I want to use frame-level ssast just for frame-level audio token extraction

[Dayoung, Baik]

in your ast_models.py, you put cluster True as Default

But if to use frame-level ssast, cluster should be False. Do I have to turn it off?

If I want to use your pretrained frame level ssast for audio token extraction, is the output of self.v.norm(x) except the first one what I have to use in finetuningcls function? because the first one is cls token....^^

One more thing I wonder....Could I get some part of fbank that is corresponding to the video frames? melspectrogram does but I don't know fbank could be....

[Yuan Gong]

Hi there,

Thanks for reaching out.

1/

In your ast_models.py, you put cluster True as Default ... But if to use frame-level ssast, cluster should be False. Do I have to turn it off?

You are correct that cluster=True is default in the model script, but we do pass cluster=False for frame-level when we instantiate the model, please see here:

ssast/src/run.py

Lines 125 to 130 in a1a3eec

 if 'pretrain' in args.task: 
     cluster = (args.num_mel_bins != args.fshape) 
     if cluster == True: 
         print('The num_mel_bins {:d} and fshape {:d} are different, not masking a typical time frame, using cluster masking.'.format(args.num_mel_bins, args.fshape)) 
     else: 
         print('The num_mel_bins {:d} and fshape {:d} are same, masking a typical time frame, not using cluster masking.'.format(args.num_mel_bins, args.fshape)) 

FYI, you can use cluster=True for frame-level AST, but from my experience, it will lead to a performance drop.

2/

If I want to use your pretrained frame level ssast for audio token extraction, is the output of self.v.norm(x) except the first one what I have to use in finetuningcls function? because the first one is cls token

You are correct, but please be cautious on that cls_token_num might not always be 1 for all models.

3/

One more thing I wonder....Could I get some part of fbank that is corresponding to the video frames? melspectrogram does but I don't know fbank could be....

This involves audio-visual learning while this paper is about pure audio research. But we do use fbank features as input in this paper, see:

ssast/src/dataloader.py

Lines 126 to 127 in a1a3eec

 fbank = torchaudio.compliance.kaldi.fbank(waveform, htk_compat=True, sample_frequency=sr, use_energy=False, 
                                           window_type='hanning', num_mel_bins=self.melbins, dither=0.0, frame_shift=10) 

Hope these help.

-Yuan

[Dayoung, Baik]

Have you tried other input types like melspectrograms or mfcc? @YuanGongND I am gonna try feeding melspectrograms to SSAST to extract audio feature....Is it okay?
Could I ask you why cls token number is not consistent as 1? It can be 2 because of dist_token? Then, what is dist_token?

[Yuan Gong]

Have you tried other input types like melspectrograms or mfcc? @YuanGongND I am gonna try feeding melspectrograms to SSAST to extract audio feature....Is it okay?

I have never tried other input features. You can pretrain your own model with other input feature, but if you plan to use our pretrained model to extract feature/embedding/token, then you have to use the same dataloader (which is fully released in this repo) with us, any input distribution shift could cause a dramatic performance difference.

Could I ask you why cls token number is not consistent as 1? It can be 2 because of dist_token? Then, what is dist_token?

dist_token stands for distillation token, please read our AST paper for details. SSAST does not need this token, but our code is compatible with old AST models.

-Yuan