Austin
I personally see this as being a nice sister project (so maybe it should get its own repo?), even if code reusability isn't totally possible. Especially since D4J doesn't necessarily...
Something to sort out before designing the actual API would probably be to figure out the Token hierarchy. I think a polymorphic hierarchy design would be good to use. For...
We could also (optionally) pass a configuration object to the method to set things like quotation behavior and I guess also toggling things like escapes.
@qwbarch I think WordToken on its own might be useful. Although I guess in my mind, a better name for it should've been something like `EntityToken`. Because my original idea...
@Mee42 That was one thought I had. But it would probably need to be configurable to meet all use cases as D4J tries not to be super opinionated.
Possibly something to consider, but I don't necessarily expect any lexer to be super complex, as I think super complex tokenization might be better left up to command libraries.
Hi, thanks for the report. Since support has been deprecated for years I cannot guarantee that this will be fixed in a timely manner. In the mean time, you could...
Ah, that's unfortunate. Thanks for the info! Feel free to close.
@fritzo I might be willing to try to tackle this, do you have any opinions on how to expose the functionality to the end user?
It looks like I might need to try option 3 since AMP-aware gradient scaling requires access to the optimizer's step() function. I could try making it a boolean flag for...