Open CrushDemo01 opened 4 months ago
不是多张相同的图片,数据集是多张不同的图片,具体的代码可以参考这个我修改的这个repo: https://github.com/Chengru-Song/LLaVA/tree/feature/multi_image_training
数据集需要使用多张不同的图片,不过可以Follow他们最新的工作:interleaved LLaVA,他们应该已经实现了多图的训练。不过感觉方法是类似的 https://llava-vl.github.io/blog/2024-06-16-llava-next-interleave/
[ { "conversations": [ {"from": "human", "value": "Please tell me what's unusual about this image: <image1><image2>"}, {"from": "gpt", "value": "A man is ironing his clothes on a vehicle. "}, {"from": "human", "value": "What's funny about this?"}, {"from": "gpt", "value": "Because people don't usually do this at home."} ], "image": "llava/image_folder(local image path)" }, { "conversations": [ {"from": "human", "value": "Please tell me what's unusual about this image: <image1><image2>"}, {"from": "gpt", "value": "A man is ironing his clothes on a vehicle. "}, {"from": "human", "value": "What's funny about this?"}, {"from": "gpt", "value": "Because people don't usually do this at home."} ], "image": "llava/image_folder(local image path)" } ]
谢谢您的回答。
LLaVA-interleaved我看了一点点,他们的构建的数据集类似也是两个不同的image token,就像上面这样。感觉会合理一些?但我看您的数据集样例还是只有一个
另外您的博客一直无法直接评论,如图 再次感谢您的回答!
谢谢你的指出,我刚支持了博客的comment功能。
prompt里面的
但我不认为他们会根据位置不一样设置不同的image_token_id,原因在于
您好,请教您一下,您对训练数据json修改的方式的是否是多张相同的图片? 因为我看只有一个。
如果要实现两个不同的图片,是不是构建数据集应该要和 这种形式?
ps:纯小白,刚开始看MLLM,像迁移到自己的任务上。见谅!