Chinese Text2video retrieval support?

[KeyaoZhao]

Thank you for contributing such outstanding work, I would like to ask InternVideo2 support Chinese text search video? What model do I need to replace the VisionEncoder and TextEncoder with? Or how to modify our finetune? Thank you very much
———————————————————————————————————————————————————————
您好，感谢贡献如此杰出的工作，我想请问InternVideo2支持中文文字检索视频吗？我需要把VisionEncoder和TextEncoder换成什么模型呢？或者需要怎么修改我们finetune吗？非常感谢

[leexinhao]

You could use https://huggingface.co/OpenGVLab/InternVideo2-CLIP-1B-224p-f8, it supports Chinese text search!

[KeyaoZhao]

You could use https://huggingface.co/OpenGVLab/InternVideo2-CLIP-1B-224p-f8, it supports Chinese text search!

Thanks for your reply. I use 'InternVideo2-stage2_1b-224p-f4.pt'+'1B_clip.pth' as the vision encoder, 'chinese_alpaca_lora_7b' as tokenizer, 'internvl_c_13b_224px.pth' as the text encoder, but I got the error:

RuntimeError: Error(s) in loading state_dict for InternVideo2_Stage2: size mismatch for text_proj.weight: copying a param with shape torch.Size([512, 1024]) from checkpoint, the shape in current model is torch.Size([512, 768]).

I found the vision feature is Lineard to 512d, but the text feature is 768d and cannot be Lineard to 512d, so how to multiply this two mat? I did something wrong?

[leexinhao]

How do you set the ckpt_path?

, if you set the ckpt of InternVideo2_Stage2 to vision_ckpt_path, it shouldn't meet size mismatch of text_proj.weight.

[zoezhu]

How do you set the ckpt_path? , if you set the ckpt of InternVideo2_Stage2 to vision_ckpt_path, it shouldn't meet size mismatch of text_proj.weight.

Can you explain this with more detail please? I am comfused about how to load the model with these parameters? Which class should I use to initialize the model?

[KeyaoZhao]

How do you set the ckpt_path? , if you set the ckpt of InternVideo2_Stage2 to vision_ckpt_path, it shouldn't meet size mismatch of text_proj.weight.

Thanks, I already solved the mismatch bug. I use the "InternVideo2_clip" to initialize the model but the logger got the following message. And I wonder if I load the model correctly? Because I got the different score answer every time /(ㄒoㄒ)/~~

2024-10-30T09:44:18 | models.internvideo2_clip: Load vision_encoder checkpoint from /root/.cache/huggingface/hub/models--OpenGVLab--InternVideo2-Stage2_1B-224p-f4/snapshots/4362e1f88a992e7edbfd7696f7f78b7f79426dfd/InternVideo2-stage2_1b-224p-f4.pt 2024-10-30T09:44:19 | models.internvideo2_clip: Load text_encoder checkpoint from /workspace/InternVideo/InternVideo/InternVideo2/multi_modality/pretrained/internvl_c_13b_224px.pth 2024-10-30T09:44:34 | models.internvideo2_clip: _IncompatibleKeys(missing_keys=['temp', 'text_encoder.transformer.layers.0.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.0.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.1.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.1.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.2.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.2.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.3.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.3.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.4.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.4.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.5.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.5.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.6.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.6.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.7.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.7.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.8.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.8.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.9.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.9.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.10.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.10.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.11.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.11.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.12.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.12.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.13.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.13.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.14.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.14.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.15.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.15.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.16.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.16.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.17.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.17.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.18.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.18.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.19.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.19.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.20.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.20.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.21.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.21.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.22.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.22.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.23.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.23.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.24.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.24.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.25.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.25.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.26.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.26.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.27.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.27.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.28.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.28.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.29.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.29.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.30.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.30.self_attn.v_proj.base_layer.weight', 'text_encoder.transformer.layers.31.self_attn.q_proj.base_layer.weight', 'text_encoder.transformer.layers.31.self_attn.v_proj.base_layer.weight'], unexpected_keys=['text_encoder.transformer.layers.0.self_attn.q_proj.weight', 'text_encoder.transformer.layers.0.self_attn.v_proj.weight', 'text_encoder.transformer.layers.1.self_attn.q_proj.weight', 'text_encoder.transformer.layers.1.self_attn.v_proj.weight', 'text_encoder.transformer.layers.2.self_attn.q_proj.weight', 'text_encoder.transformer.layers.2.self_attn.v_proj.weight', 'text_encoder.transformer.layers.3.self_attn.q_proj.weight', 'text_encoder.transformer.layers.3.self_attn.v_proj.weight', 'text_encoder.transformer.layers.4.self_attn.q_proj.weight', 'text_encoder.transformer.layers.4.self_attn.v_proj.weight', 'text_encoder.transformer.layers.5.self_attn.q_proj.weight', 'text_encoder.transformer.layers.5.self_attn.v_proj.weight', 'text_encoder.transformer.layers.6.self_attn.q_proj.weight', 'text_encoder.transformer.layers.6.self_attn.v_proj.weight', 'text_encoder.transformer.layers.7.self_attn.q_proj.weight', 'text_encoder.transformer.layers.7.self_attn.v_proj.weight', 'text_encoder.transformer.layers.8.self_attn.q_proj.weight', 'text_encoder.transformer.layers.8.self_attn.v_proj.weight', 'text_encoder.transformer.layers.9.self_attn.q_proj.weight', 'text_encoder.transformer.layers.9.self_attn.v_proj.weight', 'text_encoder.transformer.layers.10.self_attn.q_proj.weight', 'text_encoder.transformer.layers.10.self_attn.v_proj.weight', 'text_encoder.transformer.layers.11.self_attn.q_proj.weight', 'text_encoder.transformer.layers.11.self_attn.v_proj.weight', 'text_encoder.transformer.layers.12.self_attn.q_proj.weight', 'text_encoder.transformer.layers.12.self_attn.v_proj.weight', 'text_encoder.transformer.layers.13.self_attn.q_proj.weight', 'text_encoder.transformer.layers.13.self_attn.v_proj.weight', 'text_encoder.transformer.layers.14.self_attn.q_proj.weight', 'text_encoder.transformer.layers.14.self_attn.v_proj.weight', 'text_encoder.transformer.layers.15.self_attn.q_proj.weight', 'text_encoder.transformer.layers.15.self_attn.v_proj.weight', 'text_encoder.transformer.layers.16.self_attn.q_proj.weight', 'text_encoder.transformer.layers.16.self_attn.v_proj.weight', 'text_encoder.transformer.layers.17.self_attn.q_proj.weight', 'text_encoder.transformer.layers.17.self_attn.v_proj.weight', 'text_encoder.transformer.layers.18.self_attn.q_proj.weight', 'text_encoder.transformer.layers.18.self_attn.v_proj.weight', 'text_encoder.transformer.layers.19.self_attn.q_proj.weight', 'text_encoder.transformer.layers.19.self_attn.v_proj.weight', 'text_encoder.transformer.layers.20.self_attn.q_proj.weight', 'text_encoder.transformer.layers.20.self_attn.v_proj.weight', 'text_encoder.transformer.layers.21.self_attn.q_proj.weight', 'text_encoder.transformer.layers.21.self_attn.v_proj.weight', 'text_encoder.transformer.layers.22.self_attn.q_proj.weight', 'text_encoder.transformer.layers.22.self_attn.v_proj.weight', 'text_encoder.transformer.layers.23.self_attn.q_proj.weight', 'text_encoder.transformer.layers.23.self_attn.v_proj.weight', 'text_encoder.transformer.layers.24.self_attn.q_proj.weight', 'text_encoder.transformer.layers.24.self_attn.v_proj.weight', 'text_encoder.transformer.layers.25.self_attn.q_proj.weight', 'text_encoder.transformer.layers.25.self_attn.v_proj.weight', 'text_encoder.transformer.layers.26.self_attn.q_proj.weight', 'text_encoder.transformer.layers.26.self_attn.v_proj.weight', 'text_encoder.transformer.layers.27.self_attn.q_proj.weight', 'text_encoder.transformer.layers.27.self_attn.v_proj.weight', 'text_encoder.transformer.layers.28.self_attn.q_proj.weight', 'text_encoder.transformer.layers.28.self_attn.v_proj.weight', 'text_encoder.transformer.layers.29.self_attn.q_proj.weight', 'text_encoder.transformer.layers.29.self_attn.v_proj.weight', 'text_encoder.transformer.layers.30.self_attn.q_proj.weight', 'text_encoder.transformer.layers.30.self_attn.v_proj.weight', 'text_encoder.transformer.layers.31.self_attn.q_proj.weight', 'text_encoder.transformer.layers.31.self_attn.v_proj.weight'])