song = model.generate(code, img=True)[0]
if mode == '2x':
codeA = sample(1)
codeA = np.reshape(np.tile(np.expand_dims(codeA, 1), (1, 128, 1)),
(-1, codeA.shape[-1]))
codeB = sample(1)
codeB = np.reshape(np.tile(np.expand_dims(codeB, 1), (1, 128, 1)),
(-1, codeB.shape[-1]))
code = np.concatenate([codeA, codeB, codeA, codeB, codeA], axis=0)
song = model.generate([code], code_img=True, img=True)[0]
print song[song.nonzero()].mean()
final = song > -0.3
coder.decode(final, speed=1.).write('midi', 'example.mid')
coder = NoteDurationCoder(first_voice=False)
note, dura = coder.encode(ms.converter.parse('example.mid'),
force=True)
coder.decode(note, 2).write('midi', 'quantized.mid')
import matplotlib.pyplot as plt
plt.subplot(211)
plt.imshow(song.T[::-1, :])
plt.subplot(212)
plt.imshow(final.T[::-1, :])
# plt.colorbar(orientation='horizontal')
plt.savefig('example.png')
plt.show()
# t, p = (song > 0.80).nonzero()
# song = []
learning_rate=1e-3,
iterations=2000,
continued=False,
only_train_quantized_rec=False,
vocab_size=128 + 1,
debug=False,
overwrite_workdir=True,
# workdir='./temp/SeqGAN_ablation/',
clip_norm=1.,
alpha=5e-3, # 5-->2, alpha*con_loss
beta=1.00,
gamma=1e-3) # gamma*q_loss
model = ContinuousSeqAE(hparam)
model.build()
# coder = ExampleCoder()
coder = NoteDurationCoder(normalize_key='C5', )
# single=True,
# first_voice=True)
try:
data = np.load('temp/easy.npz')['data']
except:
data = np.array(
map_dir(lambda fn: coder.encode(ms.converter.parse(fn))[0],
'./datasets/easymusicnotes/'))
np.savez('temp/easy.npz', data=data)
print(len(data), map(lambda x: len(x), data))
data = filter(lambda x: len(x) > hparam.timesteps, data)
data = map(remove_con_dup, data)
print(len(data), map(lambda x: len(x), data))
output_size=16,
timesteps=10,
iterations=5000,
learning_rate=1e-4,
continued=False,
overwrite_workdir=True,
workdir='./temp/DurationNet/',
)
mode = 'train'
mode = 'refine'
mode = 'oneshot'
model = RefineNet(hparam)
model.build()
coder = NoteDurationCoder(normalize_key='C5', first_voice=False)
try:
data = np.load('temp/easy_with_duration.npz')
notes = data['notes']
durations = data['durations']
except:
data = np.array(
map_dir(lambda fn: coder.encode(ms.converter.parse(fn)),
'./datasets/easymusicnotes/'))
notes, durations = zip(*data)
np.savez('temp/easy_with_duration.npz',
notes=notes,
durations=durations)
notes = filter(lambda x: len(x) > hparam.timesteps, notes)
basic_cell=rnn.GRUCell,
# hparam
D_lr=1e-4,
D_boost=0,
G_lr=1e-3,
G_k=5,
G_clipnorm=1.0,
# train
batch_size=25,
continued=False,
overwrite_workdir=True,
iterations=50000,
workdir='./temp/RhythmGAN/')
model = SeqGAN(hparam)
model.build()
coder = NoteDurationCoder(normalize_key='C5', first_voice=True)
try:
# raise Exception
data = np.load('temp/easy.npz')['data']
except:
data = np.array(
map_dir(lambda fn: coder.encode(ms.converter.parse(fn))[0],
'./datasets/easymusicnotes/'))
np.savez('temp/easy.npz', data=data)
print(len(data), map(lambda x: len(x), data))
data = filter(lambda x: len(x) > hparam.timesteps, data)
print(len(data), map(lambda x: len(x), data))
train_data, test_data = train_test_split(data,
from DeepSymphony.utils.Music21Coder import NoteDurationCoder
from keras.optimizers import Adam, SGD
import music21 as ms
def handle(fn):
return coder.encode(ms.converter.parse(fn))
if __name__ == '__main__':
timesteps = 128
batch_size = 32
mode = 'train'
mode = 'test'
coder = NoteDurationCoder(normalize_key='C5', resolution=1. / 16.)
try:
# raise Exception
data = np.load('temp/piano-midi_duration.npz')
notes = data['notes']
durations = data['durations']
except:
data = np.array(map_dir(handle, './datasets/piano-midi.de/', cores=8))
print map(lambda x: len(x[0]), data)
data = filter(lambda x: len(x[0]) > 0, data)
print map(lambda x: len(x[0]), data)
notes, durations = zip(*data)
np.savez('temp/piano-midi_duration.npz',
notes=notes,
durations=durations)