Match token value with array and tuple

[shavit]

• Add match cases for token value
• Add BartTokenizer to known tokenizers

/Close #182

[greenrazer]

Thank you for the contribution @shavit! I tested it and facebook/bart-large-mnli (#182) is now working with the AutoTokenizer.

However this is a curious issue, because when pulling from a URL AutoTokenzier.from calls:

• LanguageModelConfigurationFromHub.init
• Then LanguageModelConfigurationFromHub.loadConfig
• Then calls this code from HubAPI.swift:

func configuration(fileURL: URL) throws -> Config {
    let data = try Data(contentsOf: fileURL)
    let parsed = try JSONSerialization.jsonObject(with: data, options: [])
    guard let dictionary = parsed as? [NSString: Any] else { throw Hub.HubClientError.parse }
    return Config(dictionary)
}

I could be missing something, but since JSON doesn't support tuples, wouldn't we expect the value entries in the config to be arrays instead of tuples? I'm a bit confused why this worked previously. @pcuenca @FL33TW00D

[shavit]

Right, I agree. The config likely decodes to a dictionary with any byte values, and the token value casts in a dynamic lookup.

I'll do a few more checks.

[pcuenca]

I think you are right @greenrazer! I don't think this could ever work with actual tokenizer files. sep and cls would be decoded as arrays, and I don't think we can cast them to tuples. The correct way to do it is the current one proposed in this PR. Actually, I don't think this would work either.

[shavit]

True, but the JSON array was decoded into Any?, not a tuple. The value is then read from a new Config and matched with the correct type.
https://github.com/huggingface/swift-transformers/blob/main/Sources/Hub/Hub.swift#L83

The problem comes from the token order, which is not index-first.

{
  "type": "RobertaProcessing",
  "sep": [
    "</s>",
    2
  ],
  "cls": [
    "<s>",
    0
  ],
}

[pcuenca]

Do we know if both versions exist in the wild?

[pcuenca]

Thanks for the test! Could we please add another one that loads a tokenizer such as facebook/bart-large-mnli? We don't need anything fancy: just show that the tokenizer loads and that it can encode a string.