def test_with_mel_lstm(self):
"""
Test that a very short training runs without exceptions
"""
model = mel_lstm.get_model()
train_generator = RandomMidiDataGenerator(
real_midis[::2], mel_lstm.preprocess, mapper, 2, 100
)
test_generator = RandomMidiDataGenerator(
real_midis[1::2], mel_lstm.preprocess, mapper, 2, 100
)
self.train(model, train_generator, test_generator)
plt.title("Mel LSTM, precomputed - Accuracy")
plt.savefig("test_mel_lstm_training_precomputed.png")
def test_with_seq_lstm(self):
"""
Test that a very short training runs without exceptions
"""
model = seq_lstm.get_model(mapper.dims)
train_generator = RandomMidiDataGenerator(
real_midis[::2], seq_lstm.make_preprocessor(mapper), mapper, 2, 100
)
test_generator = RandomMidiDataGenerator(
real_midis[1::2], seq_lstm.make_preprocessor(mapper), mapper, 2, 100
)
self.train(model, train_generator, test_generator)
plt.title("Seq LSTM, on generator - Accuracy")
plt.savefig("test_seq_lstm_training_generator.png")
def test_with_mel_lstm(self):
generator = RandomMidiDataGenerator(real_midis, mel_lstm.preprocess, mapper, 2)
got_a_batch = False
for x, y in generator():
got_a_batch = True
self.assertFalse(np.any(np.isnan(x)))
self.assertFalse(np.any(np.isnan(y)))
if generator.epochs >= 1:
break
self.assertTrue(got_a_batch)
import pretty_midi
from glob import glob
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense
from tensorflow.keras.layers import TimeDistributed
from musicrl.midi2vec import MidiVectorMapper
from musicrl.render import *
from musicrl.random_generator import generate_random_midi, resemble_midi
from musicrl.data import RandomMidiDataGenerator
from musicrl import mel_lstm
filepaths = list(glob("maestro-v2.0.0/2008/**.midi"))
real_midis = [pretty_midi.PrettyMIDI(i) for i in filepaths]
mapper = MidiVectorMapper(real_midis)
model = mel_lstm.get_model()
train_generator = RandomMidiDataGenerator(real_midis[::2], mel_lstm.preprocess,
mapper, 10, 8000)
test_generator = RandomMidiDataGenerator(real_midis[1::2], mel_lstm.preprocess,
mapper, 10, 1000)
test_data = test_generator.compute_batch()
x_train, y_train = train_generator.compute_batch()
history = model.fit(x_train, y_train, 10, validation_data=test_data, epochs=30)
mel_lstm.plot_history(history)
plt.savefig("test_mel_lstm_training.png")
from musicrl.midi2vec import MidiVectorMapper
from musicrl.render import *
from musicrl.random_generator import generate_random_midi, resemble_midi
from musicrl.data import RandomMidiDataGenerator
from musicrl import mel_lstm
filepaths = list(glob("maestro-v2.0.0/2008/**.midi"))
real_midis = [pretty_midi.PrettyMIDI(i) for i in filepaths]
mapper = MidiVectorMapper(real_midis)
model = mel_lstm.get_model()
train_generator = RandomMidiDataGenerator(
real_midis[:50], mel_lstm.preprocess, mapper, 4
)
test_generator = RandomMidiDataGenerator(
real_midis[50:60], mel_lstm.preprocess, mapper, 10, max_num_timeframes=1000
)
history = model.fit_generator(
train_generator(),
validation_data=test_generator(),
steps_per_epoch=train_generator.steps_per_epoch,
epochs=2,
validation_steps=test_generator.steps_per_epoch,
validation_freq=1,
verbose=1,
)
mel_lstm.plot_history(history)