input = dataset[0][ind][start:start + hparam.timesteps]
label = dataset[1][ind][start:start + hparam.timesteps]
inputs.append(input)
labels.append(label)
return np.array(inputs), np.array(labels)
return fetch_data
fetch_train_data = fetch_data_g(train_set)
fetch_test_data = fetch_data_g(test_set)
fetch_one_data = fetch_data_g(([train_set[0][1]], [train_set[1][1]]))
if mode == 'train':
model.train(fetch_one_data,
fetch_test_data,
continued=hparam.continued)
if mode == 'refine':
ind = train_ind[1]
song = notes[ind]
dura = durations[ind]
pred_dura = model.predict([song])[0]
print pred_dura
coder.decode(song, 2).write('midi', 'quantized.mid')
coder.decode(song, dura).write('midi', 'truth.mid')
coder.decode(song, pred_dura).write('midi', 'refine.mid')
if mode == 'oneshot':
song = np.load('temp/oneshot_example.npz')['song']
pred_dura = model.predict([song], max=8)[0]
coder.decode(song, pred_dura).write('midi', 'theone.mid')
# for c in code:
# print np.histogram(c)
if mode == 'rec':
seq = train_data[1].copy()
batch = continuous_sample(seq,
hparam.timesteps,
stride=hparam.timesteps)
ori = code = model.encode(batch, quantized=True)
print code
rec = model.generate(code, quantized=True).flatten()
print seq.shape
print batch.shape
print rec.shape
coder.decode(seq, 2).write('midi', 'truth.mid')
coder.decode(rec, 2).write('midi', 'rec.mid')
if mode == 'plot':
seq = train_data[0].copy()
batch = continuous_sample(seq,
hparam.timesteps,
stride=hparam.timesteps)
code = model.encode(batch, quantized=True) # first layer
print code
from sklearn.manifold import TSNE
tsne = TSNE(2)
code = tsne.fit_transform(code)
print code.shape
import matplotlib.pyplot as plt
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)
print 'done'