Condition in the FM training is always zeros ?

FunAudioLLM / CosyVoice

Multi-lingual large voice generation model, providing inference, training and deployment full-stack ability.

https://funaudiollm.github.io/

Apache License 2.0

4.6k stars 461 forks source link

Condition in the FM training is always zeros ? #79

Open BridgetteSong opened 2 months ago

BridgetteSong commented 2 months ago

in the Flow forward, conds is always zeros. It is different from paper randomly drop the conditions with 0.2 probability.

Can you explain why and how to mask conds?

https://github.com/FunAudioLLM/CosyVoice/blob/6a3e44242ad24f01ba64430d8f6ac5718442b0da/cosyvoice/flow/flow.py#L78-L91

aluminumbox commented 2 months ago

in training we also have code to randomly drop condition, we are still testing to determine what trick is useful and suitable for opensource, we will update it later

hongwen-sun commented 2 months ago

催更

hongwen-sun commented 2 months ago

@aluminumbox 请问这个问题后续会更新吗