Open pgyrya opened 6 years ago
So why not first embed the sparse inputs then use an MOE layer on those dense representations? What did you have in mind?
The example I previously worked on was a bag-of-words representation of text, where embedding doesn’t seem like a natural choice.
From: Ryan Sepassi notifications@github.com Sent: Thursday, May 3, 2018 10:42 PM To: tensorflow/tensor2tensor tensor2tensor@noreply.github.com Cc: Pavel Gyrya pavel.gyrya@cornell.edu; Author author@noreply.github.com Subject: Re: [tensorflow/tensor2tensor] Requesting Sparse Input Support - MOE layer (#665)
So why not first embed the sparse inputs then use an MOE layer on those dense representations? What did you have in mind?
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/tensorflow/tensor2tensor/issues/665#issuecomment-386492854 , or mute the thread https://github.com/notifications/unsubscribe-auth/AeFaQe5NzwzuH02SUe4ycfCixxjSLTB7ks5tu7_wgaJpZM4S5QYA . https://github.com/notifications/beacon/AeFaQd1Pbb3HD-Riz0K2RFh90_6rXaiDks5tu7_wgaJpZM4S5QYA.gif
feature request
Description
I would like to suggest to prominently add support for sparse input tensors - specifically for mixture-of-expert gating functionality (e.g. local_moe and noisy_top_k_gating).
Sparse input support is prominently important when installing layers in the beginning of the network that relies on sparsely encoded data - a common phenomenon in large-scale feature problems.