CODEWITHSUNDEEP
pythontensorflownlpbert-language-model
tmwstw
tmwstw

Reputation: 11

Can I use non-fine-tuned BERT model from TF HUB to serve it with TF serving?

I'm a new to TF serving and currently I have such kind of problem. I run server part using bert_en_uncased from TF HUB, but I don't understand how to implement client side correctly. I faced with a couple of articles but each of them assumes that I have a ready-made fine-tuned model with pre-assigned handlers for requests. Can anyone share some tutors or maybe API references to facilitate my task?

Some of articles I have read:

PS. I'm not trying to create QA model or something like that, I just need BERT embeddings from this particular model.

Upvotes: 0

Views: 231

Answers (1)

tmwstw
tmwstw

Reputation: 11

UPD: a've already solved this problem. The main thing was, TF.HUB model don't have any spec list or something like that, only some documentation of how you can use it with tf.hub. If you faced with similar problem do I recommend to do the following things: 1) Install/compile from source SavedModelCli, it's TensorFlow's tool to, let's say, unpack saved models and get it's specs; 2) Find some guides on TF Serving, just change some code pieces, nearly every implementation is the same; 3) Probably you might (and you WILL, believe me) face with deprecation warnings. Don't try to look for documentation, solution was here :) Good luck on serving your models!

Upvotes: 1

Related Questions