cached_file("bert-base-uncased", "pytorch_model.bin") `

get_checkpoint_shard_files(pretrained_model_name_or_path, index_filename, cache_dir=None, subfolder='', from_aistudio=False, from_hf_hub=False)[source]#

For a given model: - download and cache all the shards of a sharded checkpoint if pretrained_model_name_or_path is a model ID on the


  • returns the list of paths to all the shards, as well as some metadata.

For the description of each arg, see [PretrainedModel.from_pretrained]. index_filename is the full path to the index (downloaded and cached if pretrained_model_name_or_path is a model ID on the Hub).

class ContextManagers(context_managers: List[ContextManager])[source]#

Bases: object

Wrapper for contextlib.ExitStack which enters a collection of context managers. Adaptation of ContextManagers in the fastcore library.


Returns the size (in bytes) occupied by one parameter of type dtype.


`py >>> dtype_byte_size(paddle.float32) 4 `

class CaptureStd(out=True, err=True, replay=True)[source]#

Bases: object

Context manager to capture:

  • stdout: replay it, clean it up and make it available via obj.out

  • stderr: replay it and make it available via obj.err

  • out (bool, optional, defaults to True) – Whether to capture stdout or not.

  • err (bool, optional, defaults to True) – Whether to capture stderr or not.

  • replay (bool, optional, defaults to True) – Whether to replay or not. By default each captured stream gets replayed back on context’s exit, so that one can see what the test was doing. If this is a not wanted behavior and the captured data shouldn’t be replayed, pass replay=False to disable this feature.


```python # to capture stdout only with auto-replay with CaptureStdout() as cs:

print(“Secret message”)

assert “message” in cs.out

# to capture stderr only with auto-replay import sys

with CaptureStderr() as cs:

print(“Warning: “, file=sys.stderr)

assert “Warning” in cs.err

# to capture both streams with auto-replay with CaptureStd() as cs:

print(“Secret message”) print(“Warning: “, file=sys.stderr)

assert “message” in cs.out assert “Warning” in cs.err

# to capture just one of the streams, and not the other, with auto-replay with CaptureStd(err=False) as cs:

print(“Secret message”)

assert “message” in cs.out # but best use the stream-specific subclasses

# to capture without auto-replay with CaptureStd(replay=False) as cs:

print(“Secret message”)

assert “message” in cs.out ```