Is it possible if we train the model with another language?

[dangcaptkd2]

Hello, I'm trying to transfer your model on MSCOCO which was translated into Vietnamese. I got a prediction result that is not relative to the input picture although the training process achieve 0.35 on the Bleu_4 score. I used the Vietnamese tokenizer of PhoBert instead and I also changed the version of pytorch_transformers to 1.0.0 because of PhoBert requirement.
Please help me solve this issue, thanks.

[jontooy]

Hi dangcaptkd2,

I am trying to do Image Captioning on the Arabic language and have the same problem! Can you please share how you used PhoBert for training Oscar?

[dangcaptkd2]

Hi jontooy, We simply changed the configuration --model_name_or_path to bert-base-multilingual-uncased and added --tokenizer_name vinai/phobert-base. You need to change the special tokens to adapt your tokenizer (ex: pad_token in Bert-base-multilingual tokenizer is 0 while pad_token in phobert is 1). That's all we changed and it worked!!! Hope this can help you, good luck