def EstimateChord(idx, dnnmodel, todir=False):
#dnn = networks.FeatureDNN()
#dnn = networks.ConvnetFeatExtractor()
dnn = networks.FullCNNFeatExtractor()
#dnn = networks.NoOperation()
dnn.load(dnnmodel)
dnn.to_gpu(0)
decoder = networks.NBLSTMCRF()
decoder.load()
decoder.to_gpu(0)
cqtfilelist = np.array(find_files(const.PATH_HCQT, ext="npy"))[idx]
i = 0
chainer.config.train = False
chainer.config.enable_backprop = False
for cqtfile in cqtfilelist:
cqt = utils.Embed(utils.PreprocessSpec(np.load(cqtfile)[:, :, :]), 1)
chroma = dnn.GetFeature(cp.asarray(cqt)).data
path = decoder.argmax(chroma)
feat = cp.asnumpy(chroma)
if todir:
fname = cqtfile.split("/")[-1] + ".lab"
alb = cqtfile.split("/")[-2]
utils.SaveEstimatedLabelsFramewise(
path, const.PATH_ESTIMATE_CROSS + alb + "/" + fname, feat)
else:
utils.SaveEstimatedLabelsFramewise(
path, const.PATH_ESTIMATE + "%03d.lab" % i, feat)
i += 1
type=str, default="nblstm_crf.model", action="store")
args = parser.parse_args()
audio_list = find_files("Datas/audios_estimation")
for audiofile in audio_list:
fname = audiofile.split("/")[-1]
print("Processing: %s" % fname)
#load audio
y,sr = load(audiofile,sr=C.SR)
#extract Harmonic-CQT from audio
fmin = note_to_hz("C1")
hcqt = np.stack([np.abs(cqt(y,sr=C.SR,hop_length=C.H,n_bins=C.BIN_CNT,bins_per_octave=C.OCT_BIN,fmin=fmin*(h+1),filter_scale=2,tuning=None)).T.astype(np.float32) for h in range(C.CQT_H)])
#extract feature using trained CNN extractor
cnn_feat_extractor = N.FullCNNFeatExtractor()
cnn_feat_extractor.load(args.f)
feat = cnn_feat_extractor.GetFeature(U.PreprocessSpec(hcqt)).data
#decode label sequence
decoder = N.NBLSTMCRF()
decoder.load(args.d)
labels = decoder.argmax(feat)
#convert into .lab file
labfile = os.path.join("Datas/labs_estimated",fname+".lab")
U.SaveEstimatedLabelsFramewise(labels,labfile,feat)
def TrainNStepCRF(idx,
epoch=20,
augment=0,
featmodel=const.DEFAULT_CONVNETFILE,
path_blstm="blstm.model",
savefile="nblstm_crf.model"):
cqtfilelist = np.array(find_files(const.PATH_HCQT, ext="npy"))
chordlablist = np.array(
find_files(const.PATH_CHORDLAB, ext=["lab", "chords"]))
if idx is not None:
cqtfilelist = cqtfilelist[idx]
chordlablist = chordlablist[idx]
chainer.config.train = False
chainer.config.enable_backprop = False
#dnn = networks.TripleDNNExtractor()
#dnn = networks.FeatureDNN()
dnn = networks.FullCNNFeatExtractor()
#dnn = networks.NoOperation()
#dnn = networks.ConvnetFeatExtractor()
dnn.load(featmodel)
dnn.to_gpu(0)
rnn = networks.NBLSTMCRF()
rnn.blstm.load(path_blstm)
rnn.to_gpu(0)
opt = optimizers.MomentumSGD()
opt.setup(rnn)
#opt.add_hook(optimizer.WeightDecay(0.001))
X = []
T = []
for cqtfile, labfile in zip(cqtfilelist, chordlablist):
cqt = utils.Embed(utils.PreprocessSpec(np.load(cqtfile)[:, :, :]), 1)
feature = cp.asnumpy(dnn.GetFeature(cp.asarray(cqt)).data)
lab = utils.LoadLabelArr(labfile)
min_sz = min([feature.shape[0], lab.shape[0]])
X.append(feature[:min_sz, :])
T.append(lab[:min_sz])
sizes = np.array([x.shape[0] for x in X], dtype="int32")
print("start epoch:")
chainer.config.train = False
chainer.config.enable_backprop = True
last_loss = np.inf
for ep in range(epoch):
sum_loss = 0.0
rand_songid = np.random.randint(len(X),
size=np.sum(sizes) //
const.DECODER_TRAIN_SEQLEN * 8)
for i in range(0, rand_songid.size, const.DECODER_TRAIN_BATCH):
xbatch = []
tbatch = []
for songid in rand_songid[i:i + const.DECODER_TRAIN_BATCH]:
seq_len = sizes[songid]
idx = np.random.randint(seq_len - const.DECODER_TRAIN_SEQLEN -
1)
x_snip = X[songid][idx:idx + const.DECODER_TRAIN_SEQLEN, :]
t_snip = T[songid][idx:idx + const.DECODER_TRAIN_SEQLEN]
if augment > 0:
shift = np.random.randint(augment)
x_snip, t_snip = shift_data(x_snip, t_snip, shift)
xbatch.append(Variable(cp.asarray(x_snip)))
tbatch.append(Variable(cp.asarray(t_snip)))
rnn.cleargrads()
opt.update(rnn, xbatch, tbatch)
sum_loss += rnn.loss.data
print("epoch %d/%d loss=%.3f" % (ep + 1, epoch, sum_loss / 12800.0))
rnn.save(savefile)