Tfjs: TF Node backend does not support normalized logits passed to multinomial

Created on 14 Jan 2020  路  1Comment  路  Source: tensorflow/tfjs

TensorFlow.js version

V1.5.1

Describe the problem or feature request

Just stumbled on this error:
TF Node backend does not support normalized logits passed to multinomial

https://github.com/tensorflow/tfjs/blob/c2d556415db8d9fb611a69a77a6288678ab0f86d/tfjs-node/src/nodejs_kernel_backend.ts#L1561-L1568

I know that as a workaround I can unnormalize the inputs, but what's the actual limitation? Why the node backend doesn't support this?

node.js question others

Most helpful comment

This is because TensorFlow C library multinomial op only support unnormalized logits: https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/multinomial

FYI tfjs-node library is using TensorFlow C library under the hood.

>All comments

This is because TensorFlow C library multinomial op only support unnormalized logits: https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/multinomial

FYI tfjs-node library is using TensorFlow C library under the hood.

Was this page helpful?
0 / 5 - 0 ratings

Related issues

rumschuettel picture rumschuettel  路  3Comments

ritikrishu picture ritikrishu  路  4Comments

pranayaryal picture pranayaryal  路  4Comments

lastnod picture lastnod  路  3Comments

JasonShin picture JasonShin  路  4Comments