def data_generator():
batch_size = 32
while True:
x = []
y = []
h = []
for _ in range(batch_size):
ind = np.random.randint(len(data))
start = np.random.randint(data[ind].shape[0] - LEN - 1)
x.append(data[ind][start:start + LEN])
y.append(data[ind][start + 1:start + LEN + 1])
# global hist
# h.append(hist[ind])
# patch hist
h_ = histogram_onehot(x[-1], coder.code_to_name, NOTE_NUMBER)
h.append(norm(h_, H_TEMP))
x = np.array(x)
y = np.array(y)
h = np.array(h)
# sparse to complex
x = to_categorical(x.flatten(), num_classes=DIM_IN).\
reshape(x.shape[0], x.shape[1], DIM_IN)
x[:, :, -len(NOTE_NUMBER):] = h[:, None, :]
y = y.reshape(x.shape[0], x.shape[1], 1)
yield (x, y)
model = define_stackrnn(LEN, DIM_IN, DIM_OUT)
D = []
for _ in range(TEST):
ind = RNG.randint(len(data))
start = RNG.randint(data[ind].shape[0] - LEN - 1)
x = np.array(data[ind][start:start + LEN])
h_pre = histogram_onehot(x[:PREFIX], coder.code_to_name, NOTE_NUMBER)
h_ans = histogram_onehot(x[-ANSWER:], coder.code_to_name, NOTE_NUMBER)
x = to_categorical(x[:PREFIX], num_classes=DIM_IN).\
reshape(PREFIX, DIM_IN)
model.reset_generator()
song = model.generate(
seed=RNG.randint(1e+9),
length=ANSWER,
prefix=x[:PREFIX],
verbose=0,
addition=norm(h_ans, TEMPERATURE) if HIST else None)
hist_res = histogram_onehot(song, coder.code_to_name, NOTE_NUMBER)
d = JS_divergence(h_ans, hist_res)
print h_ans
print hist_res
print d
if d != np.inf:
D.append(d)
print np.mean(D)
timespan=LEN,
input_dim=DIM_IN,
output_dim=DIM_OUT,
cells=[512, 512, 512],
)
model.build()
# model.model.load_weights('temp/stackedrnn_hist.h5')
# model.train(data_generator(),
# opt=8e-4,
# steps_per_epoch=30,
# epochs=300,
# save_path='temp/stackedrnn_hist.h5')
C_scale = np.array([2, 0, 1, 0, 1, 1, 0, 1, 0, 1, 0, 1])
model.build_generator('temp/stackedrnn_hist.h5')
res = model.generate(seed=64, length=3000, addition=norm(C_scale, ))
events = [
coder.code_to_name(note)
for note in res[:, :DIM_IN - len(NOTE_NUMBER)].argmax(1)
]
np.set_printoptions(precision=2)
res_hist = histogram(events, NOTE_NUMBER)
res_hist /= res_hist.sum()
print res_hist
print hist[0]
mid = Song()
track = mid.add_track()
# the ``useful'' notes are more than usual
# therefore use a smaller _MIDO_TIME_SCALE
input_dim=DIM_IN,
output_dim=DIM_OUT,
cells=[512, 512, 512],
)
model.build()
# model.model.load_weights('temp/stackedrnn_hist_easy.h5')
# model.train(data_generator(),
# opt=1e-3,
# steps_per_epoch=30,
# epochs=300,
# save_path='temp/stackedrnn_hist_easy.h5')
model.build_generator('temp/stackedrnn_hist_easy.h5')
res = model.generate(seed=64,
length=3000,
addition=norm(hist[1], H_TEMP),
verbose=0)
events = [
coder.code_to_name(note)
for note in res[:, :DIM_IN - len(NOTE_NUMBER)].argmax(1)
]
np.set_printoptions(precision=2)
res_hist = histogram(events, NOTE_NUMBER)
res_hist /= res_hist.sum()
print res_hist
print hist[0]
mid = Song()
track = mid.add_track()
# the ``useful'' notes are more than usual