Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
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
Adding needed Tokenizer's APIs #7047
Adding needed Tokenizer's APIs #7047
Changes from 1 commit
ca18ba7
4832f9a
d9c1524
5041ebf
765f79a
b9e95b2
1d89506
File filter
Filter by extension
Conversations
Jump to
There are no files selected for viewing
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Existing: what token does this refer to? The only other thing specified is
text
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
token is the string word. can be any word like
dog
or it can be special token like<|endoftext|>
. I can changeIndicate if the token is a special token.
toIndicate if the text represent a special token.
or similar.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is that the parameter
text
, and does thattext
need to represent a single token? Or does it refer to all tokens withintext
?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
the text can represent multiple tokens or represent one special token.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ok, so how does
isSpecialToken
apply to 'text` in the case it is multiple tokens?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It is a flag telling if the input text is representing a special token or not so the encoder can treat it differently. Here is some example how this is used
machinelearning/src/Microsoft.ML.Tokenizers/Model/Tiktoken.cs
Line 353 in 164fde0
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ok, so we should probably refer to
text
in the docs.Here's my attempt:
Similarly it looks like the
Count
andEncodeToIds
just return default values ignoring the text if it's not a special token so they could get a slightly different version of this.machinelearning/src/Microsoft.ML.Tokenizers/Model/Tiktoken.cs
Line 313 in 164fde0
machinelearning/src/Microsoft.ML.Tokenizers/Model/Tiktoken.cs
Line 355 in 164fde0
I raised this issue because this parameter confused me when adopting the tokenizer.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm curious as to the use case for setting isSpecialToken to true...?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Seems like the only scenario is when someone wants to pass in a single special token string and get the value for that. If we're making them specify a parameter to this API to do that they might as well just call a different API to do it and avoid the confusing parameter on this API.
I wonder what happens if someone specifies a special token string but forgets to set isSpecialToken?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@tarekgh and I talked about this offline - we'll update the docs for these methods in a separate PR and discuss this during API review.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I could imagine someone wanting an API like
IEnumerable<byte> Decode(IEnumerable<ids> ids, ...)
. Presumably if that was desired we could always add it in the future.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is it ever the case that someone might have pre-normalized text and want a method that doesn't do this normalization?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, when creating the tokenizer you have the option to provide the normalizer object. If you don't then the tokenizer will not do any normalization before processing the text.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Just imagining that someone might want to call the normalizer once, then tell this method that they've already done the normalization and avoid double-normalization / allocation. My understanding of the use case for this API is to do a minimal amount of work so I was just asking myself "is there anything else that I can imagine someone might not want this method to do?" and normalization was the only thing I could imagine.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Well written normalizer will return the original text without new allocation if there is no change from the original text. But I think processing time will be counted.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Also, users can create a copy of the tokenizer without the normalization and can be used in such scenario too.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Right, but I was more concerned with this API doing two things where someone might want it to do just one. It's the same feeling @stephentoub shared offline
I don't feel like it needs to be solved now, but it may be a topic during API review.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The normalization is part of tokenization. It is optional for some scenario but important to other scenarios. So the API is not really doing two things more than communicating the encoding results including the change in the text. Anyway, I am open to any better suggestion that can make us avoid any confusion or to get cleaner API shape.