Yue Cao

> > > I have just downloaded model htc++_beit_adapter_large_fpn_3x_coco.pth and config from this github. But I cannot load model use this command: > > > from mmdet.apis import init_detector configFile...

Thank you for your feedback. You can carefully check whether the compiled script make.sh executed successfully and did not report any errors. Or would it be convenient for you to...

Thank you for your answer. How should the single-branch convnext in your paper be trained? Because I tried to use your code to train single-branch convnext, the loss effect in...

Hi, Please try set `"output_hidden_states": true,` in [config.json](https://huggingface.co/OpenGVLab/InternVL2_5-78B/blob/main/config.json#L65).

你好， 对于longvideobench，我们报告了其16, 32, 48和64帧的最优结果。对于8B你可以[尝试设置frame为48](https://github.com/open-compass/VLMEvalKit/blob/6c5f81cbb50ff780f33799f708c1a76bb26efa7f/vlmeval/dataset/video_dataset_config.py#L39)以测试其结果。 参考[技术报告](https://arxiv.org/pdf/2412.05271)中longvideobench介绍段： > We test four settings—16, 32, 48, and 64 frames—and report the best results on the validation set.

你好，请问是否有尝试过transformers 3.37.2版本？

你好， 请问是lora训练吗？lora训练后需要[合并权重](https://internvl.readthedocs.io/en/latest/tutorials/coco_caption_finetune.html#merging-lora-weights)

Hello, As stated in the original text, we force the model to output the final answer in the form of 'Final Answer: xxx' at the end. So for the final...

你好， 当前代码支持每一步保存模型权重，可以参考参数save_steps。并且支持resume_ckpt，可以参考训练代码[1097行](https://github.com/OpenGVLab/InternVL/blob/34a81000402bf8f716bab8c9b57aff1f6b436bd0/internvl_chat/internvl/train/internvl_chat_pretrain.py#L1097)。 这是Trainer实现的功能，会从对应step继续向后训练，代码seed前后一致可以避免数据重复训练。

你好， 感谢你的提问，请关注 [llava/model/multimodal_projector](https://github.com/yuecao0119/MMFuser/tree/main/llava/model/multimodal_projector) 这个文件夹，主要模型修改内容主要在这个里面。