WaveUNet

implement Wave-U-Net by pytorch

result

Wave-U-Net	my improved version
SDR=11.83	SDR=12.51

Train network

• if you just want to train the model, use commandTrain.py

python commandTrain.py --dataset both 
##(for both ccmixter and musdb18)
python commandTrain.py --dataset ccmixter 
##(for ccmixter only)
python commandTrain.py --dataset musdb18 
##(for musdb18 only)

• If you want to change the neural network model
• from modelStruct.pyramidnet import Unet[1]
• from modelStruct.unet import Unet[2])
• you can choose one between these two.

Inference

• If you have one or more mix songs, you can use inference.py to predict accompaniments by using these mix songs and a saved model.

python inference.py --checkpoint pyramid --test_number 1
##(please type checkpoints name and number of test songs)
##(please name test songs as 0.wav, 1.wav, 2.wav etc)
##(please put test songs in folder ccmixter2/x)

Dataset

• please put mix songs in folder ccmixter2/x
• please put accompaniments in folder ccmixter2/y
• please put vocal songs in folder ccmixter2/z
• first 50 songs are ccmixter and last 150 songs are musdb18.
• please name songs as 0.wav, 1.wav, 2.wav etc in folder ccmixter2/x, ccmixter2/y and ccmixter2/z respectively.
• if you only use ccmixter, you should have 0.wav, 2.wav, to 49.wav.
• if you only use musdb18, you should have 50.wav, 51.wav, to 199.wav.
• if you want to use both ccmixter and musdb18, you should have 0.wav, 1.wav, 2.wav, 3.wav, to 199.wav.
• all Audio rates I read are 16000 and Mono.
• I use all ccmixter songs and musdb18 songs, which includes 200 songs.
• training_set = Dataset(np.arange(150), 'ccmixter2/')
• test_set = Testset(np.arange(140,160), 'ccmixter2/')
• validation_set =Valtset(np.arange(150,200), 'ccmixter2/')
• as shown here, I use first 150 songs as training set, last 50 songs as validation set(to visualize loss)
• I will also write results(from 140th songs to 159th songs, which includes training set and validation set) generated from network to folder vsCorpus.

Installment

• pytorch 0.4
• tensorboardX (using tensorboard with pytorch, if you do not want to use tensorboard, set USEBOARD as False)
• soundfile
• h5py
• numpy

Describe files

Different start points

• trainForRandomGen.py (use ccmixter and musdb as dataset to train model)
• trainchinese.py (use chinese songs as dataset to train model)
• trainclassify.py (use classification instead regression, classification can also generalize as good as regression but much more noise)

Tools

• transformData.py (same as utils file)

Read Dataset

• readccmu.py (read ccmixter and musdb18)
• readchinese.py (read 20000 songs)
• readpiano.py (read piano songs which is download from youtube to train wavenet, but now it is useless)

Model structure(all in folder modelStruct)

• pyramidnet.py(in the middle of nework, use different dilation rate filters to extract features, learned from deep lab series)
• quanunet.py(use softmax as loss fuction)
• randomunet.py(my experiment, use random dilation rate, which is inspired by [3])
• unet.py(use classical wave-u-net[2])
• unetd.py(use wave-u-net with dilation filters)
• resunet.py(wanna combine unet and resnet)

Reference

• [1]. https://arxiv.org/pdf/1606.00915.pdf
• [2]. https://qmro.qmul.ac.uk/xmlui/bitstream/handle/123456789/39785/Stoller%20Wave-U-Net%202018%20Accepted.pdf?sequence=1
• [3]. https://arxiv.org/abs/1808.03578