Implement a tensorflow-native interface to the petastorm reader

[selitvin]

In its current state, the pattern for reading data into Tensoflow:

reader = Reader(...)

readout_examples = reader.tf_tensors()

with tf.Session() as sess:
  ...
reader.stop()
reader.join()

Can we make the usage pattern to have more Tensorflow native flavor?

readout_examples = dataset_toolkit_reader(...)

with tf.Session() as sess:
  ...

To do so:

Thread termination should be controlled by tf.train.Coordinator Question? Who should be allocating/deallocating Reader object and when?

[psood708]

hi @selitvin is this still open to contribute?

[selitvin]

The project is no longer actively maintained. Nevertheless I can review the pr if you’ll put it up.