The only sampling-specific method in the API is tf.contrib.data.rejection_resample(). If that doesn't work for your application, can you specify what sampling operation you'd like? Like rejection_resample() it may be possible to build it out of existing parts, like Dataset.filter() and tf.contrib.data.scan().

I think explicit methods for giving weights corresponding to the label would be highly useful. e.g. Dataset.resample(label, weight)

The original poster has replied to this issue after the stat:awaiting response label was applied.

The tf.contrib.data.rejection_resample() transformation allows you to specify a mapping from Dataset elements to an integer class, and a vector of probabilities representing the target distribution (indexed by class) for the resampling. Can you be more specific about what a Dataset.resample() method would do in addition to this?

Nagging Awaiting Response: It has been 14 days with no activityand the awaiting response label was assigned. Is this still an issue?

It has been 14 days with no activity and the awaiting response label was assigned. Is this still an issue?

Is it possible to sample rows from our dataset using a custom distribution like tf.distributions.Exponential?

Hello all,
I'm also interested if it's possible to batch elements from Dataset with distribution, so to pick last elements more often than elements in the begin of dataset.

I'm talking about numeric dataset made from tensor_slices with initial shape 1000103

In this case depending on distribution probability we'll take elements with index [999,:,:] more often than elements with index [1,:,:].

It has been 14 days with no activity and the awaiting response label was assigned. Is this still an issue?

Is there a workaround to create a weighted sampler?