Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Universal Sentence Encoder Model runs very slow after embedding large datasets

See original GitHub issue

TensorFlow.js version 2.0.1

Node Version 12.18.0

OS Windows 7

Prerequisite: yarn add @tensorflow/tfjs @tensorflow/tfjs-node @tensorflow-models/universal-sentence-encoder

Steps to Reproduce:

Download and unzip debug use model.zip
Run node checktime.js and note down the time taken to execute
Run node embedlargedata.js, this script will embed large amount of data using universal sentence encoder model (takes around 1 hr)
Run node checktime.js and note down the time taken to execute You will observe that after embedding large amount of data, the universal sentence encoder model will run very slow even for small datasets.

Issue Analytics

State:
Created 3 years ago
Comments:5

Top GitHub Comments

1reaction

annxingyuancommented, Jun 25, 2020

@pyu10055 I think the issue is that on the benchmark page we get the USE output data as part of the predict function, which is unlike all our other models (so the benchmarking script’s tensor cleanup mechanism fails to apply to the USE). I sent a fix here: https://github.com/tensorflow/tfjs/pull/3510

0reactions

google-ml-butler[bot]commented, Aug 14, 2020

Are you satisfied with the resolution of your issue? Yes No

Top Results From Across the Web

javascript - Universal Sentence Encoder tensorflowjs optimize ...

The time taken to create embeddings is too large as I need to create embedding for more than 1000 sentences which takes around...

tensorflow hub Universal Sentence Encoder prediction kept ...

As I generate more batches, it's getting slower to predict embedding. I was wondering if there a better way to do this?

Experiments Using Universal Sentence Encoder Embeddings ...

Running this model with equivalent speed to the DAN, assuming perfect linear speedup, would require 340 cores. ACCURACY. Given the enormous ...

Common issues | TensorFlow Hub

Often this is a problem specific to the machine running the code and not an issue with the library. Here is a list...

Universal Sentence Encoder - Google Research

We present models for encoding sentences into embedding vectors that specifically target transfer learning to other NLP tasks. The models are efficient and ......