Pierre Guillou

> The current logic of misspell identification relies on vocab.txt from the transformer model. For not so common words tokenizers breaks them into subwords and hence the original entire word...

Hi @kcarnold. > I hacked around this by reading the `arpa` file. Can you publish your code for doing that? Thanks.

> it should be `model_name`, not `model_name_or_path` I agree. This is exactly my point: the name of the model is important to get back our ONNX model with `get_onnx_model()`, not...

Hello @subhamkhemka. I guess you use a recent version of transformers (4.11.3 is the actual version)? Unfortunately, I think [onnx_transformers](https://github.com/patil-suraj/onnx_transformers) is no longer up to date (see this [post](https://discuss.huggingface.co/t/new-pipeline-for-zero-shot-text-classification/681/70) of...

Hi @NielsRogge, Thank you but you consider in your code that I already have the corresponding bounding boxes as input, but I don't. I have only the image of a...

Issue opened in the Optimum library: https://github.com/huggingface/optimum/issues/1024

Bonjour, En ce qui concerne le LM général, je vous conseille d'utiliser directement mon 3ème modèle qui est plus performant (il utilise la configuration MultiFit, alors que le premier utilise...

> La partie fine-tuning sur lm3-french-classifier-amazon.ipynb est bien celle comprenant Fine-tuning "forward LM" et Fine-tuning "backward LM" ? Oui, c'est cela (forward et backward pour entraîner un LM bidirectionnel). 1....

Il est certain qu'il faut passer à des Transformers du type BERT pour la génération de texte. Et si seul BERT a été entraîné en français, ça vaut le coup...

Autre conseil: regarder aussi [lm3-portuguese.ipynb](https://github.com/piegu/language-models/blob/master/lm3-portuguese.ipynb) et [lm3-portuguese-classifier-TCU-jurisprudencia.ipynb](https://github.com/piegu/language-models/blob/master/lm3-portuguese-classifier-TCU-jurisprudencia.ipynb) qui utilisent toutes les techniques MultiFiT et en particulier Label Smoothing.