def process_file(self, source, target):
#self.n_ins = 240
with open(source, 'rb') as i:
reader = csv.reader(i)
(before, chords) = util.list_spectrum_data(reader, components=self.n_ins + 60, allow_no_chord=True)
before = self.getFirstColumns(before, 0, self.n_ins)
result = self.through_sda(self.model, before)
with open(target, 'wb') as o:
writer = csv.writer(o)
writer.writerows(result)
def process_file(self, source, target):
subnotes = 1
with open(source, 'rb') as i:
reader = csv.reader(i)
(before, chords) = util.list_spectrum_data(reader,
components=self.n_ins + 12 * subnotes, allow_no_chord=True)
result = None
for offset in range(12):
#data = numpy.asmatrix(before, dtype=theano.config.floatX)
data = before
# if extra_octave:
start, end = offset * subnotes, offset * subnotes + self.n_ins
data = data[:,start:end]
# else:
# data = numpy.roll(data, -offset * subnotes, axis=1)
temp = self.through_sda(self.model, data)
temp = numpy.roll(temp, offset * subnotes, axis=1)
if result == None:
result = numpy.zeros(temp.shape)
result = numpy.add(result, temp)
# result = self.through_sda(self.model, before)
# s = result.shape[0]
# s = 0
# ss = numpy.zeros(shape=[s, s])
# for i, v1 in enumerate(result):
# for j, v2 in enumerate(result):
# ss[i][j] = numpy.linalg.norm(self.normalize(v1) - self.normalize(v2))
# for i in range(s):
# eps = numpy.sort(ss[i])[int(s * 0.03)]
# for j in range(s):
# if ss[i][j] > eps:
# ss[i][j] = 1
# result1 = numpy.zeros(shape = result.shape)
# for i in range(s):
# sumW = 0
# for j in range(s):
# w = 1 - ss[i][j]
# result1[i] = numpy.add(result1[i], numpy.multiply(result[j], w))
# sumW = sumW + w
# result1[i] = numpy.multiply(result1[i], 1.0 / sumW)
with open(target, 'wb') as o:
writer = csv.writer(o)
writer.writerows(result)
def read_data():
with open(train_file, 'rb') as f:
reader = csv.reader(f)
(array, chords) = util.list_spectrum_data(reader, components=ins, allow_non_majmin=False)
# (array, chords) = util.shuffle_2(array, chords)
# array = prepare_data(array)
# chords = prepare_chords(chords)
train = int(0.7 * len(array))
test = int(0.9 * len(array))
train_array = array[:train]
train_chords = chords_to_array(chords[:train])
test_array = array[train:test]
test_chords = chords_to_array(chords[train:test])
valid_array = array[test:]
valid_chords = chords_to_array(chords[test:])
print 'finished reading data'
return [[train_array, train_chords], [test_array, test_chords], \
[valid_array, valid_chords]]
def read_data(self):
print 'reading data from ' + self.train_file
with open(self.train_file, 'rb') as f:
reader = csv.reader(f)
(array, chords) = list_spectrum_data(reader, components=self.ins)
array = self.prepare_data(array)
chords = self.prepare_chords(chords)
# array = numpy.asarray(array, dtype=theano.config.floatX)
chords = numpy.asarray(chords, dtype=theano.config.floatX)
train = int(0.7 * len(array))
test = int(0.85 * len(array))
tr = numpy.copy(array[:train])
tr_ch = numpy.copy(chords[:train])
train_array = theano.shared(array[:train], borrow = True)
train_chords = theano.shared(chords[:train], borrow = True)
test_array = theano.shared(array[train:test], borrow = True)
test_chords = theano.shared(chords[train:test], borrow = True)
valid_array = theano.shared(array[test:], borrow = True)
valid_chords = theano.shared(chords[test:], borrow = True)
shuffle_2_numpy(tr, tr_ch)
train_shuffled = theano.shared(tr, borrow = True)
chords_shuffled = theano.shared(tr_ch, borrow = True)
if self.recurrent_layer >= 0:
return [[train_shuffled, chords_shuffled], [test_array, test_chords], \
[valid_array, valid_chords], [train_array, train_chords]]
else:
del train_array
del train_chords
return [[train_shuffled, chords_shuffled], [test_array, test_chords], \
[valid_array, valid_chords]]