Allow LabelMaker to return TFRecords from NPZ

[martham93]

@wronk and I think it will be important for the ffda-poi work as well as for future label-maker users to have an optional step to convert the NPZ files generated from label-maker package.

To address this, we can open a PR that would allow for an optional step 6 in the label maker workflow, and could work as a cli command label-maker tfrecords, to write out a.tfrecords file for train, test, and if applicable, val, using numpy array representations of the images and labels from label-maker package.

Initially this PR will just address adding tfrecords for classification data to stay within the ffda-poi scope. I do see that there is already some existing work around tfrecords creation for object detection data https://github.com/developmentseed/label-maker/blob/master/examples/utils/tf_records_generation.py.

[martham93]

the implementation I'm currently working on does require tensorflow2, and tensorflow2 has some conflicts with label-maker's requirements.txt, would it be okay if to upgrade these to work with tf2 or do we want to avoid having tensorflow as a dependency?

Happy to approach this a different way, if people have thoughts of something else to try.

cc @Geoyi @wronk @drewbo

[wronk]

@martham93, thinking more about this, I'm not sure we want to add TF as a dependency. I think the options are:

1. Code this functionality up separately and link to a public example in label-maker's documentation. However, don't explicitly support it in this codebase.
2. Include the code in label-maker, make requirements updates (maybe in a separate PR), and make the TF functionality optional. We can do this by only importing TF imports within your data->TFRecord file and only include tf with the requirements-dev.txt file.

Feel free to continue coding this separately and we can drop in the functionality later depending on what we decide to do.

@drewbo, @Geoyi

[martham93]

Right now code to convert data.npz is in the ffda-poi repo. I can work on making a public example (once I've made more progress on other ffda-poi work) and link to label-maker's documentation, to move forward with option 1 presented above.

A few things to note:

• Converting a ~1.8GB data.npz file takes about 3.5 hours to run. The corresponding .tfrecords file will be ~1GB.

• kudos to @wronk for figuring out that the image array needs to be converted to a tensor, then run through tf.image.encode_png function to reduce the size before converting to a tfrecord without this step the data size is huge!

cc @Geoyi