This PR introduces signal augmentation. During inference, a specific signal (one at a time) is augmented based on the string supplied in the config file conf['data']['signal_to_augment'] (string or None). During training, augment a random signal (again, one at a time) or do not augment at all (conf['data']['augment_during_training']).
Augmentation can either be a random normal noise or zeroing a signal.
The augmentation is implemented as a class hierarchy - a base class AbstractAugmentator and currently only one derived class Augmentator in a dedicated module plasma.preprocessor.augment. Augmentator class uses object composition and defines a data member of type Normalizer. So augmentator wraps normalizer and applies normalization to a shot before augmenting it.
To use in the driver script first create a normalizer object, and then pass it to Augmentator:
from plasma.preprocessor.normalize import Normalizer
from plasma.preprocessor.augment import Augmentator
from plasma.models.loader import Loader
from plasma.conf import conf
raw_normalizer = Normalizer(conf)
raw_normalizer.train()
is_inference= False
normalizer = Augmentator(raw_normalizer,is_inference,conf)
loader = Loader(conf,normalizer)
Switch to inference mode after training of the NN model is complete:
normalizer.set_inference(True)
Create a dedicated example mpi_augment_learn.py driver script to worry less about defaults of augmentation specific config parameters.
Side change: switch from shutil to os.system to perform bash commands, and fixes to tune_hyperparameter.py
This PR introduces signal augmentation. During inference, a specific signal (one at a time) is augmented based on the string supplied in the config file
conf['data']['signal_to_augment'](string or None). During training, augment a random signal (again, one at a time) or do not augment at all (conf['data']['augment_during_training']).Augmentation can either be a random normal noise or zeroing a signal.
The augmentation is implemented as a class hierarchy - a base class
AbstractAugmentatorand currently only one derived classAugmentatorin a dedicated moduleplasma.preprocessor.augment.Augmentatorclass uses object composition and defines a data member of typeNormalizer. So augmentator wraps normalizer and applies normalization to a shot before augmenting it.To use in the driver script first create a normalizer object, and then pass it to Augmentator:
Switch to inference mode after training of the NN model is complete:
Create a dedicated example
mpi_augment_learn.pydriver script to worry less about defaults of augmentation specific config parameters.Side change: switch from
shutiltoos.systemto perform bash commands, and fixes totune_hyperparameter.py