Experimental Callbacks

Miscellaneous experimental callbacks for timeseriesAI.

Gambler’s loss: noisy labels

gambler_loss


def gambler_loss(
    reward:int=2
):

Call self as a function.

source

GamblersCallback


def GamblersCallback(
    after_create:NoneType=None, before_fit:NoneType=None, before_epoch:NoneType=None, before_train:NoneType=None,
    before_batch:NoneType=None, after_pred:NoneType=None, after_loss:NoneType=None, before_backward:NoneType=None,
    after_cancel_backward:NoneType=None, after_backward:NoneType=None, before_step:NoneType=None,
    after_cancel_step:NoneType=None, after_step:NoneType=None, after_cancel_batch:NoneType=None,
    after_batch:NoneType=None, after_cancel_train:NoneType=None, after_train:NoneType=None,
    before_validate:NoneType=None, after_cancel_validate:NoneType=None, after_validate:NoneType=None,
    after_cancel_epoch:NoneType=None, after_epoch:NoneType=None, after_cancel_fit:NoneType=None,
    after_fit:NoneType=None
):

A callback to use metrics with gambler’s loss

from tsai.data.external import *
from tsai.data.core import *
from tsai.models.InceptionTime import *
from tsai.models.layers import *
from tsai.learner import *
from fastai.metrics import *
from tsai.metrics import *

X, y, splits = get_UCR_data('NATOPS', return_split=False)
tfms = [None, TSCategorize()]
dsets = TSDatasets(X, y, tfms=tfms, splits=splits)
dls = TSDataLoaders.from_dsets(dsets.train, dsets.valid, bs=[64, 128])
loss_func = gambler_loss()
learn = ts_learner(dls, InceptionTime(dls.vars, dls.c + 1), loss_func=loss_func, cbs=GamblersCallback, metrics=[accuracy])
learn.fit_one_cycle(1)

epoch	train_loss	valid_loss	accuracy	time
0	1.840055	1.945397	0.166667	00:05

Uncertainty-based data augmentation

source

UBDAug


def UBDAug(
    batch_tfms:list, N:int=2, C:int=4, S:int=1
):

A callback to implement the uncertainty-based data augmentation.

from tsai.models.utils import *

X, y, splits = get_UCR_data('NATOPS', return_split=False)
tfms = [None, TSCategorize()]
dsets = TSDatasets(X, y, tfms=tfms, splits=splits)
dls = TSDataLoaders.from_dsets(dsets.train, dsets.valid, batch_tfms=[TSStandardize()])
model = build_ts_model(InceptionTime, dls=dls)
TS_tfms = [TSMagScale(.75, p=.5), TSMagWarp(.1, p=0.5),  TSWindowWarp(.25, p=.5), 
           TSSmooth(p=0.5), TSRandomResizedCrop(.1, p=.5), 
           TSRandomCropPad(.3, p=0.5), 
           TSMagAddNoise(.5, p=.5)]

ubda_cb = UBDAug(TS_tfms, N=2, C=4, S=2)
learn = ts_learner(dls, model, cbs=ubda_cb, metrics=accuracy)
learn.fit_one_cycle(1)

epoch	train_loss	valid_loss	accuracy	time
0	1.817080	1.791119	0.077778	00:14

BatchLossFilter

source

BatchLossFilter


def BatchLossFilter(
    loss_perc:float=1.0, schedule_func:Optional=None
):

Callback that selects the hardest samples in every batch representing a percentage of the total loss

RandomWeightLossWrapper

source

RandomWeightLossWrapper


def RandomWeightLossWrapper(
    after_create:NoneType=None, before_fit:NoneType=None, before_epoch:NoneType=None, before_train:NoneType=None,
    before_batch:NoneType=None, after_pred:NoneType=None, after_loss:NoneType=None, before_backward:NoneType=None,
    after_cancel_backward:NoneType=None, after_backward:NoneType=None, before_step:NoneType=None,
    after_cancel_step:NoneType=None, after_step:NoneType=None, after_cancel_batch:NoneType=None,
    after_batch:NoneType=None, after_cancel_train:NoneType=None, after_train:NoneType=None,
    before_validate:NoneType=None, after_cancel_validate:NoneType=None, after_validate:NoneType=None,
    after_cancel_epoch:NoneType=None, after_epoch:NoneType=None, after_cancel_fit:NoneType=None,
    after_fit:NoneType=None
):

Basic class handling tweaks of the training loop by changing a Learner in various events

BatchMasker

source

BatchMasker


def BatchMasker(
    r:float=0.15, lm:int=3, stateful:bool=True, sync:bool=False, subsequence_mask:bool=True,
    variable_mask:bool=False, future_mask:bool=False, schedule_func:Optional=None
):

Callback that applies a random mask to each sample in a training batch

Args:

r: probability of masking. subsequence_mask: apply a mask to random subsequences. lm: average mask len when using stateful (geometric) masking. stateful: geometric distribution is applied so that average mask length is lm. sync: all variables have the same masking. variable_mask: apply a mask to random variables. Only applicable to multivariate time series. future_mask: used to train a forecasting model. schedule_func: if a scheduler is passed, it will modify the probability of masking during training.

SamplerWithReplacement

source

SamplerWithReplacement


def SamplerWithReplacement(
    after_create:NoneType=None, before_fit:NoneType=None, before_epoch:NoneType=None, before_train:NoneType=None,
    before_batch:NoneType=None, after_pred:NoneType=None, after_loss:NoneType=None, before_backward:NoneType=None,
    after_cancel_backward:NoneType=None, after_backward:NoneType=None, before_step:NoneType=None,
    after_cancel_step:NoneType=None, after_step:NoneType=None, after_cancel_batch:NoneType=None,
    after_batch:NoneType=None, after_cancel_train:NoneType=None, after_train:NoneType=None,
    before_validate:NoneType=None, after_cancel_validate:NoneType=None, after_validate:NoneType=None,
    after_cancel_epoch:NoneType=None, after_epoch:NoneType=None, after_cancel_fit:NoneType=None,
    after_fit:NoneType=None
):

Callback that modify the sampler to select a percentage of samples and/ or sequence steps with replacement from each training batch