try:
max_len = float(args.max_len)
section = [0, max_len]
except:
max_len = None
section = None
print("Reading MIDI files")
prs = []
for file in glob.glob(os.path.join(args.data_path, "*.mid")):
if args.step == "beat":
data = DataMapsBeats()
data.make_from_file(file,
args.beat_gt,
args.beat_subdiv,
section,
acoustic_model=args.acoustic_model)
else:
data = DataMaps()
data.make_from_file(file,
args.step,
section,
acoustic_model=args.acoustic_model)
prs.append(data.target)
print("Calculating priors and transition probabilities")
# import matplotlib.pyplot as plt
# plt.imshow(mask, aspect='auto')
# plt.show(block=[bool])
if args.step in ['quant', 'quant_short', 'event']:
data_quant = DataMaps()
data_quant.make_from_file(filename_target,
args.step, [0, 30],
acoustic_model='kelz')
input_roll = convert_note_to_time(input_roll,
data_quant.corresp,
data_quant.input_fs,
max_len=30)
if args.step == 'beat':
data_quant = DataMapsBeats()
data_quant.make_from_file(filename_target,
args.beat_gt,
args.beat_subdiv, [0, 30],
acoustic_model='kelz')
input_roll = convert_note_to_time(input_roll,
data_quant.corresp,
data_quant.input_fs,
max_len=30)
if args.step == 'time' and args.with_quant:
data_quant = DataMaps()
data_quant.make_from_file(filename_target,
'quant', [0, 30],
acoustic_model='kelz')
input_roll = align_matrix(input_roll,
data_quant.corresp,
def weight_search(params, num=0, verbose=False):
global global_params
print(params)
sys.stdout.flush()
# Parse params
min_diff = params[0]
history = int(params[1])
num_layers = int(params[2])
is_weight = params[3]
features = params[4]
warnings.filterwarnings("ignore", message="tick should be an int.")
max_len = 30
section = [0, max_len]
# Load model
model = model_dict['model']
sess = model_dict['sess']
# Get weight_model data
pkl = data_dict['blending_data']
X = pkl['X']
Y = pkl['Y']
D = pkl['D']
max_history = pkl['history']
features_available = pkl['features']
with_onsets = pkl['with_onsets']
# Filter data for min_diff
X, Y = filter_data_by_min_diff(
X, Y,
np.maximum(D[:, 0], D[:, 1]) if with_onsets else D, min_diff)
if len(X) == 0:
print("No training data generated.")
sys.stdout.flush()
return 0.0
# Filter X for desired input fields
X = filter_X_features(X, history, max_history, features,
features_available, with_onsets)
# Ablate X
X = ablate(X, global_params['ablate'], with_onsets=with_onsets)
history = min(history, max_history)
if features and not features_available:
features = False
# Train weight model
print("Training weight model")
sys.stdout.flush()
layers = []
for i in range(num_layers):
layers.append(10 if with_onsets else 5)
weight_model = train_model(X,
Y,
layers=layers,
weight=is_weight,
with_onsets=with_onsets)
# Save model
global most_recent_model
most_recent_model = {
'model': weight_model,
'history': history,
'features': features,
'weight': is_weight,
'with_onsets': with_onsets,
'ablate': global_params['ablate']
}
weight_model_name = get_filename(min_diff, history, num_layers, features,
with_onsets, is_weight,
global_params['step'])
# Write out weight model
with open(os.path.join(global_params['model_out'], weight_model_name),
"wb") as file:
pickle.dump(most_recent_model, file)
# Evaluation
results = {}
frames = np.zeros((0, 3))
notes = np.zeros((0, 3))
for filename in sorted(glob.glob(os.path.join(data_dict['valid'],
"*.mid"))):
print(filename)
sys.stdout.flush()
if global_params['step'] == 'beat':
data = DataMapsBeats()
data.make_from_file(filename,
global_params['beat_gt'],
global_params['beat_subdiv'],
section,
acoustic_model=global_params['acoustic'],
with_onsets=with_onsets)
else:
data = DataMaps()
data.make_from_file(filename,
global_params['step'],
section,
acoustic_model=global_params['acoustic'],
with_onsets=with_onsets)
# Decode
input_data = data.input
if with_onsets:
input_data = np.zeros(
(data.input.shape[0] * 2, data.input.shape[1]))
input_data[:data.input.shape[0], :] = data.input[:, :, 0]
input_data[data.input.shape[0]:, :] = data.input[:, :, 1]
# Add noise
input_data = add_noise_to_input_data(input_data, data_dict['noise'],
data_dict['noise_gauss'])
pr, priors, weights, combined_priors = decode(
input_data,
model,
sess,
branch_factor=5,
beam_size=50,
weight=[[0.8], [0.2]],
out=None,
hash_length=12,
weight_model=weight_model,
verbose=verbose,
weight_model_dict=most_recent_model)
# Evaluate
if with_onsets:
target_data = pm.PrettyMIDI(filename)
corresp = data.corresp
[P_f, R_f,
F_f], [P_n, R_n, F_n
], _, _ = compute_eval_metrics_with_onset(pr,
corresp,
target_data,
double_roll=True,
min_dur=0.05,
section=section)
else:
if global_params['step'] in [
'quant', 'event', 'quant_short', 'beat'
]:
pr = convert_note_to_time(pr,
data.corresp,
data.input_fs,
max_len=max_len)
data = DataMaps()
if global_params['step'] == "20ms" or with_onsets:
data.make_from_file(filename,
"20ms",
section=section,
with_onsets=False,
acoustic_model="kelz")
else:
data.make_from_file(filename,
"time",
section=section,
with_onsets=False,
acoustic_model="kelz")
target = data.target
#Evaluate
P_f, R_f, F_f = compute_eval_metrics_frame(pr, target)
P_n, R_n, F_n = compute_eval_metrics_note(pr, target, min_dur=0.05)
print(
f"Frame P,R,F: {P_f:.3f},{R_f:.3f},{F_f:.3f}, Note P,R,F: {P_n:.3f},{R_n:.3f},{F_n:.3f}"
)
sys.stdout.flush()
frames = np.vstack((frames, [P_f, R_f, F_f]))
notes = np.vstack((notes, [P_n, R_n, F_n]))
if F_n < global_params['early_exit']:
print("Early stopping, F-measure too low.")
sys.stdout.flush()
return 0.0
P_f, R_f, F_f = np.mean(frames, axis=0)
P_n, R_n, F_n = np.mean(notes, axis=0)
print(
f"Frame P,R,F: {P_f:.3f},{R_f:.3f},{F_f:.3f}, Note P,R,F: {P_n:.3f},{R_n:.3f},{F_n:.3f}"
)
print(str(F_n) + ": " + str(params))
sys.stdout.flush()
return -F_n
filename = os.path.join(folder,filename)
data = DataMaps()
data.make_from_file(filename,'time',section=[0,30], acoustic_model='kelz')
roll_gt = data.target
data_est = DataMapsBeats()
data_est.make_from_file(filename,False,section=[0,30], acoustic_model='kelz')
est_roll = np.loadtxt(os.path.join(folder_est,os.path.basename(filename).replace('.mid','_pr.csv')))
# est_roll = (data_est.input>0.5).astype(float)
roll_time_est = convert_note_to_time(est_roll,data_est.corresp,25,30)
data_gt = DataMapsBeats()
data_gt.make_from_file(filename,True,section=[0,30], acoustic_model='kelz')
gt_roll = np.loadtxt(os.path.join(folder_gt,os.path.basename(filename).replace('.mid','_pr.csv')))
# gt_roll = (data_gt.input>0.5).astype(float)
roll_time_gt = convert_note_to_time(gt_roll,data_gt.corresp,25,30)
print(est_roll.shape)
print(gt_roll.shape)
# plt.subplot(221)
def test(params):
global priors
global step
global beat_gt
global beat_subdiv
global order
global with_offset
global data_path
global model_dir
print(params)
transitions = np.zeros((88, 2**order, 2))
for i in range(88):
transitions[i, :, 0] = params
transitions[i, :, 1] = 1 - transitions[i, :, 0]
frames = np.zeros((0, 3))
notes = np.zeros((0, 3))
folder = data_path
for file in glob.glob(os.path.join(folder, "*.mid")):
print(file)
sys.stdout.flush()
if step == "beat":
data = DataMapsBeats()
data.make_from_file(file,
beat_gt,
beat_subdiv, [0, 30],
acoustic_model='kelz')
else:
data = DataMaps()
data.make_from_file(file, step, [0, 30], acoustic_model='kelz')
pr = hmm_eval.decode_all_pitches(data.input, priors, transitions)
if step != "time":
pr = convert_note_to_time(pr,
data.corresp,
data.input_fs,
max_len=30)
data = DataMaps()
data.make_from_file(file,
"time",
section=[0, 30],
acoustic_model="kelz")
target = data.target
#Evaluate
P_f, R_f, F_f = compute_eval_metrics_frame(pr, target)
P_n, R_n, F_n = compute_eval_metrics_note(pr,
target,
min_dur=0.05,
with_offset=with_offset)
print(
f"Frame P,R,F: {P_f:.3f},{R_f:.3f},{F_f:.3f}, Note P,R,F: {P_n:.3f},{R_n:.3f},{F_n:.3f}"
)
sys.stdout.flush()
frames = np.vstack((frames, [P_f, R_f, F_f]))
notes = np.vstack((notes, [P_n, R_n, F_n]))
P_f, R_f, F_f = np.mean(frames, axis=0)
P_n, R_n, F_n = np.mean(notes, axis=0)
print(
f"Frame P,R,F: {P_f:.3f},{R_f:.3f},{F_f:.3f}, Note P,R,F: {P_n:.3f},{R_n:.3f},{F_n:.3f}"
)
print(str(F_n) + ": " + str(params))
sys.stdout.flush()
out = os.path.join(model_dir,
step + "." + str(order) + "." + str(F_n) + ".pkl")
with open(out, "wb") as file:
pickle.dump({"priors": priors, "transitions": transitions}, file)
return -F_n