def __init__(self, mode, timesteps=128):
self.mode = mode
self.timesteps = timesteps
self.l_conv = lambda label, tid, **kwargs: label_conversion(
label, tid, timesteps=timesteps, **kwargs)
self.mode_mapping = {
"frame": {
"conversion_func": self.get_frame,
"out_classes": 2
},
"frame_onset": {
"conversion_func": self.get_frame_onset,
"out_classes": 3
},
"frame_onset_offset": {
"conversion_func": self.get_frame_onset_offset,
"out_classes": 4
}
}
self._update_mode()
if mode not in self.mode_mapping:
raise ValueError(
f"Available mode: {self.mode_mapping.keys()}. Provided: {mode}"
)
def gen_onsets_info_from_label(label, inst_num=1, t_unit=0.02, mpe=False):
roll = label_conversion(label,
0,
timesteps=len(label),
onsets=True,
mpe=mpe,
ori_feature_size=88,
feature_num=88)
roll = np.where(roll > 0.5, 1, 0)
midi_ch_mapping = sorted([v for v in MusicNetMIDIMapping.values()])
ch = midi_ch_mapping.index(inst_num) + 1
interval, pitches = gen_onsets_info(roll[:, :, ch], t_unit=t_unit)
if len(interval) == 0:
return interval, pitches
off_roll = label_conversion(label,
0,
timesteps=len(label),
mpe=mpe,
ori_feature_size=88,
feature_num=88)
off_roll = np.where(off_roll > 0.5, 1, 0)
for i in range(off_roll[:, :, ch].shape[1]):
notes = find_occur(roll[:, i, ch], t_unit=t_unit)
for nn in notes:
on_idx = nn["onset"]
off_roll[on_idx - 1:on_idx, i, ch] = 0
off_interval, _ = gen_onsets_info(off_roll[:, :, ch], t_unit=t_unit)
if len(interval) != len(off_interval):
l_on = len(interval)
l_off = len(off_interval)
print("WARNING!! Interval length inconsistent. Diff: {}".format(
abs(l_on - l_off)))
print("On len: {}, Off len: {}".format(l_on, l_off))
min_l = min(l_on, l_off)
interval = interval[:min_l]
off_interval = off_interval[:min_l]
interval[:, 1] = off_interval[:, 1]
inconsist = np.where(interval[:, 1] - interval[:, 0] <= 0)
interval[inconsist, 1] += t_unit * 2
return interval, pitches
def gen_frame_info_from_label(label, inst_num=1, t_unit=0.02, mpe=False):
roll = label_conversion(label,
0,
timesteps=len(label),
mpe=mpe,
ori_feature_size=88,
feature_num=88)
midi_ch_mapping = sorted([v for v in MusicNetMIDIMapping.values()])
ch = midi_ch_mapping.index(inst_num) + 1
return gen_frame_info(roll[:, :, ch], t_unit=t_unit)
def __init__(self,
mode:str,
timesteps=128):
self.mode = mode
self.timesteps = timesteps
self.l_conv = lambda label, tid, **kwargs: label_conversion(label, tid, timesteps=timesteps, **kwargs)
if mode == "frame":
self.conversion_func = self.get_frame
self.out_classes = 2
elif mode == "frame_onset":
self.conversion_func = self.get_frame_onset
self.out_classes = 3
elif mode == "frame_onset_offset":
self.conversion_func = self.get_frame_onset_offset
self.out_classes = 4
elif self.customized_mode(mode):
pass
else:
raise ValueError("Available mode: 'frame', 'frame_onset', 'frame_onset_offset'. Provided: {}".format(mode))
def FullTest(model_path,
test_path,
label_path=None,
pred_save_path="./predictions",
use_ram=True,
MAX_FRAME=1800):
# Load files
print("Loading files")
features = parse_path(test_path)
for ff in features:
if not ff.endswith(".hdf"):
idx = features.index(ff)
del features[idx]
if label_path is not None:
# Assume there are exactly label files corresponding to the test audios
#labels = parse_path(label_path, label=True)
labels = []
for ff in features:
ext = ff[ff.rfind("."):]
if ext != ".hdf" and ext != ".pickle":
continue
ll = ff[(ff.rfind("/") + 1):]
if "_label" not in ll:
ll = ll[:ll.rfind(".")] + "_label.pickle"
labels.append(os.path.join(label_path, ll))
labels = load_files(labels, use_ram=use_ram)
features = load_files(features, use_ram=use_ram)
model = load_model(model_path)
# Validate on model/feature configurations
f_type, channels, out_classes, timesteps = model_info(model_path)
key = list(features.keys())
if f_type == "HCFP" and features[key[0]].shape[2] < 12:
assert (
False
), "The model uses HCFP as input feature, but loaded features are not."
if f_type == "CFP" and features[key[0]].shape[2] == 12:
assert (len(channels) == 2 and 1 in channels
and 3 in channels), """The
The given feature are HCFP, but the model uses more feature types.
Model input feature types: """ + str(
channels) + " ({0: Z, 1: Spec, 2: GCoS, 3: Ceps})"
channels = [0, 6]
mpe = False
if out_classes == 2:
mpe = True
# To avoid running out of memory.
# 9000 is suitable for 32G RAM with one instrument only and all 4 channels used. (Max ram usage almost 100%)
#MAX_FRAME = 1800
print("Max frame per prediction: ", MAX_FRAME)
# Start to predict
pred_out = h5py.File(os.path.join(pred_save_path, "pred.hdf"), "w")
label_out = h5py.File(os.path.join(pred_save_path, "label.hdf"), "w")
len_data = len(features)
for idx in trange(len_data, desc='Dataset'):
i = key[idx]
feature = features[i][:]
pred = predict(feature,
model,
MAX_FRAME=MAX_FRAME,
channels=list(channels),
instruments=out_classes - 1,
timesteps=timesteps)
# Save to output
pred_out.create_dataset(str(i),
data=pred,
compression="gzip",
compression_opts=5)
del feature, features[i]
# Process corresponding label
if label_path is not None:
ll = labels[0]
if type(ll) != np.ndarray:
ll = label_conversion(ll, 352, 128, mpe=mpe)[:, :, 1:]
label_out.create_dataset(str(i),
data=ll,
compression="gzip",
compression_opts=5)
del labels[0]
pred_out.close()
label_out.close()
def gen_onsets_info_from_label(label, inst_num=1, t_unit=0.02):
roll = label_conversion(label, 0, timesteps=len(label), onsets=True, ori_feature_size=88, feature_num=88)
roll = np.where(roll>0.5, 1, 0)
midi_ch_mapping = sorted([v for v in MusicNetMIDIMapping.values()])
ch = midi_ch_mapping.index(inst_num)+1
return gen_onsets_info(roll[:,:,ch], t_unit=t_unit)