Open ongdisheng opened 1 year ago
Have you solved it yet?
Unfortunately no
Unfortunately no
Have you solved it yet?
Nope, I still couldn't figure it out
Hello, I am sorry to hear this news. Do you have the file MOSEI_cmu_audio_CLS.pickle? The http://immortal.multicomp.cs.cmu.edu/raw_datasets/processed_data/ URL posted by the author has expired. Can you share this file with me? good luck!
你好,请问可以分享一下MOSEI_cmu_audio_CLS.pickle吗?官方的这个数据貌似不共享了
好,请问可以分享一下MOSEI_cmu_audio_CLS.pickle吗?官方的这个数据你不分享了
后面我没有找到官方的文件,但是你可以参考同样是清华大学发布的https://drive.google.com/drive/folders/1A2S4pqCHryGmiqnNSPLv7rEg63WvjCSk?usp=sharing中的MOSEI数据集,并修改相应的代码即可进行模型的训练, 我与@ongdisheng最后进行实验的结果相似, Acc-2指标仅有73.7,Acc-7仅有51.8,F1值74.2。
Hi @Kaicheng-Yang0828,
I have tried to run CM-BERT on CMU-MOSEI dataset. However, I only manage to get around 74.15 for acc2 instead of 83.6 as mentioned in https://github.com/thuiar/Cross-Modal-BERT/issues/1#issuecomment-843100902.
Below are the steps that I have taken:
Changed audio pickle file from MOSI_cmu_audio_CLS.pickle to MOSEI_cmu_audio_CLS.pickle
Changed dev, test and train tsv files to hold CMU-MOSEI data
Modified loaded pickle file to MOSEI_cmu_audio_CLS.pickle in _runclassifier.py file![image](https://user-images.githubusercontent.com/63136897/211031271-78e954dc-4be2-41c8-a4de-2c524678c921.png)
Modified dimension of audio modality from 5 to 74 in model.py file![image](https://user-images.githubusercontent.com/63136897/211031558-2b14469c-d69a-4a74-a3ab-caf284e45385.png)
Is there any step that I have done wrongly or missed out? Just wondering is it also possible to have the code files for CMU-MOSEI too?
Your help is very much appreciated. Thanks!