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))
train_data, test_data = train_test_split(data,
test_size=0.22,
random_state=32)
def fetch_tri_data_g(dataset):
def fetch_tri_data(batch_size):
src_seqs, nei_seqs, int_seqs = [], [], []
# model.compile('Adam', 'MSE')
model.compile(Adam(5e-5), 'categorical_crossentropy',
['categorical_accuracy'])
model.load_weights('temp/duration.h5')
model.fit_generator(fetch_data(),
steps_per_epoch=20,
epochs=10,
verbose=1)
model.save('temp/duration.h5')
if mode == 'test':
model = load_model('temp/duration.h5')
model.summary()
coder = NoteDurationCoder(resolution=1. / 16.)
notes, duras = coder.encode(ms.converter.parse('example.mid'),
force=True)
# notes, duras = coder.encode(
# ms.converter.parse('datasets/piano-midi.de/schumann/schum_abegg.mid'),
# force=True)
print notes
print notes.shape
duration = model.predict([np.array([notes]),
np.array([duras])])[0].argmax(-1) + 1
print duration - duras
print 'old', duras
print 'new', duration
print duration[-1]
print 'sqr diff', ((duras - duration)**2).sum()
coder.decode(notes, duration).write('midi', 'example.d.mid')
# print zip(notes, duras)