batch_tfms not applied to valid batch?

[yttuncel]

trafficstars

Hi,

I've came across a strange issue with the batch transform. The after_batch transform is not applied to dls.valid. Is this intended behavior?

I've replicated the issue with the below sample code:

from tsai.all import *
X, y, splits = get_UCR_data('LSST', split_data=False)
tfms = [None, TSClassification()]
batch_tfms = [TSSmooth(filt_len=10000)]
dls = get_ts_dls(X, y, splits=splits, tfms=tfms, batch_tfms=batch_tfms)
dls.train.show_batch()
dls.valid.show_batch()

Basically I set a very long smoothing filter to destroy the data. I expect to observe flat lines as a result of this: Output of dls.train.show_batch():

Output of dls.valid.show_batch():

As you can see, the train batch is all flat. However, the transform is not applied to the valid batch. I'd appreciate any help on this.

Thanks, Yigit

[vrodriguezf]

Does that happen only with augmentation Transforms (those that inherit from fastai's RandTranform) or it happens with normal preprocessing transforms too?

[yttuncel]

@vrodriguezf I didn't do a thorough testing of this, but using TSRollingMean() instead of TSSmooth() addressed the problem. TSRollingMean() inherits Transform and not RandTransform, so you might be correct with your above guess.

[oguiza]

Hi @yttuncel , Victor is right. This works as design by fastai. Here's an extract from their documentation:

A RandTransform is only applied to the training set by default, so you have to pass split_idx=0 if you are calling it directly and not through a Datasets. That behavior can be changed by setting the attr split_idx of the transform to None.

So if you want to use TSSmooth(filt_len=10000) on both training, validation and testing you should do it this way:

from tsai.all import *
X, y, splits = get_UCR_data('LSST', split_data=False)
tfms = [None, TSClassification()]
tsmooth = TSSmooth(filt_len=10000)
tsmooth.split_idx = None
batch_tfms = [tsmooth]

dls = get_ts_dls(X, y, splits=splits, tfms=tfms, batch_tfms=batch_tfms)
dls.train.show_batch()
dls.valid.show_batch()

I've tested it and it works well. Please, try it if you can. If it does, feel free to close this issue.

[oguiza]

Issue closed due to lack of response.