def __init__(self, bar_embedder, context_size,
enc_lstm_size, dec_lstm_size,
enc_use_meta=False, dec_use_meta=False, compile_now=False):
self.n_voices = 6
prev_bars = [Input(shape=(None,), name="context_" + str(i))
for i in range(context_size)]
if enc_use_meta or dec_use_meta:
meta_cat = Input(shape=(None,), name="metaData")
embeddings = [bar_embedder(pb) for pb in prev_bars]
embed_size = bar_embedder.embedding_size
if enc_use_meta:
embeddings = [Concat([emb, meta_cat]) for emb in embeddings]
embed_size = bar_embedder.embedding_size + self.n_voices
embeddings_stacked = Lambda(lambda ls: K.stack(ls, axis=1),
output_shape=(context_size,
embed_size)
)(embeddings)
# encode
embeddings_processed = LSTM(enc_lstm_size)(embeddings_stacked)
encoded_size = enc_lstm_size
# decode
if dec_use_meta:
encoded_size += self.n_voices
embeddings_processed = Concat([embeddings_processed, meta_cat])
repeated = Lambda(self._repeat, output_shape=(None, encoded_size))\
([prev_bars[0], embeddings_processed])
decoded = LSTM(dec_lstm_size,
return_sequences=True, name='dec_lstm')(repeated)
preds = TimeDistributed(Dense(bar_embedder.vocab_size, activation='softmax'),
name='softmax_layer')(decoded)
if enc_use_meta or dec_use_meta:
super().__init__(inputs=[*prev_bars, meta_cat], outputs=preds)
else:
super().__init__(inputs=prev_bars, outputs=preds)
self.params = [context_size, enc_lstm_size, dec_lstm_size,
enc_use_meta, dec_use_meta]
self.use_meta = enc_use_meta or dec_use_meta
if compile_now:
self.compile_default()
def __init__(self, melody_encoder,
rhythm_embed_size, dec_lstm_size, V,
enc_use_meta=False, dec_use_meta=False,
compile_now=False):
self.n_voices = 9
self.use_meta = enc_use_meta or dec_use_meta
m = melody_encoder.m
prev_melodies = Input(shape=(None, m), name="contexts")
bar_embedding = Input(shape=(rhythm_embed_size, ), name="bar_rhythm_embedded")
if self.use_meta:
meta_cat = Input(shape=(self.n_voices,), name="metaData")
lead = Input(shape=(None, m), name="lead")
lead_enc = self.get_lead_encoder(melody_encoder)
#encode
processed = melody_encoder(prev_melodies)
processed_with_rhythms = Concat([processed, bar_embedding])
# decode
if dec_use_meta:
processed_with_rhythms = Concat([processed_with_rhythms, meta_cat])
lead_processed = lead_enc(lead)
processed_with_lead = Concat([processed_with_rhythms, lead_processed])
proc_repeated = RepeatVector(m)(processed_with_lead)
lstm_outputs = LSTM(dec_lstm_size, return_sequences=True)(proc_repeated)
preds = TimeDistributed(Dense(V, activation="softmax"))(lstm_outputs)
self.params = [melody_encoder.params, rhythm_embed_size,
dec_lstm_size, V, enc_use_meta, dec_use_meta]
# self.params = [V, rhythm_embed_size,
# dec_lstm_size,
# enc_use_meta, dec_use_meta]
if self.use_meta:
super().__init__(inputs=[prev_melodies, bar_embedding,
meta_cat, lead],
outputs=preds, name=repr(self))
else:
super().__init__(inputs=[prev_melodies, bar_embedding, lead],
outputs=preds, name=repr(self))
self.encoder = melody_encoder
if compile_now:
self.compile_default()
def __init__(self,
m,
V,
rhythm_embed_size,
conv_f,
conv_win_size,
enc_lstm_size,
dec_lstm_1_size,
dec_lstm_2_size,
meta_len,
compile_now=False):
self.n_voices = 6
prev_melodies = Input(shape=(None, m), name="contexts")
bar_embedding = Input(shape=(rhythm_embed_size, ),
name="bar_rhythm_embedded")
meta_data = Input(shape=(meta_len, ), name="metaData")
meta_cat = Dense(self.n_voices, activation="softmax")(meta_data)
#encode
conved = Conv1D(filters=conv_f,
kernel_size=conv_win_size)(prev_melodies)
meta_repeated = Lambda(self._repeat,
output_shape=(None, self.n_voices))(
[conved, meta_cat])
conved_with_meta = Concat([conved, meta_repeated], axis=-1)
processed = LSTM(enc_lstm_size)(conved_with_meta)
# decode
proc_rhythm_concat = Concat([processed, bar_embedding])
proc_repeated = RepeatVector(m)(proc_rhythm_concat)
lstm1_outputs = LSTM(dec_lstm_1_size,
return_sequences=True)(proc_repeated)
lstm2_outputs = LSTM(dec_lstm_2_size,
return_sequences=True)(lstm1_outputs)
preds = TimeDistributed(Dense(V, activation="softmax"))(lstm2_outputs)
super().__init__(inputs=[prev_melodies, bar_embedding, meta_data],
outputs=preds)
self.params = [
m, V, rhythm_embed_size, conv_f, conv_win_size, enc_lstm_size,
dec_lstm_1_size, dec_lstm_2_size, meta_len
]
if compile_now:
self.compile_default()
def __init__(self, melody_encoder, dense_size, V, compile_now=False):
m = melody_encoder.m
self.n_voices = 10
root_note = Input(shape=(1, ), name="root_note")
melody_context = Input(shape=(None, m), name="bar_melody")
meta_embedded = Input(shape=(self.n_voices, ), name="meta_embedded")
root_encoded = Dense(dense_size)(root_note)
context_encoded = melody_encoder(melody_context)
inputs_concat = Concat([root_encoded, context_encoded, meta_embedded])
decoded = Dense(dense_size)(inputs_concat)
preds = Dense(V, activation="softmax")(decoded)
super().__init__(inputs=[root_note, melody_context, meta_embedded],
outputs=preds)
self.params = [dense_size, V]
self.melody_encoder = melody_encoder
if compile_now:
self.compile_default()
def __init__(self, rhythm_encoder, dec_lstm_size, V,
enc_use_meta=False, dec_use_meta=False, compile_now=False):
self.n_voices = 8
context_size = rhythm_encoder.context_size
encoded_size = rhythm_encoder.encoding_size
bar_embedder = rhythm_encoder.bar_embedder
bar_embed_size = rhythm_encoder.bar_embedder.embedding_size
prev_bars = [Input(shape=(None,), name="context_" + str(i))
for i in range(context_size)]
if enc_use_meta or dec_use_meta:
meta_cat = Input(shape=(None,), name="metaData")
lead = Input(shape=(None, ), name="lead")
lead_embedded = bar_embedder(lead)
# encode
embeddings_processed = rhythm_encoder(prev_bars)
# decode
if dec_use_meta:
encoded_size += self.n_voices
encoded_size += bar_embed_size
embeddings_processed = Concat([embeddings_processed, meta_cat, lead_embedded])
repeated = Lambda(self._repeat, output_shape=(None, encoded_size))\
([prev_bars[0], embeddings_processed])
decoded = LSTM(dec_lstm_size,
return_sequences=True, name='dec_lstm')(repeated)
preds = TimeDistributed(Dense(V, activation='softmax'),
name='softmax_layer')(decoded)
self.params = [rhythm_encoder.params, dec_lstm_size, V,
enc_use_meta, dec_use_meta]
# self.params = [V, context_size, dec_lstm_size,
# enc_use_meta, dec_use_meta]
self.use_meta = enc_use_meta or dec_use_meta
if enc_use_meta or dec_use_meta:
super().__init__(inputs=[*prev_bars, meta_cat, lead], outputs=preds,
name=repr(self))
else:
super().__init__(inputs=[prev_bars, lead], outputs=preds,
name=repr(self))
if compile_now:
self.compile_default()
def inception_3b(input_t, fc1, fc2_in, fc2_out, fc3_in, fc3_out, fc4_out):
layer_1x1 = Conv2D(fc1, (1,1), padding='same', activation='relu')(input_t)
layer_3x3 = Conv2D(fc2_in, (1,1), padding='same', activation='relu')(input_t)
layer_3x3 = Conv2D(fc2_out, (3,3), padding='same', activation='relu')(layer_3x3)
layer_5x5 = Conv2D(fc3_in, (1,1), padding='same', activation='relu')(input_t)
layer_5x5 = Conv2D(fc3_out, (5,5), padding='same', activation='relu')(layer_5x5)
layer_pool = MaxPooling2D((3,3), strides=(1,1), padding='same')(input_t)
layer_pool = Conv2D(fc4_out, (1,1), padding='same', activation='relu')(layer_pool)
output = Concat([layer_1x1, layer_3x3, layer_5x5, layer_pool], axis = -1)
return output
def __init__(self, bar_embedder, context_size,
enc_lstm_size, dec_lstm_size, meta_len, compile_now=False):
self.num_calls = 0
self.n_voices = 6
prev_bars = [Input(shape=(None,), name="context_" + str(i))
for i in range(context_size)]
meta_data = Input(shape=(meta_len,), name="metaData")
meta_cat = Dense(self.n_voices, activation="softmax")(meta_data)
# embed
embeddings_with_meta = [Concat([bar_embedder(pb), meta_cat]) for pb in prev_bars]
embeddings_stacked = Lambda(lambda ls: K.stack(ls, axis=1),
output_shape=(context_size,
bar_embedder.embedding_size + self.n_voices)
)(embeddings_with_meta)
# encode
embeddings_processed = LSTM(enc_lstm_size)(embeddings_stacked)
# decode
repeated = Lambda(self._repeat, output_shape=(None, enc_lstm_size))\
([prev_bars[0], embeddings_processed])
decoded = LSTM(dec_lstm_size,
return_sequences=True, name='dec_lstm')(repeated)
pred = TimeDistributed(Dense(bar_embedder.vocab_size, activation='softmax'),
name='softmax_layer')(decoded)
super().__init__(inputs=[*prev_bars, meta_data], outputs=pred)
self.params = [context_size, enc_lstm_size, dec_lstm_size, meta_len]
if compile_now:
self.compile_default()
x13 = Conv2D(32, 3, activation="relu", padding="same")(x12) x1 = Dropout(0.15 * dropout)(x13) x21 = Conv2D(8, 5, activation="relu", padding="same")(normal) xmp2 = MaxPooling2D(pool_size=(1, 2))(x21) x22 = Conv2D(16, 5, activation="relu", padding="same")(xmp2) x23 = Conv2D(32, 5, activation="relu", padding="same")(x22) x2 = Dropout(0.15 * dropout)(x23) x31 = Conv2D(8, 5, activation="relu", padding="same")(normal) xmp3 = MaxPooling2D(pool_size=(1, 2))(x31) x32 = Conv2D(16, 3, activation="relu", padding="same")(xmp3) x33 = Conv2D(32, 1, activation="relu", padding="same")(x32) x3 = Dropout(0.15 * dropout)(x33) inception = Concat([x1, x2, x3], axis=3) x = Conv2D(32, 6, activation="relu", padding="same")(inception) x = MaxPooling2D(pool_size=(1, 2))(x) x = Dropout(0.3 * dropout)(x) x = Conv2D(32, 5, activation="relu", padding="same")(x) x = Dropout(0.3 * dropout)(x) x = Conv2D(32, 4, activation="relu", padding="same")(x) x = MaxPooling2D(pool_size=(1, 2))(x) x = Dropout(0.3 * dropout)(x) x = Conv2D(32, 3, activation="relu", padding="same")(x) x = Dropout(0.3 * dropout)(x)