Training

__init__(training_name, from_config=False, model=None, dataset=None, from_preset='adaptive', pytest=False, verbose=False, **kwargs_optional)

Parameters:

Name	Type	Description	Default
training_name	str	e.g. 'my_training'	required
from_config	bool	True => read parameters from config file (static definition). False => use Training arguments (dynamic definition)	False
model	Optional[str]	[equivalent to model_name] e.g. 'bert-base-cased'	None
dataset	Optional[str]	[equivalent to dataset_name] e.g. 'conll2003'	None
from_preset	Optional[str]	True => use parameters from preset	'adaptive'
pytest	bool	only for testing, don't specify	False
verbose	bool	True => verbose output	False
**kwargs_optional	Any	parameters	{}

get_result(metric='f1', level='entity', label='micro', phase='test', average=True)

Parameters:

Name	Type	Description	Default
metric	str	"f1", "precision", "recall"	'f1'
level	str	"entity" or "token"	'entity'
label	str	"micro", "macro", "PER", ..	'micro'
phase	str	"val" or "test"	'test'
average	bool	if True, return average result of all runs. if False, return result of best run.	True

Returns:

Name	Type	Description
result	Optional[str]	e.g. "0.9011 +- 0.0023" (average = True) or "0.9045" (average = False)

run()

run a single training.

Note:

• from_config == True -> training config file is used, no other optional arguments will be used

• from_config == False -> training config file is created dynamically, optional arguments will be used

  • model and dataset are mandatory.

  • All other arguments relate to hyperparameters and are optional. They are determined using the following hierarchy:

    1) optional argument

    2) from_preset (adaptive, original, stable), which specifies e.g. the hyperparameters "max_epochs", "early_stopping", "lr_schedule"

    3) default training configuration

show_config()

print training config