def from_network_parameters(cls,
context_size,
melody_bar_len,
meta_len,
bar_embed_params,
rhythm_net_params,
melody_net_params,
generation=False,
compile_now=True):
bar_embedder = BarEmbedding(*bar_embed_params, compile_now=False)
rhythm_net = RhythmNetwork(bar_embedder,
*rhythm_net_params,
compile_now=False)
melody_net = MelodyNetwork(*melody_net_params, compile_now=False)
combined_net = cls(context_size,
melody_bar_len,
meta_len,
bar_embedder,
rhythm_net,
melody_net,
generation=generation,
compile_now=compile_now)
return combined_net
def from_saved_custom(cls, save_dir, generation=False, compile_now=True):
with open(save_dir + "/parameters", "r") as handle:
param_dict = json.load(handle)
bar_embedder = BarEmbedding(*param_dict["bar_embed_params"],
compile_now=False)
rhythm_net = RhythmNetwork(bar_embedder,
*param_dict["rhythm_net_params"],
compile_now=False)
melody_net = MelodyNetwork(*param_dict["melody_net_params"],
compile_now=False)
bar_embedder.load_weights(save_dir + "/bar_embedding_weights")
rhythm_net.load_weights(save_dir + "/rhythm_net_weights")
melody_net.load_weights(save_dir + "/melody_net_weights")
combined_net = cls(param_dict["context_size"],
param_dict["melody_bar_len"],
param_dict["meta_len"],
bar_embedder,
rhythm_net,
melody_net,
generation=generation,
compile_now=compile_now)
return combined_net
conv_win_size = 3
melody_enc_lstm_size = 52
melody_dec_lstm_1_size = 32
melody_dec_lstm_2_size = 32
meta_data_len = 10
# INDIVIDUAL NETS
be = BarEmbedding(V=V_rhythm,
beat_embed_size=beat_embed_size,
embed_lstm_size=embed_lstm_size,
out_size=out_size)
rhythm_net = RhythmNetwork(bar_embedder=be,
context_size=context_size,
enc_lstm_size=rhythm_enc_lstm_size,
dec_lstm_size=rhythm_dec_lstm_size,
enc_use_meta=False,
dec_use_meta=True)
melody_net = MelodyNetwork(m=m,
V=V_melody,
rhythm_embed_size=out_size,
conv_f=conv_f,
conv_win_size=conv_win_size,
enc_lstm_size=melody_enc_lstm_size,
dec_lstm_1_size=melody_dec_lstm_1_size,
enc_use_meta=False,
dec_use_meta=True)
print("Individual networks set up...\n")
conv_f = 4
conv_win_size = 3
melody_enc_lstm_size = 52
melody_dec_lstm_size = 32
# INDIVIDUAL NETS
bar_embedder = BarEmbedding(V=V_rhythm,
beat_embed_size=beat_embed_size,
embed_lstm_size=embed_lstm_size,
out_size=out_size)
rhythm_encoder = RhythmEncoder(bar_embedder=bar_embedder,
context_size=rc_size,
lstm_size=rhythm_enc_lstm_size)
rhythm_net = RhythmNetwork(rhythm_encoder=rhythm_encoder,
dec_lstm_size=rhythm_dec_lstm_size,
V=V_rhythm,
dec_use_meta=True,
compile_now=True)
# ATTENTION: conv_win_size must not be greater than context size!
melody_encoder = MelodyEncoder(m=m,
conv_f=conv_f,
conv_win_size=min(mc_size, conv_win_size),
enc_lstm_size=melody_enc_lstm_size)
melody_net = MelodyNetwork(melody_encoder=melody_encoder,
rhythm_embed_size=out_size,
dec_lstm_size=melody_dec_lstm_size,
V=V_melody,
dec_use_meta=True,
compile_now=True)
#%%
with open("bach_beats.pkl", "rb") as handle:
data = pickle.load(handle)
data = rand.permutation(data)
parts = [list(map(tuple, part)) for score in data for part in score]
_, label_d = label([b for p in parts for b in p], start=1)
pad_symb = "<s>"
label_d[pad_symb] = 0
#%%
r = RhythmGenerator(4, label_d, "<s>")
r_gen = r.generate_data(parts, shuffle=False)
#%%
rnet = RhythmNetwork(num_categories=len(label_d),
embed_size=32,
lstm_size=64)
#%%
rnet.fit_generator(r_gen, steps_per_epoch=len(parts), epochs=1)
#%%
rnet.save_weights("rnet_weights.h5")
#%%
rnet2 = RhythmNetwork.from_weights("rnet_weights.h5", num_categories=len(label_d),
embed_size=32,
lstm_size=64)
#%%
rnet2.fit_generator(r_gen, steps_per_epoch=len(parts), epochs=1)