yangangcao / trunet Goto Github PK

It will be good for the researchers / students if you share the working version of training script

我最近在使用 TRUNet 进行训练，和文中不同，为了避免相位估计的问题，我直接用复数谱的实部和虚部拼接作为输入特征。也即是输入特征维度为（T, 2, F），模型输出对实部和虚部的mask估计，虽然最终的性能正常，但是在训练过程中，train loss 下降很平稳， 验证 loss 震荡相当严重，增大了batch或者调小了学习率后也没有改善，不知道你在训练的时候遇到过这种情况吗？

T-GRU和F-GRU的输入shape好像是一样的，所以这里怎么提现这个T-GRU？

The shape of the output is [1, 5, 257], 1 for the batch, 257 for the frequency. Could you please tell me what 5 means? Thank you so much.

the model input shape is 【1，4，257】，while the output shape is 【1，5，257】，

Hi

As per the paper, 4 features must be concatenated as input to TRUNET,

1. log spectrum
2. PCEN
3. real part of demodulated phase
4. imaginary part of demodulated phase

so the input will become (Batchsize, 4 features, No.of frames in STFT, No.of STFT bins) , so it is a 4 dimesional one

But in the sample code you are showing input as 3 dimension (1,4,257), since first layer is conv1d

I'm confused whether the input to TRUNET is 3 dimension or 4 dimension ?

Regards
Yugesh