note_seq.sequence_proto_to_midi_file(x1, 'input_mel.mid')
note_seq.sequence_proto_to_midi_file(y1, 'target_bass.mid')
#%%
# Set seed
np.random.seed(42)
path_to_midi_dir = folder_name
#instruments_to_extract = (54, 34) # voice and bass
instruments_to_extract = (0, 1) # voice and bass
# Hyper-parameters
num_epochs = 200
training_set, validation_set, test_set, tokenizer \
= create_dataset_from_midi(path_to_midi_dir, instruments_to_extract, print_info=True)
#%%
encoder_decoder = tokenizer.encoder_decoder
num_sequences = tokenizer.song_count
vocab_size = tokenizer.vocab_size
#training_set, validation_set, test_set, word_to_idx, idx_to_word, num_sequences, vocab_size = load_dummy_dataset(True)
# Initialize a new LSTM network
net = MyRecurrentNet(vocab_size)
training_loss, validation_loss = train_lstm(net, num_epochs, training_set,
validation_set, vocab_size,
encoder_decoder)
# Plot training and validation loss
length_per_seq = tokenizer_kwargs[
'steps_per_quarter'] * 4 * tokenizer_kwargs['split_in_bar_chunks']
print('length_per_seq: ', length_per_seq)
# Dataset creation, splitting and batching parameters
dataset_split_kwargs = {
'p_train': 0.6,
'p_val': 0.3,
'p_test': 0.0,
'batch_size': 1,
'eval_batch_size': 1
}
train, val, test, t = create_dataset_from_midi(sequences,
lead_instrument,
accomp_instrument,
print_info=False,
**tokenizer_kwargs,
**dataset_split_kwargs)
encoder_decoder = t.encoder_decoder
num_sequences = t.song_count
#LSTM hyperparameters
vocab_size = t.vocab_size
num_epochs = 30 #100
learning_rate = 1e-4
mel_notes, bass_notes = get_histograms_from_dataloader(test, vocab_size=vocab_size, \
plot=False)
#print('Melody histogram: ', Counter(mel_notes))
import torch.optim as optim
import utils.paths as paths
# Set seed such that we always get the same dataset
np.random.seed(42)
instruments = [0, 1]
lead_instrument = ('melody', instruments[0])
accomp_instrument = ('bass', instruments[1])
split_in_bar_chunks = 8
# Hyper-parameters
num_epochs = 100
training_set, validation_set, test_set, tokenizer \
= create_dataset_from_midi(paths.midi_dir, lead_instrument, accomp_instrument, split_in_bar_chunks, print_info=True)
encoder_decoder = tokenizer.encoder_decoder
num_sequences = tokenizer.song_count
vocab_size = tokenizer.vocab_size
# Initialize a new LSTM network
net = MusicLSTMNet(vocab_size)
training_loss, validation_loss = train_lstm(net, num_epochs, training_set,
validation_set, vocab_size,
encoder_decoder)
torch.save(net, paths.model_serialized_dir + 'music_lstm.pt')
# Plot training and validation loss
epoch = np.arange(len(training_loss))
plt.figure()
import utils.paths as paths
# Set seed such that we always get the same dataset
np.random.seed(42)
instruments = [0,1]
lead_instrument = ('melody',instruments[0])
accomp_instrument = ('bass',instruments[1])
split_in_bar_chunks = 8
# Hyper-parameters
num_epochs = 100
train, val, test, t = create_dataset_from_midi(paths.midi_dir_small,
lead_instrument,
accomp_instrument,
split_in_bar_chunks,
print_info=True)
encoder_decoder = t.encoder_decoder
num_sequences = t.song_count
# Hyperparameters
num_epochs = 100
batch_size = 16
vocab_size = t.vocab_size
#%%
length_per_seq = tokenizer_kwargs['steps_per_quarter'] * 4 * tokenizer_kwargs[
'split_in_bar_chunks']
print('length_per_seq: ', length_per_seq)
# Dataset creation, splitting and batching parameters
dataset_split_kwargs = {
'p_train': 0.7,
'p_val': 0.3,
'p_test': 0.0,
'batch_size': 16,
'eval_batch_size': 16
}
train, val, test, t = create_dataset_from_midi(dataset_path,
lead_instrument,
accomp_instrument,
print_info=True,
**tokenizer_kwargs,
**dataset_split_kwargs)
encoder_decoder = t.encoder_decoder
num_sequences = t.song_count
#LSTM hyperparameters
vocab_size = t.vocab_size
num_epochs = 20
learning_rate = 0.7e-4
dropout_prob = 0.5
num_layers = 2
hidden_size = 256
print('train Histogram')