def TrainDNNExtractor(trainidx, epoch=20, saveas="dnn.model"):
cqtfilelist = np.array(find_files(const.PATH_MIDIHCQT,
ext="npz"))[trainidx]
#midifilelist = find_files(const.PATH_MIDI,ext="mid")[:filecnt]
filecnt = cqtfilelist.size
chainer.config.train = True
chainer.config.enable_backprop = True
dnn = networks.FeatureDNN()
model = networks.DNNModel(dnn)
model.to_gpu(0)
opt = optimizers.MomentumSGD()
opt.setup(model)
spl = np.arange(0, filecnt, 2000)
if spl[-1] < filecnt:
spl = np.append(spl, filecnt)
print("split count:%d" % (spl.size - 1))
print(spl)
print("start epochs...")
S = []
T = []
for cqtfile in cqtfilelist:
dat = np.load(cqtfile)
spec = utils.PreprocessSpec(dat["spec"][:, :])
targ = GetConvnetTargetFromPianoroll(dat["target"]).astype(np.int32)
assert (spec.shape[0] == targ.shape[0])
S.append(spec)
T.append(targ)
S = np.concatenate(S, axis=0)
T = np.concatenate(T, axis=0)
for ep in range(epoch):
sum_loss = 0
assert (S.shape[0] == T.shape[0])
randidx = np.random.permutation(S.shape[0])
for i in range(0, randidx.size, const.CONV_TRAIN_BATCH):
x_batch = S[randidx[i:i + const.CONV_TRAIN_BATCH], :]
t_batch = T[randidx[i:i + const.CONV_TRAIN_BATCH], :]
x_in = cp.asarray(x_batch)
t_in = cp.asarray(t_batch)
model.cleargrads()
loss = model(x_in, t_in)
loss.backward()
opt.update()
sum_loss += loss.data * const.CONV_TRAIN_BATCH
dnn.save(saveas)
print("epoch: %d/%d loss:%.04f" %
(ep + 1, epoch, sum_loss / const.CONV_TRAIN_BATCH))
def EvaluateConvnet(modelfile, cqtfilelist):
predicted_chroma = []
target_chroma = []
predicted_bass = []
target_bass = []
predicted_top = []
target_top = []
#dnn = networks.ConvnetFeatExtractor()
dnn = networks.FeatureDNN()
#dnn = networks.FullCNNFeatExtractor()
dnn.load(modelfile)
dnn.to_gpu(0)
chainer.config.train = False
chainer.config.enable_backprop = False
for cqtfile in cqtfilelist:
#print(cqtfile)
dat = np.load(cqtfile)
cqt = utils.PreprocessSpec(dat["spec"][:, :])
t = chromatemplate.GetConvnetTargetFromPianoroll(dat["target"])
#assert(cqt.shape[0]==t.shape[0])
out = cp.asnumpy(dnn.GetFeature(cp.asarray(cqt)).data)
#out = dnn.GetFeature(cqt).data
out_len = out.shape[0]
pred_chroma = np.zeros((out_len, 12), dtype="int32")
pred_bass = np.zeros((out_len, 12), dtype="int32")
pred_top = np.zeros((out_len, 12), dtype="int32")
pred_chroma[out[:, 12:24] > 0.5] = 1
pred_bass[np.arange(out_len), np.argmax(out[:, :12], axis=1)] = 1
#pred_bass[out[:,:12]<=0.5] = 0
pred_top[np.arange(out_len), np.argmax(out[:, 24:36], axis=1)] = 1
#pred_top[out[:,24:36]<=0.5] = 0
predicted_chroma.append(pred_chroma[:, :].astype(np.bool))
target_chroma.append(t[:out_len, 12:24].astype(np.bool))
predicted_bass.append(pred_bass[:out_len, :].astype(np.bool))
target_bass.append(t[:out_len, :12].astype(np.bool))
predicted_top.append(pred_top[:out_len, :].astype(np.bool))
target_top.append(t[:out_len, 24:36].astype(np.bool))
eval_chroma = evaluate(predicted_chroma, target_chroma)
eval_bass = evaluate(predicted_bass, target_bass)
eval_top = evaluate(predicted_top, target_top)
return eval_chroma, eval_bass, eval_top
def TrainTranscribeDNNChord(idx, epoch=20, saveas="dnn_deepchroma.model"):
cqtfilelist = np.array(find_files(const.PATH_HCQT, ext="npy"))[idx]
chordlablist = np.array(
find_files(const.PATH_CHORDLAB, ext=["lab", "chords"]))[idx]
featurelist = []
targetlist = []
chainer.config.train = True
chainer.config.enable_backprop = True
for cqtfile, labfile in zip(cqtfilelist, chordlablist):
cqt = np.load(cqtfile)[0, :, :]
chroma = voc.LoadChromaTarget(labfile)
min_sz = min([cqt.shape[0], chroma.shape[0]])
cqt = utils.Embed(utils.PreprocessSpec(cqt[:min_sz]), size=7)
chroma = chroma[:min_sz]
featurelist.append(cqt)
targetlist.append(chroma.astype(np.int32))
featurelist = np.concatenate(featurelist)
targetlist = np.concatenate(targetlist)
itemcnt = targetlist.shape[0]
print("DNN Training begin...")
dnn = networks.FeatureDNN()
dnn.train = True
model = networks.DNNModel(predictor=dnn)
model.to_gpu()
opt = optimizers.AdaDelta()
opt.setup(model)
for ep in range(epoch):
randidx = np.random.permutation(itemcnt)
sumloss = 0.0
for i in range(0, itemcnt, const.DNN_TRAIN_BATCH):
X = cp.asarray(featurelist[randidx[i:i + const.DNN_TRAIN_BATCH]])
T = cp.asarray(targetlist[randidx[i:i + const.DNN_TRAIN_BATCH]])
opt.update(model, X, T)
sumloss += model.loss.data * const.DNN_TRAIN_BATCH
print("epoch %d/%d loss=%.3f" % (ep + 1, epoch, sumloss / itemcnt))
dnn.save(saveas)
n_bins=144,
bins_per_octave=24,
filter_scale=2,
tuning=None)).T.astype(np.float32)),
size=1)
#dat = np.load("/media/wuyiming/TOSHIBA EXT/midihcqt_12/000005.npy")
#dat_24 = np.load("/media/wuyiming/TOSHIBA EXT/midihcqt_24/000005.npz")
#spec_dnn = U.Embed(U.PreprocessSpec(dat_24["spec"]),size=7)
spec = spec[:, :250, :]
spec_dnn = spec_dnn[:250, :]
cnn = networks.FullCNNFeatExtractor()
cnn.load("fullcnn_crossentropy_6000.model")
deepchroma = networks.FeatureDNN()
deepchroma.load(
"/home/wuyiming/Projects/TranscriptionChordRecognition/dnn3500.model")
chroma_cnn = cnn.GetFeature(spec).data[:, 12:24].T
chroma_dnn = deepchroma.GetFeature(spec_dnn).data[:, 12:24].T
chroma = np.log(
1 + chroma_cqt(wav, sr=C.SR, hop_length=C.H, bins_per_octave=24)[:, :250])
target = chromatemplate.GetConvnetTargetFromPianoroll(
U.GetPianoroll(
"/media/wuyiming/TOSHIBA EXT/AIST.RWC-MDB-P-2001.SMF_SYNC/RM-P051.SMF_SYNC.MID"
))
target = target[10:260, 12:24].T
plt.subplot(4, 1, 1)