bumblebee
bumblebee copied to clipboard
Support MarkupLM
I'm trying to use this huggingface model but I'm getting:
** (Mix) Could not start application extractor: exited in: Extractor.Application.start(:normal, [])
** (EXIT) an exception was raised:
** (RuntimeError) could not infer model type from the configuration, please specify the :module and :architecture options
(bumblebee 0.2.0) lib/bumblebee.ex:297: Bumblebee.load_spec/2
(bumblebee 0.2.0) lib/bumblebee.ex:411: Bumblebee.load_model/2
(extractor 0.1.0) lib/extractor/application.ex:20: Extractor.Application.start/2
(kernel 8.1.2) application_master.erl:293: :application_master.start_it_old/4
{:ok, model} = Bumblebee.load_model({:hf, "microsoft/markuplm-base"})
{:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "microsoft/markuplm-base-finetuned-websrc"})
I can't seem to find the information anywhere in the huggingface docs or config. Can someone guide someone not very versed in ML.
I believe it means the model is not supported. We should probably improve the error message to say something like "could not infer model type from the configuration, this model is not supported out of the box by Bumblebee, please specify the :module and :architecture of a custom implementation".
Hey @chrisgreg! MarkupLM is not supported currently. The error should be more specific, that's what I get for microsoft/markuplm-base:
** (RuntimeError) could not match the class name "MarkupLMForPretraining" to any of the supported models, please specify the :module and :architecture options
Ahh.. thanks @josevalim and @jonatanklosko - I guess I'll wait until it is unless either of you know of any currently supported models that let me pass it some HTML and detect which is the "main" part of the document with the important content?
I'd love to contribute but I'm very new to anything ML related so I'm far more of a consumer than a contributor at this point 😅
I've been looking for models to port over from huggingface transformers. Maybe I'll take a crack at MarkupML
I've been looking for models to port over from huggingface transformers. Maybe I'll take a crack at MarkupML
Please do 😄