def create_midi_file(gt_midi, model_output_dir, midi_output_dir, filename, step='time', max_len=30):
"""
Convert csv piano rolls into MIDI files, for a single given MIDI file.
Parameters
==========
gt_midi : string
The directory containing all of the ground truth MIDI files there are csv files for.
model_output_dir : string
The directory containing the output csv files to convert.
midi_output_dir : string
The directory to write the resulting MIDI files to.
filename : string
The MAPS filename of the MIDI file to convert.
step : string
The frame step type to use for conversion. Either "time" (default), "quant", or "event".
max_len : int
The number of seconds of each file to convert. Defaults to 30.
"""
data = DataMaps()
data.make_from_file(os.path.join(gt_midi, filename), step, [0, max_len], acoustic_model="kelz")
csv_path = os.path.join(model_output_dir, filename.replace('.mid', '_pr.csv'))
roll = np.loadtxt(csv_path)
roll_time = convert_note_to_time(roll, data.corresp, data.input_fs, max_len=max_len)
midi_data = make_midi_from_roll(roll_time, 25)
output_filename = os.path.join(midi_output_dir, get_name_from_maps(filename) + '_' + filename[-6:-4])
save_midi(midi_data, output_filename + '.mid')
def test(params):
global priors
global step
global order
global with_offset
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/outputs-20/valid"
for file in glob.glob(os.path.join(folder, "*.mid")):
print(file)
sys.stdout.flush()
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 = "hmm/models/" + step + "." + str(order) + "." + str(F_n) + ".pkl"
with open(out, "wb") as file:
pickle.dump({"priors" : priors,
"transitions" : transitions}, file)
return -F_n
def weight_search(params, num=0, verbose=False):
print(params)
sys.stdout.flush()
gt = params[0]
min_diff = params[1]
history = int(params[2])
num_layers = int(params[3])
is_weight = params[4]
features = params[5]
history_context = 0
prior_context = 0
if len(params) > 6:
history_context = params[6]
prior_context = params[7]
use_lstm = True
if len(params) > 8:
use_lstm = params[8]
warnings.filterwarnings("ignore", message="tick should be an int.")
max_len = 30
section = [0, max_len]
note_range = [21, 109]
note_min = note_range[0]
note_max = note_range[1]
# Load model
model = model_dict['model']
sess = model_dict['sess']
# Get weight_model data
pkl = data_dict['gt' if gt else 'beam']
X = pkl['X']
Y = pkl['Y']
D = pkl['D']
max_history = pkl['history']
no_mlm = pkl['no_mlm'] if 'no_mlm' in pkl else False
if np.max(D) < min_diff:
print("No training data generated")
sys.stdout.flush()
return 0.0
data_points = np.where(D > min_diff)
data_features = []
if history > 0:
data_features.extend(range(max_history - history, max_history))
if features:
data_features.extend(range(max_history, len(X[0]) - 2))
data_features.append(-2)
if use_lstm:
data_features.append(-1)
X = X[:, data_features]
if prior_context + history_context > 0:
X_new = np.zeros(
(X.shape[0],
X.shape[1] + prior_context * 4 + 2 * history_context * history))
for i in range(int(X.shape[0] / 88)):
x_frame = X[88 * i:88 * (i + 1), :]
X_new[88 * i:88 * (i + 1), :] = pad_x(x_frame, x_frame[:, -2],
x_frame[:, -1],
x_frame[:, :history],
history, history_context,
prior_context)
X = X_new
X = X[data_points]
Y = Y[data_points]
if len(X) == 0:
print("No training data generated")
sys.stdout.flush()
return 0.0
# Train weight model
print("Training weight model")
sys.stdout.flush()
layers = []
for i in range(num_layers):
layers.append(5)
weight_model = train_model(X, Y, layers=layers, weight=is_weight)
global most_recent_model
most_recent_model = {
'model': weight_model,
'history': history,
'features': features,
'weight': is_weight,
'history_context': history_context,
'prior_context': prior_context,
'use_lstm': use_lstm,
'no_mlm': no_mlm
}
weight_model_name = "weight_model."
weight_model_name += "gt" if gt else "b10"
weight_model_name += "_md" + str(min_diff)
weight_model_name += "_h" + str(history)
weight_model_name += "_l" + str(num_layers)
if features:
weight_model_name += "_f"
weight_model_name += "_hc" + str(history_context)
weight_model_name += "_pc" + str(prior_context)
if not use_lstm:
weight_model_name += "_noLSTM"
if no_mlm:
weight_model_name += "_noMLM"
weight_model_name += "_weight" if is_weight else "_prior"
weight_model_name += "." + global_params['step'] + "." + str(num) + ".pkl"
# 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)
results = {}
frames = np.zeros((0, 3))
notes = np.zeros((0, 3))
for filename in glob.glob(os.path.join(data_dict['valid'], "*.mid")):
print(filename)
sys.stdout.flush()
data = DataMaps()
data.make_from_file(filename,
global_params['step'],
section,
acoustic_model=global_params['acoustic'])
# Decode
pr, priors, weights, combined_priors = decode(
data.input,
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)
if global_params['step'] != "time":
pr = convert_note_to_time(pr,
data.corresp,
data.input_fs,
max_len=max_len)
data = DataMaps()
data.make_from_file(filename,
"time",
section=section,
acoustic_model=global_params['acoustic'])
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
input_roll = np.loadtxt(filename_input)
target_roll = data.target
mask = data.get_key_profile()
mask_octave = data.get_key_profile_octave()
# 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],
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)
# plt.imshow(data_est.input,aspect='auto',origin='lower')
# plt.subplot(222)
# plt.imshow(est_roll,aspect='auto',origin='lower')
#
data = DataMapsBeats()
data.make_from_file(file,args.beat_gt,args.beat_subdiv,section, acoustic_model="kelz")
else:
data = DataMaps()
data.make_from_file(file, args.step, section,acoustic_model="kelz")
pr = decode_all_pitches(data.input, priors, transitions)
# Save output
if not args.save is None:
np.save(os.path.join(args.save, base.replace('.mid','_pr')), pr)
np.savetxt(os.path.join(args.save, base.replace('.mid','_pr.csv')), pr)
if args.step in ['quant','event','beat']:
pr = convert_note_to_time(pr, data.corresp, data.input_fs, max_len=max_len)
data = DataMaps()
data.make_from_file(file, "time", section=section, 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)
frames = np.vstack((frames, [P_f, R_f, F_f]))
notes = np.vstack((notes, [P_n, R_n, F_n]))
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()
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
