Do you commit the latest code?

[payne4handsome]

@HaozheZhao
Hi, HeozhaZhao, Thanks for your great job at MLLM. I cost a lot of time to run your code MIC. But I find some errors in your code.
First, just like bebow

I can't even find the function save_pred_label implement in dataset.py.
Besides, There two addiational problems.

1. When train with jsonl data format that download from MIC_full and set done_preprocess==False, your code does't work. It prompt IterableDataset hasn't len method.
2. When train with arrow data format I use data_preprocess.py to generate and use Flan-T5 as language model, I got dimension mismatch error. Becase T5's tokenizer.model_max_length is 512, The length of one sample with few-shot is much longer than 512 tokens. Your truncate the input_ids, resulting in image_placeholder (T5's image_placeholder is 图) after 512 tokens in input_ids also truncated. But image num in pixel value not truncated.

So, I think you didn't upload the latest code. Am I right? Or am I wrong about something.

Any suggestions will help me