tft.quantiles

Computes the quantile boundaries of a Tensor over the whole dataset.

Quantile boundaries are computed using approximate quantiles, and error tolerance is specified using epsilon. The boundaries divide the input tensor into approximately equal num_buckets parts. See go/squawd for details, and how to control the error due to approximation. NaN input values and values with NaN weights are ignored.

x An input Tensor.
num_buckets Values in the x are divided into approximately equal-sized buckets, where the number of buckets is num_buckets. The number of returned quantiles is num_buckets - 1.
epsilon Error tolerance, typically a small fraction close to zero (e.g. 0.01). Higher values of epsilon increase the quantile approximation, and hence result in more unequal buckets, but could improve performance, and resource consumption. Some measured results on memory consumption: For epsilon = 0.001, the amount of memory for each buffer to hold the summary for 1 trillion input values is ~25000 bytes. If epsilon is relaxed to 0.01, the buffer size drops to ~2000 bytes for the same input size. The buffer size also determines the amount of work in the different stages of the beam pipeline, in general, larger epsilon results in fewer and smaller stages, and less time. For more performance trade-offs see also http://web.cs.ucla.edu/~weiwang/paper/SSDBM07_2.pdf
weights (Optional) Weights tensor for the quantiles. Tensor must have the same batch size as x.
reduce_instance_dims By default collapses the batch and instance dimensions to arrive at a single output vector. If False, only collapses the batch dimension and outputs a vector of the same shape as the input.
name (Optional) A name for this operation.

The bucket boundaries represented as a list, with num_bucket-1 elements, unless reduce_instance_dims is False, which results in a Tensor of shape x.shape + [num_bucket-1]. See code below for discussion on the type of bucket boundaries.