전력 AI 모델 허브포털

Filters

Collection
mobilebert-emobilebe
MobileBert-EdgeTPU is a set of language pretraining models that are optimized for Pixel 6.
  • Published by : Google
  • Last Updated : 2022-03-04
  • Models : 3
Overview

By co-designing the neural networks with the Edge TPU hardware accelerator in Tensor (Google's first phone SoC and the heart of Pixel 6), we have built EdgeTPU-customized MobileBERT models that demonstrate datacenter model quality meanwhile outperforms baseline MobileBERT's latency.

We set up our model architecture search space based on MobileBERT and leverage AutoML algorithms to find models with up to 2x better hardware utilization. With higher utilization, we are able to bring larger and more accurate models on chip, and meanwhile the models can still outperform the baseline MobileBERT latency. The quantized MobileBERT-EdgeTPU models (int8 quantized tflite will be available soon) establish a new pareto-frontier for the question answering tasks and overshot the accuracy of the float BERT_base model which is 400+MB and too large to run on edge devices.

Models
Name Language Vocabulary size Cased
en_sp35k_cased English 35k cased
en_sp35k_cased English 35k cased
en_sp35k_cased English 35k cased