ZebangCheng

Sorry for the delay in responding. If you have successfully reproduced the results, you can proceed with training, but some detailed steps are involved. First, you need to prepare the...

Sorry, we haven't experimented with multi-GPU inference for large models. During inference, we typically use a single GPU, which requires around 30GB of memory. Your issue might be due to...

Apologies for the confusion. To ensure clarity in the filenames, I have renamed the two files to MERR_coarse_grained.json and MERR_fine_grained.json. Below are the download links: > https://drive.google.com/drive/folders/1LSYMq2G-TaLof5xppyXcIuWiSN0ODwqG?usp=sharing Recently, we have...

> Hello, when running the code, I cannot find the file '/home/user/selected_face/face_emotion/transcription_en_all.csv'. Can you provide this file? [transcription_en_all.csv](https://drive.google.com/file/d/19Pc2YiPnIhcePISUrI313hUYEzXJH41o/view)

The MAE and VideoMAE models we used are not the original pre-trained models; instead, they are model parameters that our team pre-trained using unsupervised mask recovery on the MER2023-SEMI dataset....

Sorry for the delayed response—I was busy recently and couldn’t reply in time. During training, we did not evaluate on a validation set because the dataset we used does not...

MER2024-best这个模型主要用于参加比赛，进行情绪分类任务，所以在后期微调时，只训练了情绪识别任务，情绪推理任务完全没有涉及。建议将MER2024-best这个参数更换为Emotion-LLaMA.pth参数，这个模型参数能够进行情绪推理。

1.编码文件指的是特征吗？我们都开源了相关特征： > https://drive.google.com/drive/folders/1ModyjVKWcWjsacrChDkaTSLpWjnC_lo8?usp=sharing 2.对于mer2024数据中无标签的数据样本，我们是使用MER2024-baseline中的代码，融合7-8个多模态特征，进行情绪类别分类，将分类结果作为伪标签。即我们训练了baseline模型来打伪标签。更具体的，我们先用baseline模型打伪标签用于训练Emotion-LLaMA，再用Emotion-LLaMA打伪标签反过来训练baseline模型，就这样迭代大概两轮后，分数就提升不上去了。 3.我们就是对无标签的两万个样本打伪标签。F1-score为71.37的分数是MER2024-Noise测试集的分数吗？

ok，mer2024中5030个视频样本是训练集，MER2024的训练集的5030个样本来自于MER2023，所以我们没有专门去开源这一部分的特征。这一部分的特征是和MER2023的所有特征一起开源，在以下链接可以找到： > https://drive.google.com/drive/folders/1fudMoAC2IXeInuhAEYbGEES13ungLm_G?usp=sharing MER2024-Noise中那20000个样本没有真实的label，怎么进行评测？对伪标签进行评测？感觉这样有点怪怪的。

抱歉，前段时间我在忙没有及时回复。 1.我查看了你的日志文件log.txt, 你将"image_size"修改为224，我认为这个对模型的性能影响较大; 2.我们没有测试过在24GB上训练模型的经验，要想从头训练，难道较大； 3.MER2024-OV的实验结果是直接用的经过MER2023-SEMI数据集微调后的模型（yes）。我们直接使用Emotion-LLaMA对MER2024-OV的模型进行zore-shot情绪推理，然后从推理的结果中抽取情绪关键词。对于MER2024-OV我们没有做针对性的微调。 有问题可以发邮件联系我，我可以指导一下。我也想知道在24GB上能否完全复现。