downstream input and output about s3prl

[andi611]

Hi,

the shape of the label is (batch_size, seq_len),
the shape of the feature is (batch_size, seq_len, feature_dim).

maybe 1544 means the loaded feature is consist of 1544 phones(after zero padding?)

The 1544 is the sequence length of that particular phone sequence,
and also the sequence length of that particular feature sequence.

and each 160-dim feature represents one phone.

Each 160-dim acoustic feature may NOT represent one phone, since each 160-dim is just a single acoustic frame, a phone usually spans over several frames.

After the upstream model, it becomes a [12,1544,768] tensor which is the output of transformer.

Yes.

[liu-x-p]

Thank you for your reply. So does the seq_len mean the number of frames?

[andi611]

Thank you for your reply. So does the seq_len mean the number of frames?

Yes!

[liu-x-p]

Thanks! Got it!