def load_jsb(path):
mkdir_p(path)
d = os.path.join(path, 'JSB Chorales')
if not os.path.isdir(d):
download_jsb(path)
train_filenames = os.path.join(path, 'JSB Chorales', 'train', '*.mid')
valid_filenames = os.path.join(path, 'JSB Chorales', 'valid', '*.mid')
test_filenames = os.path.join(path, 'JSB Chorales', 'test', '*.mid')
train_files = glob.glob(train_filenames)
valid_files = glob.glob(valid_filenames)
test_files = glob.glob(test_filenames)
train_datasets = [
midiread(f, r=(21, 109),
dt=0.3).piano_roll.astype(theano.config.floatX)
for f in train_files
]
valid_datasets = [
midiread(f, r=(21, 109),
dt=0.3).piano_roll.astype(theano.config.floatX)
for f in valid_files
]
test_datasets = [
midiread(f, r=(21, 109),
dt=0.3).piano_roll.astype(theano.config.floatX)
for f in test_files
]
return (train_datasets, [None]), (valid_datasets, [None]), (test_datasets,
[None])
def train(self, file_name, weight_save_file, batch_size=1, num_epoch=200):
print('load data ---------------')
file_train=os.path.join(os.path.split(os.path.dirname(__file__))[0],
'data',file_name,'train','*.mid')
dataset = [midiread(f, self.r, self.dt).piano_roll.astype(theano.config.floatX) for f in glob.glob(file_train)]
file_test=os.path.join(os.path.split(os.path.dirname(__file__))[0],
'data',file_name,'test','*.mid')
testdataset = [midiread(f, self.r, self.dt).piano_roll.astype(theano.config.floatX) for f in glob.glob(file_test)]
print('load done --------------')
try:
for epoch in range(num_epoch):
t0 = time.time()
numpy.random.shuffle(dataset)
costs = []
accuracys = []
for s, sequence in enumerate(dataset):
y = numpy.hstack((sequence,numpy.zeros((sequence.shape[0],1)) ))
x = numpy.roll(y, 1, axis=0)
x[0,:]=0
x[0,self.maxFeatures-1]=1
cost, accuracy= self.rnnModel.train_on_batch(numpy.array([x]), numpy.array([y]), accuracy=True)
costs.append(cost)
accuracys.append(accuracy)
print('epoch: %i/%i\tcost: %.5f\taccu: %.5f\ttime: %.4f s' % (epoch+1, num_epoch, numpy.mean(costs), numpy.mean(accuracys),time.time()-t0))
sys.stdout.flush()
test_accu=self.evaluate(testdataset)
print('test_accu: %.5f' % ( numpy.mean(test_accu)) )
self.rnnModel.save_weights(weight_save_file)
except KeyboardInterrupt:
print('interrupt by user !')
def load_generation(N, name):
re_gen = os.path.join(os.path.split(os.path.dirname(__file__))[0], name, '*.mid')
files_gen = glob.glob(re_gen)
dataset_gen = [midiread(f, (21,109), 0.3).piano_roll.astype(theano.config.floatX) for f in files_gen]
all_gram_gen = dict()
for i in range(N):
all_gram_gen[i+1]=dict()
for i in range(N):
n = i+1
n_gram = all_gram_gen.get(n)
for s, sequence in enumerate(dataset_gen):
count = dict()
for l in range(len(sequence)-n+1):
gram = sequence[l : l+n]
if count.has_key(str(gram)):
count[str(gram)] += 1
else:
count[str(gram)] = 1
for k in count:
if n_gram.has_key(k):
n_gram[k]=max(count.get(k), n_gram.get(k))
else:
n_gram[k]=count.get(k)
return all_gram_gen
def train(self, files, batch_size=100, num_epochs=200):
'''Train the RNN-RBM via stochastic gradient descent (SGD) using MIDI
files converted to piano-rolls.
files : list of strings
List of MIDI files that will be loaded as piano-rolls for training.
batch_size : integer
Training sequences will be split into subsequences of at most this
size before applying the SGD updates.
num_epochs : integer
Number of epochs (pass over the training set) performed. The user
can safely interrupt training with Ctrl+C at any time.'''
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files]
try:
for epoch in range(num_epochs):
numpy.random.shuffle(dataset)
costs = []
for s, sequence in enumerate(dataset):
for i in range(0, len(sequence), batch_size):
cost = self.train_function(sequence[i:i + batch_size])
costs.append(cost)
print('Epoch %i/%i' % (epoch + 1, num_epochs))
print(numpy.mean(costs))
sys.stdout.flush()
except KeyboardInterrupt:
print('Interrupted by user.')
def modeling_n_gram(n, files):
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
for f in files:
print 'parsing', f
each = midiread(f).piano_roll.astype(theano.config.floatX)
numNote = len(each[0])
# each [ time ] [ note ]
for timeSlice in range(n-1, len(each)):
for noteDest in range(numNote):
valueDest = int(each[timeSlice][noteDest])
for noteFrom in range(numNote):
valueFrom = int(each[timeSlice-1][noteFrom])
#print noteDest,valueDest,noteFrom,valueFrom
probability[noteDest][valueDest][noteFrom][valueFrom] \
= probability[noteDest][valueDest][noteFrom][valueFrom] + 1.0
print 'learning_done'
pkl.dump(probability, open("bi-gram-count.dat", "wb"))
print 'bi-gram-count.dat saved'
def train(self, files, batch_size=100, num_epochs=200):
'''Train the RNN-RBM via stochastic gradient descent (SGD) using MIDI
files converted to piano-rolls.
files : list of strings --- List of MIDI files that will be loaded as piano-rolls for training.
batch_size : integer --- Training sequences will be split into subsequences of at most this size before applying the SGD updates.
num_epochs : integer --- Number of epochs (pass over the training set) performed. The user can safely interrupt training with Ctrl+C at any time.'''
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files]
try:
for epoch in xrange(num_epochs):
numpy.random.shuffle(dataset)
costs = []
for s, sequence in enumerate(dataset):
for i in xrange(0, len(sequence), batch_size):
cost = self.train_function(sequence[i:i + batch_size])
costs.append(cost)
print 'Epoch %i/%i' % (epoch + 1, num_epochs),
print numpy.mean(costs)
sys.stdout.flush()
except KeyboardInterrupt:
print 'Interrupted by user.'
def train(r, dt, files, f_grad_shared, f_grad_update, f_pred, \
batch_size=100, num_epochs=20, lr = 1E-6):
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [
midiread(f, r, dt).piano_roll.astype(theano.config.floatX)
for f in files
]
# try:
seq = None
for epoch in range(num_epochs):
numpy.random.shuffle(dataset)
costs = []
for s, sequence in enumerate(dataset):
for i in range(0, len(sequence), batch_size):
if i + 1 + batch_size < len(sequence):
loss = f_grad_shared(sequence[i:i + batch_size],
sequence[i + 1:i + batch_size + 1])
costs.append(loss)
f_grad_update(lr)
if seq is None:
seq = sequence[i:i + batch_size]
print('Epoch %i/%i' % (epoch + 1, num_epochs))
print(numpy.mean(costs))
sys.stdout.flush()
numpy.savetxt('matrix.txt', seq, delimiter=',')
return seq
def load_dataset(self):
re = ["data/JSB Chorales/train/*.mid"]
re.append("data/JSB Chorales/test/*.mid")
re.append("data/JSB Chorales/valid/*.mid")
files = []
for r in re:
files += glob.glob(r)
assert(len(files) > 0)
print "generating dataset..."
dataset = [midiread(f, self.r, self.dt).piano_roll.astype(theano.config.floatX) for f in files]
memorization_dataset = [[]] # memorize the first unit for 100 time-steps with binary noise
n = 0
for seq in dataset:
for i in range(0, len(seq), self.seq_length):
memorization_dataset[0].append(seq[i:i + self.seq_length])
while False: #n < 100000:
sequence = random.choice(dataset)
if len(sequence) < self.seq_length:
print " to short !"
i = random.choice(range(0, len(sequence), self.seq_length))
memorization_dataset[0].append(sequence[i:i + self.seq_length])
n = n + 1
print "number of sequence for training : ", len(memorization_dataset[0])
self.train = [memorization_dataset[0][:-1000]]
self.valid = [memorization_dataset[0][-1000:]]
self.gradient_dataset = SequenceDataset(self.train, batch_size=None, number_batches=1000)
self.cg_dataset = SequenceDataset(self.train, batch_size=None, number_batches=500)
self.valid_dataset = SequenceDataset(self.valid, batch_size=None, number_batches=500)
def load_dataset(N, name):
re = os.path.join(os.path.split(os.path.dirname(__file__))[0], 'data', name, 'train', '*.mid')
files = glob.glob(re)
dataset = [midiread(f, (21,109), 0.3).piano_roll.astype(theano.config.floatX) for f in files]
all_gram = dict()
for i in range(N):
all_gram[i+1]=dict()
for i in range(N):
n = i+1
print ('======'+str(n)+'======')
n_gram = all_gram.get(n)
for s, sequence in enumerate(dataset):
print (s)
count = dict()
for l in range(len(sequence)-n+1):
gram = sequence[l : l+n]
if count.has_key(str(gram)):
count[str(gram)] += 1
else:
count[str(gram)] = 1
for k in count:
if n_gram.has_key(k):
n_gram[k]=max(count.get(k), n_gram.get(k))
else:
n_gram[k]=count.get(k)
return all_gram
def __init__(self, dir_path, set_name, pitch_range, dt, batch_size, shuffle=True): self.dir_path = dir_path self.set_name = set_name self.pitch_range = pitch_range self.dt = dt self.batch_size = batch_size self.shuffle = shuffle self.curr_epoch = 0 self.batch_idx = 0 name_list = os.listdir(self.dir_path+'/'+self.set_name) self.data_num = len(name_list) self.data = [] for m in name_list: f = self.dir_path+'/'+self.set_name+'/'+m wave = midiread(f, self.pitch_range, self.dt) self.data.append(wave.piano_roll) self.data_list = np.array(range(self.data_num)) if self.shuffle: np.random.shuffle(self.data_list) if set_name=='train': self.batch_num = self.data_num/self.batch_size else: self.batch_num = int(np.ceil(1.0*self.data_num/self.batch_size))
def train(self, file_name, weight_save_file, batch_size=1, num_epoch=200):
print('load data ---------------')
file_train = os.path.join(
os.path.split(os.path.dirname(__file__))[0], 'data', file_name,
'train', '*.mid')
dataset = [
midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in glob.glob(file_train)
]
file_test = os.path.join(
os.path.split(os.path.dirname(__file__))[0], 'data', file_name,
'test', '*.mid')
testdataset = [
midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in glob.glob(file_test)
]
print('load done --------------')
try:
for epoch in range(num_epoch):
t0 = time.time()
numpy.random.shuffle(dataset)
costs = []
accuracys = []
for s, sequence in enumerate(dataset):
y = numpy.hstack(
(sequence, numpy.zeros((sequence.shape[0], 1))))
x = numpy.roll(y, 1, axis=0)
x[0, :] = 0
x[0, self.maxFeatures - 1] = 1
cost, accuracy = self.rnnModel.train_on_batch(
numpy.array([x]), numpy.array([y]), accuracy=True)
costs.append(cost)
accuracys.append(accuracy)
print('epoch: %i/%i\tcost: %.5f\taccu: %.5f\ttime: %.4f s' %
(epoch + 1, num_epoch, numpy.mean(costs),
numpy.mean(accuracys), time.time() - t0))
sys.stdout.flush()
test_accu = self.evaluate(testdataset)
print('test_accu: %.5f' % (numpy.mean(test_accu)))
self.rnnModel.save_weights(weight_save_file)
except KeyboardInterrupt:
print('interrupt by user !')
def train(self, files, batch_size=100, num_epochs=200):
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files]
def accuracy (v, v_sample):
accs = []
t, n = v.shape
for time in range(t):
tp = 0 # true positive
fp = 0 # false positive
fn = 0 # false negative
for note in range(n):
if v[time][note] == 1 and v_sample[time][note] == 1:
tp += 1.
if v[time][note] == 0 and v_sample[time][note] == 1:
fp += 1.
if v[time][note] == 1 and v_sample[time][note] == 0:
fn += 1.
if tp + fp + fn != 0:
a = tp / (tp + fp + fn)
else:
a = 0
accs.append(a)
acc = numpy.mean(accs)
return acc
try:
print ('lstm_rbm, dataset=Nottingham, lr=%f, epoch=%i' %(self.lr, num_epochs))
for epoch in xrange(num_epochs):
numpy.random.shuffle(dataset)
costs = []
accs = []
for s, sequence in enumerate(dataset):
for i in xrange(0, len(sequence), batch_size):
v = sequence[i:i + batch_size]
(cost, v_sample) = self.train_function(v)
costs.append(cost)
acc = accuracy(v, v_sample)
accs.append(acc)
p = 'Epoch %i/%i LL %f ACC %f time %s' % (epoch + 1, num_epochs, numpy.mean(costs), numpy.mean(accs), datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"))
print (p)
if (epoch%100 == 0 or epoch==num_epochs-1):
piano_roll = self.generate_function()
midiwrite('sample/lstm_rbm%i.mid' %(epoch), piano_roll, self.r, self.dt)
sys.stdout.flush()
except KeyboardInterrupt:
print 'Interrupted by user.'
def load_jsb(path):
mkdir_p(path)
d = os.path.join(path, 'JSB Chorales')
if not os.path.isdir(d):
download_jsb(path)
train_filenames = os.path.join(path, 'JSB Chorales', 'train', '*.mid')
valid_filenames = os.path.join(path, 'JSB Chorales', 'valid', '*.mid')
test_filenames = os.path.join(path, 'JSB Chorales', 'test', '*.mid')
train_files = glob.glob(train_filenames)
valid_files = glob.glob(valid_filenames)
test_files = glob.glob(test_filenames)
train_datasets = [midiread(f, r=(21, 109), dt=0.3).piano_roll.astype(theano.config.floatX) for f in train_files]
valid_datasets = [midiread(f, r=(21, 109), dt=0.3).piano_roll.astype(theano.config.floatX) for f in valid_files]
test_datasets = [midiread(f, r=(21, 109), dt=0.3).piano_roll.astype(theano.config.floatX) for f in test_files]
return (train_datasets, [None]), (valid_datasets, [None]), (test_datasets, [None])
def dataset_load(path,
r=(21, 109), #Key span
dt=0.3): #Timestep for conversion into piano roll
#Read files
files = glob.glob('data/Nottingham/train/*')
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
pianorolls = [midiread(f, r, dt).piano_roll for f in files]
#Concatenate
dataset = np.concatenate(pianorolls, axis = 0) #TODO add some zeros, so that one music piece is clearly different from another.
return dataset
def load_midi_data(data_dir):
import midi.utils as utils
from midi import MidiInFile as mf
from midi import MidiToText as mt
f = open(data_dir, 'rb')
midiIn = mf.MidiInFile(mt.MidiToText(), f)
midiIn.read()
f.close()
midi_data = utils.midiread(data_dir, dt=0.5)
return midi_data.piano_roll
def loadDataSet(files):
#File e il path della carterlla contente i file (*.mid)
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
#mi calcolo quel'el 'esempio di lunghezza massima
maxLen=0
dataset=[]
length=[]
for f in files:
currentMidi=midiread(f, (21, 109),0.3).piano_roll.astype(np.float64)
dataset.append(currentMidi)
length.append(currentMidi.shape[0])
if maxLen<currentMidi.shape[0]:
maxLen=currentMidi.shape[0]
print "MAXLEN: ",maxLen
return dataset,maxLen,len(dataset),length
def LoadDataForPreTraining(r=(21, 109), dt=0.3):
assert len(trainingSet) > 0, 'Training set is empty!' \
' (did you download the data files?)'
sampleLen=[]
dataset=[]
maxLen=0
nSample=0
for f in trainingSet:
currentMidi=midiread(f, (21, 109),0.3).piano_roll
dataset.append(currentMidi)
sampleLen.append(currentMidi.shape[0])
if maxLen< currentMidi.shape[0]:
maxLen=currentMidi.shape[0]
nSample=nSample+1
return (dataset, sampleLen, nSample, maxLen)
def train(self, files, batch_size=100, num_epochs=200):
'''Train the RNN-RBM via stochastic gradient descent (SGD) using MIDI
files converted to piano-rolls.
files : list of strings
List of MIDI files that will be loaded as piano-rolls for training.
batch_size : integer
Training sequences will be split into subsequences of at most this
size before applying the SGD updates.
num_epochs : integer
Number of epochs (pass over the training set) performed. The user
can safely interrupt training with Ctrl+C at any time.'''
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [
midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files
]
print "Start training process of the recurrent network RBM machine with the given dataset...",
print "Lenght of the Dataset: ", len(files)
print "Interrupt if necessariy by pressing Ctrl+C",
print "...Might take some time :) ..."
costst = []
try:
for epoch in xrange(num_epochs):
numpy.random.shuffle(dataset)
costs = []
for s, sequence in enumerate(dataset):
for i in xrange(0, len(sequence), batch_size):
cost = self.train_function(sequence[i:i + batch_size])
costs.append(cost)
print 'Epoch %i/%i' % (epoch + 1, num_epochs),
print "Current mean Energy cost:", numpy.mean(costs)
costst.append(numpy.mean(costs))
print "Training %i percent done; interrupt training by pressing Crtl+C." % (
float((float(epoch) + 1.0) * 100.0 / float(num_epochs)))
sys.stdout.flush()
except KeyboardInterrupt:
print 'Training Interrupted.'
return self, files, costst
def fetch_nottingham():
url = "http://www.iro.umontreal.ca/~lisa/deep/data/Nottingham.zip"
data_path = "Nottingham.zip"
if not os.path.exists(data_path):
download(url, data_path)
key_range = (21, 109)
dt = 0.3
all_data = []
with zipfile.ZipFile(data_path, "r") as f:
for name in f.namelist():
if ".mid" not in name:
# Skip README
continue
data = midiread(f, key_range,
dt).piano_roll.astype(theano.config.floatX)
all_data.extend(data)
return key_range, dt, all_data
def loadDataSet(files):
#File e il path della carterlla contente i file (*.mid)
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
#mi calcolo quel'el 'esempio di lunghezza massima
maxLen=0
dataset=[]
for f in files:
currentMidi=midiread(f, (21, 109),0.3).piano_roll.astype(theano.config.floatX)
dataset.append(currentMidi)
if maxLen< currentMidi.shape[0]:
maxLen=currentMidi.shape[0]
#porto tutte le tracce a masima lunghezza aggiongendo silenzio
for i, seq in enumerate(dataset):
if seq.shape[0]<maxLen:
dataset[i]=np.concatenate([seq, np.zeros((maxLen-seq.shape[0], 88))])
#print dataset[0].shape
return np.array(dataset, dtype=theano.config.floatX)
def train(self, files, batch_size=100, num_epochs=200):
'''Train the RNN-RBM via stochastic gradient descent (SGD) using MIDI
files converted to piano-rolls.
files : list of strings
List of MIDI files that will be loaded as piano-rolls for training.
batch_size : integer
Training sequences will be split into subsequences of at most this
size before applying the SGD updates.
num_epochs : integer
Number of epochs (pass over the training set) performed. The user
can safely interrupt training with Ctrl+C at any time.'''
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files]
print "Start training process of the recurrent network RBM machine with the given dataset..." ,
print "Lenght of the Dataset: ", len(files)
print "Interrupt if necessariy by pressing Ctrl+C",
print "...Might take some time :) ..."
costst = []
try:
for epoch in xrange(num_epochs):
numpy.random.shuffle(dataset)
costs = []
for s, sequence in enumerate(dataset):
for i in xrange(0, len(sequence), batch_size):
cost = self.train_function(sequence[i:i + batch_size])
costs.append(cost)
print 'Epoch %i/%i' % (epoch + 1, num_epochs),
print "Current mean Energy cost:", numpy.mean(costs)
costst.append(numpy.mean(costs))
print "Training %i percent done; interrupt training by pressing Crtl+C." % (float((float(epoch) + 1.0)*100.0 / float(num_epochs)))
sys.stdout.flush()
except KeyboardInterrupt:
print 'Training Interrupted.'
return self, files, costst
def loadDataSetMin(files):
#File e il path della carterlla contente i file (*.mid)
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
#mi calcolo quel'el 'esempio di lunghezza massima
minLen=sys.maxint
dataset=[]
for f in files:
currentMidi=midiread(f, (21, 109),0.3).piano_roll.astype(theano.config.floatX)
dataset.append(currentMidi)
if minLen> currentMidi.shape[0]:
minLen=currentMidi.shape[0]
#porto tutte le tracce a masima lunghezza aggiongendo silenzio
for i, seq in enumerate(dataset):
if seq.shape[0]>minLen:
dataset[i]=seq[0:minLen,:]
#print dataset[0].shape
#print "MINLEN: ", minLen
return np.array(dataset, dtype=theano.config.floatX)
def train(self, files, batch_size=500, num_epochs=300):
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files]
try:
for epoch in xrange(num_epochs):
numpy.random.shuffle(dataset)
costs = []
for s, sequence in enumerate(dataset):
for i in xrange(0, len(sequence), batch_size):
cost = self.train_function(sequence[i:i + batch_size])
costs.append(cost)
print 'Epoch %i/%i' % (epoch + 1, num_epochs),
print numpy.mean(costs)
sys.stdout.flush()
except KeyboardInterrupt:
print 'Interrupted by user.'
def LoadDataForPreTraining(r=(21, 109), dt=0.3):
assert len(trainingSet) > 0, 'Training set is empty!' \
' (did you download the data files?)'
sampleLen=[]
dataset=[]
maxLen=0
nSample=0
for f in trainingSet:
currentMidi=midiread(f, (21, 109),0.3).piano_roll
print f,": lenght= ",currentMidi.shape[0]
sampleLen.append(currentMidi.shape[0])
if maxLen< currentMidi.shape[0]:
maxLen=currentMidi.shape[0]
nSample=nSample+1
print "#-----dataset data-----#"
print "number of sample: ", len(trainingSet)
print "max lenght:", maxLen
print "# rows of matrix M:",sum(sampleLen)
print "# columns of matrix M:", maxLen*88
return (dataset, sampleLen, nSample, maxLen)
def load_dataset(style='Nottingham',
data_type='train',
max_length=256,
max_key_number=1):
"""
Load dataset from files
for mono case: max_key_number would be 1
for poly case: max_key_number default is 4
"""
file_paths = get_file_path(style, data_type)
dataset = [
midiread(f, R, DT).piano_roll.astype(np.float32) for f in file_paths
]
dataset = [
np.array([
get_index(dataset[i][j], data_type, max_key_number)
for j in xrange(len(dataset[i]))
]) for i in xrange(len(dataset))
]
# print('min length:', sum([len(d) for d in dataset])/len(dataset))
dataset = np.array(
[np.array(padding(key2frequency(d), max_length)) for d in dataset])
return dataset
def train(self, session, num_epochs=200):
dataset = [
midiread(f, self.r, self.dt).piano_roll.astype(numpy.float32)
for f in self.train_files
][0:5]
num_samples = len(dataset)
print 'total samples: %d' % num_samples
for epoch in xrange(num_epochs):
numpy.random.shuffle(dataset)
for s, sequence in enumerate(dataset):
total_cost = list()
for i in range(0,
len(sequence) - self.batch_size + 1,
self.batch_size):
v_batch = sequence[i:i + self.batch_size]
train_result = session.run([self.cost, self.train_op],
feed_dict={self.v: v_batch})
total_cost.append(train_result[0])
if s % 100 == 0:
if len(total_cost) > 0:
mean_cost = sum(total_cost) / float(len(total_cost))
print 'epoch %d, %dth sample length %d, cost: %f' % (
epoch, s, len(sequence), mean_cost)
dt = 0.2
train_files = glob.glob(r'../rbmrnn/Nottingham/train/*.mid')
n_input = r[1] - r[0]
n_hidden = 64
n_output = r[1] - r[0]
n_steps = 50
learning_rate = 0.5
learning_rate_decay = 0.99
batch_size = 32
num_epochs = 2000
check_epochs = 1
generation_path = 'generated/'
filename = '001.mid'
dataset = [
midiread(f, r, dt).piano_roll.astype(numpy.float32) for f in train_files
]
num_samples = len(dataset)
print 'total samples: %d' % num_samples
x = tf.placeholder(tf.float32, [None, n_steps, n_input])
state = tf.placeholder(tf.float32, [None, n_hidden])
output = tf.placeholder(tf.float32, [None, n_hidden])
learning_rate_variable = tf.Variable(float(learning_rate), trainable=False)
learning_rate_decay_op = learning_rate_variable.assign(learning_rate_variable *
learning_rate_decay)
cost, train_op, generator = model.lstm(x, state, output, n_steps, n_input,
n_hidden, n_output,
learning_rate_variable)
def generation_length():
re_gen = os.path.join(os.path.split(os.path.dirname(__file__))[0], 'generation', '*.mid')
files_gen = glob.glob(re_gen)
dataset_gen = [midiread(f, (21,109), 0.3).piano_roll.astype(theano.config.floatX) for f in files_gen]
return len(dataset_gen[0])
def train(self, files, batch_size=100, num_epochs=200):
'''Train the RNN-RBM via stochastic gradient descent (SGD) using MIDI
files converted to piano-rolls.
files : list of strings
List of MIDI files that will be loaded as piano-rolls for training.
batch_size : integer
Training sequences will be split into subsequences of at most this size
before applying the SGD updates.
num_epochs : integer
Number of epochs (pass over the training set) performed. The user can
safely interrupt training with Ctrl+C at any time.'''
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [
midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files
]
print "DONE DOWNLOADING"
#self.pro.start()
try:
count = 0
processed = 0
inVecs = [
self.inVecQueue, self.inVecQueue2, self.inVecQueue3,
self.inVecQueue4, self.inVecQueue5, self.inVecQueue6,
self.inVecQueue7, self.inVecQueue8
]
for epoch in xrange(num_epochs):
numpy.random.shuffle(dataset)
costs = []
''' self.l1.acquire()
self.l2.acquire()
self.l3.acquire()
self.l4.acquire()
self.l5.acquire()
self.l6.acquire()
self.l7.acquire()
self.l8.acquire()'''
#ds = chunks(dataset, len(dataset)/4)
count = 0
processed = 0
for s, sequence in enumerate(dataset):
inVecs[s % 8].put(sequence)
#count = count + len(sequence)/100
for j in xrange(0, len(sequence), batch_size):
#print "GIVING DATA"
count = count + 1
#inVecs[s%8].put(sequence[j:j+batch_size])
'''self.l1.release()
self.l2.release()
self.l3.release()
self.l4.release()
self.l5.release()
self.l6.release()
self.l7.release()
self.l8.release()'''
while processed != count:
if self.costQueue.qsize() > 0:
cost = self.costQueue.get()
costs.append(cost)
processed = processed + 1
sums = []
items = self.weightQueue.qsize()
#while self.weightQueue.qsize() > 0:
# d = self.weightQueue.get()
# sums.append(d)
adder = [None, None, None, None, None, None, None, None]
for i in xrange(0, 8):
q = self.weights[i].get()
for j in xrange(0, 8):
if adder[j] == None:
adder[j] = q[j]
else:
adder[j] += q[j]
for i in range(0, 8):
adder[i] /= float(8)
p = adder[0], adder[1], adder[2], adder[3], adder[4], adder[
5], adder[6], adder[7]
if epoch != num_epochs:
self.weightQueue.put(p)
self.weightQueue2.put(p)
self.weightQueue3.put(p)
self.weightQueue4.put(p)
self.weightQueue5.put(p)
self.weightQueue6.put(p)
self.weightQueue7.put(p)
self.weightQueue8.put(p)
print 'Epoch %i/%i' % (epoch + 1, num_epochs),
print numpy.mean(costs)
#print profmode.print_summary()
sys.stdout.flush()
except KeyboardInterrupt:
print 'Interrupted by user.'
self.pro.terminate()
self.pro2.terminate()
self.pro3.terminate()
self.pro4.terminate()
self.pro5.terminate()
self.pro6.terminate()
self.pro7.terminate()
self.pro8.terminate()
self.weightQueue.close()
self.weightQueue2.close()
self.weightQueue3.close()
self.weightQueue4.close()
self.weightQueue5.close()
self.weightQueue6.close()
self.weightQueue7.close()
self.weightQueue8.close()
#print "NOW FETCHING",self.weightQueue.qsize()
p2 = self.args[0]['params']
#print "AFTER",len(p2)
print p2[0] == self.params[0]
self.W.set_value(p2[0])
self.bv.set_value(p2[1])
self.bh.set_value(p2[2])
self.Wuh.set_value(p2[3])
self.Wuv.set_value(p2[4])
self.Wvu.set_value(p2[5])
self.Wuu.set_value(p2[6])
self.bu.set_value(p2[7])
from scipy.io import wavfile
import numpy as np
THIS_DATA_DIR = dirname(realpath(__file__))
DOWNLOADED_ZIP = join(THIS_DATA_DIR, "dataset.zip")
DOWNLOADED_DIR = join(THIS_DATA_DIR, "dataset")
FILE_URL="http://c4dm.eecs.qmul.ac.uk/rdr/bitstream/handle/123456789/13/Score-informed%20Piano%20Transcription%20Dataset.zip?sequence=1"
if __name__ == '__main__':
if not exists(DOWNLOADED_ZIP):
execute_bash("wget -O {path} {url}".format(url=FILE_URL, path=DOWNLOADED_ZIP))
if exists(DOWNLOADED_DIR) and isdir(DOWNLOADED_DIR):
execute_bash("rm -rf %s" % (DOWNLOADED_DIR))
execute_bash("rm %s " % (join(THIS_DATA_DIR, "*.npy")))
makedirs(DOWNLOADED_DIR)
execute_bash("unzip %s -d %s" % (DOWNLOADED_ZIP, DOWNLOADED_DIR))
files = collect_files_with_ext(DOWNLOADED_DIR, ".wav")
for subpath, name in files:
if name.endswith(".wav") and "Chromatic" not in name:
sampling_rate, music = wavfile.read(subpath)
np.save(join(THIS_DATA_DIR, name.replace(".wav", ".npy")), music)
piece = midiread(str(subpath).replace(".wav", "_correct.mid"))
np.save(join(THIS_DATA_DIR, name.replace(".wav", ".mid.npy")), piece.piano_roll)
execute_bash("rm -rf %s" % (DOWNLOADED_DIR))
execute_bash("rm -rf %s" % (DOWNLOADED_ZIP))
EPOCHS = 1000
BATCH_SIZE = 128
TICKS_PER_INPUT = MEASURES*TICKS_PER_MEASURE
GEN_LENGTH = TICKS_PER_INPUT*8
DT = 0.3
if __name__ == '__main__':
### read MIDI
#data_dir = '../Data/Cirriculum/easy/'
data_dir = '../biaxial-rnn-music-composition/music/'
files = os.listdir(data_dir)
files = [data_dir + f for f in files if '.mid' in f]
dataset = [midiread(f, MIDI_RANGE, DT).piano_roll for f in files]
X = []
y = []
for song in dataset:
for i in range(0, len(song) - TICKS_PER_INPUT, STEP):
X.append(song[i: i + TICKS_PER_INPUT])
y.append(song[i + TICKS_PER_INPUT])
max_samples = (len(X) // BATCH_SIZE) * BATCH_SIZE
X = X[:max_samples]
y = y[:max_samples]
X = np.array(X)
y = np.array(y)
save_dir = './'
### read MIDI
#data_dir = '../Data/Cirriculum/easy/'
#data_dir = '../biaxial-rnn-music-composition/music/'
files = os.listdir(data_dir)
files = [data_dir + f for f in files if '.mid' in f or '.MID' in f]
print files
dataset = []
for f in files:
try:
dataset.append(midiread(f, MIDI_RANGE, DT).piano_roll)
print "{} loaded".format(f)
except IndexError:
print "Skipping {}".format(f)
pass
print np.shape(dataset)
X = []
y = []
for song in dataset:
for i in range(0, len(song) - TICKS_PER_INPUT, STEP):
X.append(song[i: i + TICKS_PER_INPUT])
y.append(song[i + TICKS_PER_INPUT])
# In[109]:
model.train(onlyfiles)
# In[117]:
model.generate('testHaydn3.mid')
# In[23]:
get_ipython().magic(u'pylab inline')
f = 'testNotthingham6.mid'
extent = (0, 0.25 * len(piano_roll)) + (21, 109)
piano_roll = midiread(f, (21, 109), 0.25).piano_roll.astype(theano.config.floatX)
pylab.figure()
pylab.imshow(piano_roll.T, origin='lower', aspect='auto',
interpolation='nearest', cmap=pylab.cm.gray_r,
extent=extent)
pylab.xlabel('time (s)')
pylab.ylabel('MIDI note number')
pylab.title('generated piano-roll')
pylab.show()
# In[ ]:
def train(self, files, batch_size=100, num_epochs=200):
'''Train the RNN-RBM via stochastic gradient descent (SGD) using MIDI
files converted to piano-rolls.
files : list of strings
List of MIDI files that will be loaded as piano-rolls for training.
batch_size : integer
Training sequences will be split into subsequences of at most this size
before applying the SGD updates.
num_epochs : integer
Number of epochs (pass over the training set) performed. The user can
safely interrupt training with Ctrl+C at any time.'''
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files]
print "DONE DOWNLOADING"
#self.pro.start()
try:
count = 0
processed = 0
inVecs = [self.inVecQueue,self.inVecQueue2,self.inVecQueue3,self.inVecQueue4,self.inVecQueue5,self.inVecQueue6,self.inVecQueue7,self.inVecQueue8]
for epoch in xrange(num_epochs):
numpy.random.shuffle(dataset)
costs = []
''' self.l1.acquire()
self.l2.acquire()
self.l3.acquire()
self.l4.acquire()
self.l5.acquire()
self.l6.acquire()
self.l7.acquire()
self.l8.acquire()'''
#ds = chunks(dataset, len(dataset)/4)
count = 0
processed = 0
for s, sequence in enumerate(dataset):
inVecs[s%8].put(sequence)
#count = count + len(sequence)/100
for j in xrange(0, len(sequence), batch_size):
#print "GIVING DATA"
count = count + 1
#inVecs[s%8].put(sequence[j:j+batch_size])
'''self.l1.release()
self.l2.release()
self.l3.release()
self.l4.release()
self.l5.release()
self.l6.release()
self.l7.release()
self.l8.release()'''
while processed != count:
if self.costQueue.qsize()>0:
cost = self.costQueue.get()
costs.append(cost)
processed = processed + 1
sums = []
items = self.weightQueue.qsize()
#while self.weightQueue.qsize() > 0:
# d = self.weightQueue.get()
# sums.append(d)
adder = [None,None,None,None,None,None,None,None]
for i in xrange(0,8):
q = self.weights[i].get()
for j in xrange(0,8):
if adder[j] == None:
adder[j] = q[j]
else:
adder[j] += q[j]
for i in range(0,8):
adder[i] /= float(8)
p = adder[0],adder[1],adder[2],adder[3],adder[4],adder[5],adder[6],adder[7]
if epoch != num_epochs:
self.weightQueue.put(p)
self.weightQueue2.put(p)
self.weightQueue3.put(p)
self.weightQueue4.put(p)
self.weightQueue5.put(p)
self.weightQueue6.put(p)
self.weightQueue7.put(p)
self.weightQueue8.put(p)
print 'Epoch %i/%i' % (epoch + 1, num_epochs),
print numpy.mean(costs)
#print profmode.print_summary()
sys.stdout.flush()
except KeyboardInterrupt:
print 'Interrupted by user.'
self.pro.terminate()
self.pro2.terminate()
self.pro3.terminate()
self.pro4.terminate()
self.pro5.terminate()
self.pro6.terminate()
self.pro7.terminate()
self.pro8.terminate()
self.weightQueue.close()
self.weightQueue2.close()
self.weightQueue3.close()
self.weightQueue4.close()
self.weightQueue5.close()
self.weightQueue6.close()
self.weightQueue7.close()
self.weightQueue8.close()
#print "NOW FETCHING",self.weightQueue.qsize()
p2 = self.args[0]['params']
#print "AFTER",len(p2)
print p2[0]==self.params[0]
self.W.set_value(p2[0])
self.bv.set_value(p2[1])
self.bh.set_value(p2[2])
self.Wuh.set_value(p2[3])
self.Wuv.set_value(p2[4])
self.Wvu.set_value(p2[5])
self.Wuu.set_value(p2[6])
self.bu.set_value(p2[7])
from scipy.io import wavfile
import numpy as np
THIS_DATA_DIR = dirname(realpath(__file__))
DOWNLOADED_ZIP = join(THIS_DATA_DIR, "dataset.zip")
DOWNLOADED_DIR = join(THIS_DATA_DIR, "dataset")
FILE_URL = "http://c4dm.eecs.qmul.ac.uk/rdr/bitstream/handle/123456789/13/Score-informed%20Piano%20Transcription%20Dataset.zip?sequence=1"
if __name__ == '__main__':
if not exists(DOWNLOADED_ZIP):
execute_bash("wget -O {path} {url}".format(url=FILE_URL,
path=DOWNLOADED_ZIP))
if exists(DOWNLOADED_DIR) and isdir(DOWNLOADED_DIR):
execute_bash("rm -rf %s" % (DOWNLOADED_DIR))
execute_bash("rm %s " % (join(THIS_DATA_DIR, "*.npy")))
makedirs(DOWNLOADED_DIR)
execute_bash("unzip %s -d %s" % (DOWNLOADED_ZIP, DOWNLOADED_DIR))
files = collect_files_with_ext(DOWNLOADED_DIR, ".wav")
for subpath, name in files:
if name.endswith(".wav") and "Chromatic" not in name:
sampling_rate, music = wavfile.read(subpath)
np.save(join(THIS_DATA_DIR, name.replace(".wav", ".npy")), music)
piece = midiread(str(subpath).replace(".wav", "_correct.mid"))
np.save(join(THIS_DATA_DIR, name.replace(".wav", ".mid.npy")),
piece.piano_roll)
execute_bash("rm -rf %s" % (DOWNLOADED_DIR))
execute_bash("rm -rf %s" % (DOWNLOADED_ZIP))
def train(self, files, batch_size=128, num_epochs=200):
def downsampling(sample):
# only keep the note on even index position (0, 2, 4 ...)
downsample = []
for i, s in enumerate(sample):
if i % 2 == 0:
downsample.append(s)
return downsample
def upsampling(sample, length):
'''
double each notes in sample
sample: the melody to be upsampled
length: the length of the original melody M, sample=Downsampling(M)
'''
upsample = []
for s in sample:
upsample.append(s)
if (len(upsample) >= length):
# upsampling melody length cannot longer than original melody length
break
upsample.append(s)
return upsample
def sampling(sample):
# to be one hot (to be 0,1)
ixes = []
for i, s in enumerate(sample):
n_step = []
for n in xrange(20):
ix = np.random.choice(range(88), p=s)
n_step.append(ix)
count = Counter(n_step)
ix = count.most_common(1)[0][0]
x = np.zeros((88, ))
x[ix] = 1
ixes.append(x)
return ixes
def accuracy(v, v_sample):
# ACC
accs = []
t = len(v)
n = len(v[0])
for time in range(t):
tp = 0 # true positive
fp = 0 # false positive
fn = 0 # false negative
for note in range(n):
if v[time][note] == 1 and v_sample[time][note] == 1:
tp += 1.
if v[time][note] == 0 and v_sample[time][note] == 1:
fp += 1.
if v[time][note] == 1 and v_sample[time][note] == 0:
fn += 1.
if tp + fp + fn != 0:
a = tp / (tp + fp + fn)
else:
a = 0
accs.append(a)
acc = numpy.mean(accs)
return acc
def generate():
# sampling procedure, generating music
# layer 4
generate_sharp_4 = self.generate_4()[0]
generate_sharp_4 = sampling(generate_sharp_4)
# layer 3
upsample_3 = upsampling(generate_sharp_4,
2 * len(generate_sharp_4) + 1)
generate_sharp_3 = self.generate_3(upsample_3, upsample_3)[0]
generate_sharp_3 = sampling(generate_sharp_3)
# layer 2
upsample_2 = upsampling(generate_sharp_3,
2 * len(generate_sharp_3) + 1)
generate_sharp_2 = self.generate_2(upsample_2, upsample_2)[0]
generate_sharp_2 = sampling(generate_sharp_2)
# layer 1
upsample_1 = upsampling(generate_sharp_2,
2 * len(generate_sharp_2) + 1)
generate_sharp_1 = self.generate_1(upsample_1, upsample_1)[0]
generate_sharp_1 = sampling(generate_sharp_1)
# layer 0
upsample_0 = upsampling(generate_sharp_1,
2 * len(generate_sharp_1) + 1)
generate_sharp_0 = self.generate_0(upsample_0, upsample_0)[0]
generate_sharp_0 = sampling(generate_sharp_0)
return generate_sharp_0
# load data
dataset = [
midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files
]
print('pyramid_mono, dataset=nottingham_mono, lr=%f, epoch=%i' %
(self.lr, num_epochs))
for epoch in range(num_epochs):
numpy.random.shuffle(dataset)
monitors_0 = []
accs_0 = []
monitors_1 = []
accs_1 = []
monitors_2 = []
accs_2 = []
monitors_3 = []
accs_3 = []
monitors_4 = []
accs_4 = []
for s, sequence in enumerate(dataset):
for i in range(0, len(sequence), batch_size):
batch_music = sequence[i:i + batch_size] # 128
# layer 0
downsample_0 = downsampling(batch_music) # 64
upsample_0 = upsampling(downsample_0,
len(batch_music)) # 128
# layer 1
downsample_1 = downsampling(downsample_0) # 32
upsample_1 = upsampling(downsample_1,
len(downsample_0)) # 64
# layer 2
downsample_2 = downsampling(downsample_1) # 16
upsample_2 = upsampling(downsample_2,
len(downsample_1)) # 32
# layer 3
downsample_3 = downsampling(downsample_2) # 8
upsample_3 = upsampling(downsample_3,
len(downsample_2)) # 16
# layer 0
sharp_0, monitor_0 = self.sharp_fun_0(
batch_music, upsample_0)
monitors_0.append(monitor_0)
accs_0.append(accuracy(batch_music, sampling(sharp_0)))
# layer 1
sharp_1, monitor_1 = self.sharp_fun_1(
downsample_0, upsample_1)
monitors_1.append(monitor_1)
accs_1.append(accuracy(downsample_0, sampling(sharp_1)))
# layer 2
sharp_2, monitor_2 = self.sharp_fun_2(
downsample_1, upsample_2)
monitors_2.append(monitor_2)
accs_2.append(accuracy(downsample_1, sampling(sharp_2)))
# layer 3
sharp_3, monitor_3 = self.sharp_fun_3(
downsample_2, upsample_3)
monitors_3.append(monitor_3)
accs_3.append(accuracy(downsample_2, sampling(sharp_3)))
# layer 4
if (len(downsample_3) == 1):
sharp_4, monitor_4 = self.sharp_fun_4(
downsample_3, downsample_3)
accs_4.append(accuracy(downsample_3,
sampling(sharp_4)))
else:
sharp_4, monitor_4 = self.sharp_fun_4(
downsample_3[1:],
downsample_3[:len(downsample_3) - 1])
accs_4.append(
accuracy(downsample_3[1:], sampling(sharp_3)))
monitors_4.append(monitor_4)
p = 'Epoch %i/%i layer0:LL %f ACC %f layer1:LL %f ACC %f layer2:LL %f ACC %f layer3:LL %f ACC %f ' \
'layer4:LL %f ACC %f time %s' % \
(epoch + 1, num_epochs, numpy.mean(monitors_0), numpy.mean(accs_0), numpy.mean(monitors_1), numpy.mean(accs_1),
numpy.mean(monitors_2), numpy.mean(accs_2), numpy.mean(monitors_3), numpy.mean(accs_3), numpy.mean(monitors_4),
numpy.mean(accs_4), datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"))
print(p)
if (epoch % 50 == 0 or epoch == num_epochs - 1):
piano_roll = generate()
midiwrite('sample/pyramid_mono_%i.mid' % (epoch), piano_roll,
self.r, self.dt)
save_dir = './'
### read MIDI
#data_dir = '../Data/Cirriculum/easy/'
#data_dir = '../biaxial-rnn-music-composition/music/'
files = os.listdir(data_dir)
files = [data_dir + f for f in files if '.mid' in f or '.MID' in f]
print files
dataset = []
for f in files:
try:
dataset.append(midiread(f, MIDI_RANGE, DT).piano_roll)
print "{} loaded".format(f)
except IndexError:
print "Skipping {}".format(f)
pass
print np.shape(dataset)
X = []
y = []
for song in dataset:
for i in range(0, len(song) - TICKS_PER_INPUT, STEP):
X.append(song[i:i + TICKS_PER_INPUT])
y.append(song[i + TICKS_PER_INPUT])
def load_midi_data(data_dir):
import midi.utils as utils
midi_data = utils.midiread(data_dir, dt=0.5)
return midi_data.piano_roll
# Compatible with Python 2.7. not 3.x
# Recursively load midi files, extract piano rolls and save as *.mid (file) and *.npy (matrix)
load_root = './MIDI_Data/'
save_root = './MIDI_Data_PianoRolls/'
for dirpath, dirs, files in os.walk(load_root):
for name in files:
if name.endswith('.mid'):
print dirpath, name
print dirpath.replace(load_root, save_root), name
load_dirpath = dirpath
save_dirpath = dirpath.replace(load_root, save_root)
load_filepath = os.path.join(load_dirpath, name)
save_filepath = os.path.join(save_dirpath, name)
# Read MIDI file
piano_roll = midiread(load_filepath).piano_roll
if not os.path.exists(save_dirpath):
os.makedirs(save_dirpath)
# Save the piano roll as MIDI
midiwrite(save_filepath, piano_roll=piano_roll)
# Save the piano roll as *.npy file
np.save(save_filepath.replace('.mid', '.npy'), piano_roll)
def main():
#--- import data ---#
sizeOfMiniBatch = 5 #how many tunes per miniBatch
noOfEpoch = 100
noOfEpochPerMB = 2
lengthOfMB = 100
sparseParam = np.float32(0.01) #increases with no. of cells
path = './Piano-midi.de/train-individual/hpps'
#path = './Piano-midi.de/train'
files = os.listdir(path)
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
#pitch range is from 21 to 109
dataset = [midiread((path + "/" + f), (21, 109),0.3).piano_roll.astype(theano.config.floatX) for f in files]
#check number of notes for each tune:
print(str([np.array(dataset[n]).shape[0] for n in np.arange(np.array(dataset).shape[0])]))
# set "silent" to zero in 1-hot format
for k in np.arange(np.array(dataset).shape[0]):
for n in np.arange(0,np.array(dataset[k]).shape[0],1):
if np.sum(dataset[k][n], dtype=theano.config.floatX) == 0 :
dataset[k][n][0] = np.float32(1.0)
#--- training with data ---#
myRNN4Music = RNN4Music(h1_length=176, h2_length=176, h3_length=176, io_length=88, R1=np.float32(0.001), R2=np.float32(0.001), R3=np.float32(0.001), Rout=np.float32(0.001))
#myRNN4Music.loadParameters('120_120_120_0_001_xEn_150epoch_hpps')
#myRNN4Music.loadParameters('120_120_120_0_001_sqr_150epoch_hpps')
myRNN4Music.loadParameters('176_176_176_0_001_xEn_L1_0_01_100epoch_hpps')
#myRNN4Music.saveParameters('176_176_176_0_001_xEn_L1_0_1_300epoch_hpps')
#myRNN4Music.train(dataset, noOfEpochPerMB, noOfEpoch, sizeOfMiniBatch, lengthOfMB, sparseParam)
#myRNN4Music.saveParameters('176_176_176_0_001_xEn_L1_0_01_100epoch_hpps')
#myRNN4Music.train(dataset, noOfEpochPerMB, noOfEpoch, sizeOfMiniBatch, lengthOfMB, sparseParam)
#myRNN4Music.saveParameters('176_176_176_0_001_xEn_L1_0_01_200epoch_hpps')
#myRNN4Music.train(dataset, noOfEpochPerMB, noOfEpoch, sizeOfMiniBatch, lengthOfMB, sparseParam)
#myRNN4Music.saveParameters('176_176_176_0_001_xEn_L1_0_01_300epoch_hpps')
#myRNN4Music.train(dataset, noOfEpochPerMB, noOfEpoch, sizeOfMiniBatch, lengthOfMB, sparseParam)
#myRNN4Music.saveParameters('176_176_176_0_001_xEn_L1_0_01_400epoch_hpps')
#--- plot some genearted tunes ---#
for baseSample in np.array([0, 20, 15, 31]):
exampleLength = 50
myRNN4Music.resetStates()
generatedTune = myRNN4Music.genMusic(np.float32(dataset[baseSample][0:exampleLength]), 300)
midiwrite('176_176_176_0_001_sqr_hpps150_' + str(baseSample) + '.mid', generatedTune[0], (21, 109),0.3)
#generatedTune[0] is the tune, generatedTune[1] is the probability at each iteration
#plot genearted probability
plt.figure(0 + baseSample*100)
plt.imshow(np.array(generatedTune[1][0:50,25:65]), origin = 'lower', extent=[25,65,0,50], aspect=0.5,
interpolation = 'nearest', cmap='gist_stern_r')
plt.title('probability of generated midi note piano-roll')
plt.xlabel('midi note')
plt.ylabel('sample number (time steps)')
plt.colorbar()##
#plot leading example for generation
plt.figure(1 + baseSample*100)
plt.imshow(np.transpose(dataset[baseSample]), origin='lower', aspect='auto',
interpolation='nearest', cmap=pylab.cm.gray_r)
plt.colorbar()
plt.title('original piano-roll')
plt.xlabel('sample number (time steps)')
plt.ylabel('midi note')
#plot generated tune
plt.figure(2 + baseSample*100)
plt.imshow(np.transpose(np.array(generatedTune[0][0:500])), origin='lower', aspect='auto',
interpolation='nearest', cmap=pylab.cm.gray_r)
plt.colorbar()
plt.title('generated piano-roll')
plt.xlabel('sample number (time steps)')
plt.ylabel('midi note')
plt.show()
def train(self, files, batch_size=100, num_epochs=200):
assert len(files) > 0, 'Training set is empty!' \
' (did you download the data files?)'
dataset = [
midiread(f, self.r,
self.dt).piano_roll.astype(theano.config.floatX)
for f in files
]
def accuracy(v, v_sample):
accs = []
t, n = v.shape
for time in range(t):
tp = 0 # true positive
fp = 0 # false positive
fn = 0 # false negative
for note in range(n):
if v[time][note] == 1 and v_sample[time][note] == 1:
tp += 1.
if v[time][note] == 0 and v_sample[time][note] == 1:
fp += 1.
if v[time][note] == 1 and v_sample[time][note] == 0:
fn += 1.
if tp + fp + fn != 0:
a = tp / (tp + fp + fn)
else:
a = 0
accs.append(a)
acc = numpy.mean(accs)
return acc
def sampling(sample): # 01
s = T.matrix()
b = rng.binomial(size=s.shape,
n=1,
p=s,
dtype=theano.config.floatX)
fun = theano.function(inputs=[s], outputs=[b])
return fun(sample)[0]
try:
print('rnn, dataset=Nottingham, lr=%f, epoch=%i' %
(self.lr, num_epochs))
for epoch in range(num_epochs):
numpy.random.shuffle(dataset)
costs = []
monitors = []
accs = []
for s, sequence in enumerate(dataset):
for i in range(0, len(sequence), batch_size):
if i + batch_size + 1 >= len(sequence):
break
v = sequence[i:i + batch_size]
targets = sequence[i + 1:i + batch_size + 1]
(cost, monitor,
sample) = self.train_function(v, targets)
costs.append(cost)
monitors.append(monitor)
sample = sampling(sample)
acc = accuracy(v, sample)
accs.append(acc)
p = 'Epoch %i/%i LL %f ACC %f Cost %f time %s' \
% (epoch + 1, num_epochs, numpy.mean(monitors), numpy.mean(accs), numpy.mean(costs),
datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"))
print(p)
if (epoch % 100 == 0 or epoch == num_epochs - 1):
piano_roll = sampling(self.generate_function()[0])
midiwrite('sample/rnn_%i.mid' % (epoch), piano_roll,
self.r, self.dt)
sys.stdout.flush()
except KeyboardInterrupt:
print('Interrupted by user.')
def load_midi_data(data_dir):
import midi.utils as utils
midi_data = utils.midiread(data_dir, dt=0.5)
return midi_data.piano_roll