-
Notifications
You must be signed in to change notification settings - Fork 101
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Change to the new deberta model #735
Conversation
This shows how to change from the roberta to the deberta model. However, using it will result in #21. |
Hey @kwalcock, any updates on this? Thank you! |
I'm still digging deep into the Python code. It is going to be difficult. The most recent changes are in the kwalcock/types branch FWIW. |
That message is a warning, and it is possible that the tokens look similar. Did you compare the tokens produced by the Python and Rust tokenizers? |
In any case, if not all tokenizers are supported, it's not the end of the world. We just need to know, so we don't use the deberta model. |
All of the output coming from Python matches whether the value for use_fast is True or False. I'm assuming that False will result in the Python code running instead of Rust, but I have not yet seen exactly how that works. So far all of it has matched the Rust code called from Scala as long as the tokenizer is one of those available directly from Rust. |
The top tokenizers here (and maybe more) work from Scala via Rust without the stop in Python. The trick will be to get the ones that include some kind of Python assistance to work. Aren't the tokenizers paired with models so that this limits the models that can be used? class SentencesTest extends Test {
// See also test_clu_tokenizer.py.
val tokenizerNames = Seq(
"bert-base-cased",
"distilbert-base-cased",
"roberta-base",
"xlm-roberta-base" // ,
// All of these latter ones will not just fail, but cause a
// fatal runtime error and end the testing completely.
// "google/bert_uncased_L-4_H-512_A-8",
// "google/electra-small-discriminator",
// "microsoft/deberta-v3-base"
) |
Yes, limiting the tokenizers will limit the models we have access to. In particular, deberta is an important one for processors. |
Nice!! Ok to merge? |
Yes. Just finished testing. |
This works great! |
No description provided.