happycolor / speechformer2 Goto Github PK

Thank you for your research. I am trying to reproduce your code, but I encountered a problem. How did you cut the speech in the DAIC-WOZ database? Can you provide the relevant data preprocessing code?

Thank you very much for your sharing. What are the advantages of the v2 version over the V1 version?

Did you fill in the name and label of the self created. csv file you mentioned in readme? What do they refer to respectively

You don't seem to provide the meta_csv_file in the feature_config.json file of each dataset, how did you generate it?

Could you provide me with the well trained model file for Alzheimer's disease detection (Pitt) tasks and code for inference? Many thanks!

Hi,

First of all thank you for sharing your code! I would like to use SpeechFormer on a smaller dataset and I was wondering whether you could share the weights of a training you did on one of the datasets. Let me know if that would be easy!

Hello,
You have defined a method as lmdb, and it is an additional method in your code, can you describe how can we run the train_model.py and pass the lmdb to train model based on spec features and transfromer2
Thanks

Could you please provide some information about file: meta_info.pkl? What's "shape"?
Thank you!

I've successfully run 'train_model.py' with the epoch set to 30, and the results have been printed out as expected. In addition, a new folder labeled 'exp' has been created.
The problem arises with the saving of the trained model and checkpoint. It should be saved in 'exp/SpeechFormer++/datasetname_e30_b32_lr0.0005_hubert12_L2244_expa111/folder_1/checkpoint', however, this directory is empty.

Could you please help me solve this problem? Thank you!