def __init__(self,
model_path,
prediction_mode='sampling',
first_chords=[1, 3, 2, 1, 1, 3, 2, 1],
resample='none',
dim_factor=2,
temperature=1.0):
print('loading chord model ...')
self.model = keras.models.load_model(model_path)
self.model.reset_states()
self.embed_layer_output = K.function([self.model.layers[0].input],
[self.model.layers[0].output])
# self.embed_model = keras.ModModel(inputs=model.input,outputs=model.get_layer(.output)
self.embed_model = keras.models.Model(
inputs=self.model.input,
outputs=self.model.get_layer(index=0).output)
self.chord_to_index, self.index_to_chords = get_chord_dict()
self.prediction_mode = prediction_mode
self.temperature = temperature
self.resample = resample
self.dim_factor = dim_factor
self.song = []
for chord in first_chords[:-1]:
# print(chord)
self.model.predict(array([[chord]]))
self.song.append(chord)
chord = first_chords[-1]
self.song.append(chord)
self.current_chord = array([[chord]])
def __init__(self, model_path):
print('loading chord model ...')
model = keras.models.load_model(model_path)
model.reset_states()
self.embed_layer_output = K.function([model.layers[0].input], [model.layers[0].output])
self.chord_to_index, self.index_to_chords = get_chord_dict()
# Parameters for song generation:
BPM = 100
note_cap = 5
chord_temperature = 1
# Params for seed:
# length of the predicted song in bars:
num_bars = 64
# The first seed_length number of bars from the seed will be used:
seed_length = 4
#pred_song_length = 8*16-seed_length
with_seed = True
chord_to_index, index_to_chord = data_processing.get_chord_dict()
def sample_probability_vector(prob_vector):
# Sample a probability vector, e.g. [0.1, 0.001, 0.5, 0.9]
sum_probas = sum(prob_vector)
if sum_probas > note_cap:
prob_vector = (prob_vector / sum_probas) * note_cap
note_vector = np.zeros((prob_vector.size), dtype=np.int8)
for i, prob in enumerate(prob_vector):
note_vector[i] = np.random.multinomial(1, [1 - prob, prob])[1]
return note_vector
if not os.path.exists(LOG_DIR):
os.makedirs(LOG_DIR)
# extract embedding from chord model
# chord_model = "models/chords/1564233418-Shifted_False_Lr_1e-05_EmDim_10_opt_Adam_bi_False_lstmsize_512_trainsize_4_testsize_1_samples_per_bar8/model_Epoch100_4.pickle"
chord_model = "models/chords/1564241436-Shifted_False_Lr_1e-05_EmDim_10_opt_Adam_bi_False_lstmsize_512_trainsize_4_testsize_1_samples_per_bar8_dataset_data/Nottingham_unshifted_dataset/original/model_Epoch30_4.pickle"
model = keras.models.load_model(chord_model)
embeddings = model.layers[0].get_weights()
embeddings = np.array(embeddings)
# embeddings shape: (1, 50, 10)
embeddings = embeddings.reshape(-1, embeddings.shape[2])
# find the chord embeddings with the name
chords_with_name = {}
index_to_chords = get_chord_dict()[1]
total_data_num = embeddings.shape[0]
for i in range(total_data_num):
triad_tuple = index_to_chords[i]
if triad_tuple in chord_triad_tuple_to_chord_name:
chords_with_name[i] = chord_triad_tuple_to_chord_name[triad_tuple]
new_embeddings = []
for i in chords_with_name.keys():
new_embeddings.append(embeddings[i])
# high_dimensional_vectors = np.random.rand(50, 10)
# tensors_to_visualize = tf.Variable(high_dimensional_vectors, name="high_dim_vectors")
tensors_to_visualize = tf.Variable(np.array(new_embeddings),