Question about the embed_positions.weight in Whisper model

[bobqianic]

In the encoder of Whisper models, the encoder.embed_positions.weight was initialized using sine and cosine functions of different frequencies. However, in the decoder of Whisper models, the decoder.embed_positions.weight consists of trainable parameters that are initialized using random noise. But why is this the case?

encoder.embed_positions.weight

decoder.embed_positions.weight

[bobqianic]

Maybe...

[ggerganov]

Haha yes, it's voodoo magic