Would you please share the demo for denoising_model training?
youngsuenXMLY opened this issue · 3 comments
First, what brilliant work you have done!
I found your denoising demo without model training application. Would you please share a m-file for denoising model training. Your reply will be highly appreciated! Thank you.
I'll work on it recently :) It's a bit messy now. I'll update the thread once it is done.
Hi ,posenhuang
I mixed 2hours music and speech data at SNR=0 for training, and get satisfactory result. But there is still some problem in internal memory usage. I used a PC with 64GB internal memory, after formulate_data function in formulate_data .m, the internal memory is full, even worse the internal memory swap data with hard disk. I think the main reason is that in formulate_data function, the background music data(variable 'train2') is shift-repeated floor(lenghth(train2)/shift_size ) times, so the
feature data correspondingly increased.
Is there any solution to reduce internal memory usage?
You can try to adjust the shift size to reduce the memory. Also are you using drnn_mini? It is a mini-batch lbfgs. It should use less memory. FYI: I just pushed the training code here (merged to master): #18