Different data owners

How can we use this library for “actual” different data owners?

For example, in the “simple-average” example: https://github.com/tf-encrypted/tf-encrypted/blob/master/examples/simple-average/run.py

It works fine because each owner’s data is randomly generated in runtime: def provide_input() -> tf.Tensor: return tf.random_normal(shape=(10,))

Hence each machine will generate its own data during graph execution.

But what if the data should come from external sources, like a .csv file? For example, ‘inputter-0’ has “data0.csv” on its machine 0, ‘inputter-1’ has “data1.csv” on its machine 1, etc… NOTE that these files are NOT on the machine that executes the script “run.py”.

[zicofish]

Hi @mortendahl , thanks for the response!

I tried some simple code with tf.data, indeed, it solves my problem 👍

[jvmncs]

From what I can tell, TensorFlow IO aims to produce a tf.data.Dataset object from a variety of file system backends. Since we don’t make any assumptions about how this object is created, just that it exists, it should work without any adjustments to TFE (in theory!). Would love to see an example/POC around this.