TensorFlow Datasets (TF Dev Summit '19)

Speaker:
Ryan Sepassi, Software Engineer

event: TensorFlow Dev Summit 2019; re_ty: Publish; product: TensorFlow - General; fullname: Ryan Sepassi;

[samg]

2:40 - 5:34 bug fix at second to last line:
train_ds =

[rcampbellbassac]

I’m currently working on a project for classifying Cambodian crop diseases/pests. I plan on collecting a large amount of images and labelling them with the diseases/conditions/pests using a smartphone (a phone a Cambodian farmer could afford, example: Nokia 1). My idea is to create a dataset I can then use to develop an app/pwa that can allow a user to figure out what is wrong with their crop using an inexpensive smartphone.

I think it would be awesome if the Tensorflow Datasets team could build a simple image collection app that allows you to take a shot and easily add a label (either while you are taking the photos or after the photos have been taken and analyzed). Even sweeter if it allowed you to export your collection in a TF datasets format, and that the app would then help you obtain a link you could use in code with TF Datasets. Sweeter yet, it could store it with Google Cloud Storage or Drive, and it all just kinda worked together… Add in their some Colab support, etc.

[thaihocnguyen]

How can i create a my own dataset with tfds ?

[JacekBrauner]

Any example for sequence to point prediction for Lstm that data source is located in many files. My current pipeline has to loop through the files create sequences using window method and return as dataset then loop throug it and do forward and back pass . I would like to know if there is better way to do the same.

[bilalchandio]

How can i convert my text dataset into encoded subwords8k like tensor dataset?

[shreyanshmehta]

I wanted to know that "MNIST" data set is present in which directory of my system. (it should be store in the default directory
but it is not present in my case.
can someone help me out with this?

[frankl]

Hi,
How to do k-fold cross validation with Tensorflow Dataset ?

[firexgodx]

Not very useful for me, because I often have to transform labels into various different formats. TFRecords on the other hand are very cool.