Train, Test and Val TFRecords files partitions always created

[jmarrietar]

Describe the bug
Train, Test and Val TFRecords files always created even if I only specify TRAIN on CSV file

To Reproduce
Change Split to only TRAIN

split,image_uri,label
TRAIN,../tfrecorder/test_data/images/cat/cat-640x853-1.jpg,cat
TRAIN,../tfrecorder/test_data/images/cat/cat-800x600-2.jpg,cat
TRAIN,../tfrecorder/test_data/images/cat/cat-800x600-3.jpg,cat
TRAIN,../tfrecorder/test_data/images/goat/goat-640x640-1.jpg,goat
TRAIN,../tfrecorder/test_data/images/goat/goat-320x320-2.jpg,goat
TRAIN,../tfrecorder/test_data/images/goat/goat-640x427-3.jpg,goat

Expected behavior
Only create a partitions for TRAIN TFRecords

Screenshots

System (please complete the following information):

• OS: [ iOS]
• Python Version: [3.7.4]
• TensorFlow Version: [2.2.0]

Additional context
Not sure if this a Bug or this indeed is expected behavior. But as a user If my CSV partitions file only specified TRAIN is strange to create the other files (test and val) are also created, but without images.

[mbernico]

This behavior makes sense, knowing the code, but it does seem unexpected/weird to a user. I think we should probably address it. We are getting ready to add some features for 0.2's input schema parsing. I think it would be good to address this as part of that change. @cfezequiel awareness.

[cfezequiel]

Hi @jmarrietar , thanks for the feedback. Yes it makes sense to just write TFRecord files for splits with data. I've created a fix in #25 .

One caveat is that Apache Beam doesn't really support conditional pipeline transforms (e.g. if dataset is not empty, run transform), so the check for train/validation/test counts is done at the input (DataFrame) level. The issue is that when the pipeline is run, it may flag some erroneous samples in a split as 'discarded', which may still result in an empty TFRecord file.

For instance, let's say all samples in the test dataset are erroneous. The pipeline may still generate a test*.tfrecords.gz file as there were valid samples in the test dataset when it was counted, even though all samples were marked as 'discarded'. This would probably not be a common case however, and it would be up to the user to ensure that their data is correct.