andabi / music-source-separation

Deep neural networks for separating singing voice from music written in TensorFlow
796 stars 149 forks source link

训练loss一直居高不下 #44

Open suntingfeng opened 5 years ago

suntingfeng commented 5 years ago

你好,我使用MIR-1k中Wavfile作为训练集,使用UndividedWavfile作为验证集,训练的epoch为100000,但loss一直居高不下,稳定在3.0左右,请问这是为什么呢?我应该修改哪些配置呢? 谢谢!

ucasiggcas commented 5 years ago

不需要修改。请直接看测试结果。你会惊讶的。