# embedding list
idx_to_emb = [w2v_model[word] for word in w2v_model.wv.index2word]
# get index of period
pd_idx = w2v_model.wv.vocab['.'].index
# get indices to remove
words_to_remove = ['<unk>']
remove_idxs = [
w2v_model.wv.vocab[word].index for word in words_to_remove
if word in w2v_model.wv.vocab
]
# initialize Logger
log_dir = config['generate']['log_dir']
fn_suffix = config['generate']['log_fn_suffix']
logger = ut.Logger(log_dir, fn_suffix)
# instantiate LSTM model
lstm_rnn = rnn_keras.LyricsLSTM(config, logger, generate_mode=True)
# perform prediction
if not input_words:
for i in range(num_stanzas):
#cur_word_idx = pd_idx
cur_word_idx = np.random.randint(low=0,
high=len(w2v_model.wv.index2word))
lyrics_str = w2v_model.wv.index2word[cur_word_idx]
lstm_rnn.model.reset_states()
# generate lyrics until period is encountered
# as long as length is within min_words:max_words
weight_decay=args.weight_decay)
scheduler = ReduceLROnPlateau(optimizer,
factor=args.gamma,
patience=args.patience,
min_lr=args.min_lr)
# scheduler = StepLR(optimizer, 100, gamma=0.1)
elif args.optimizer == "adam":
optimizer = optim.Adam(model.parameters(),
lr=args.learning_rate,
weight_decay=args.weight_decay)
try:
os.mkdir("logs")
except:
pass
logger = utils.Logger("logs/{}.log".format(args.name))
with open("logs/{}.log".format(args.name), "a") as f:
f.write(str(args))
f.write("\nParameters : " + str(n_parameters))
if hasattr(model, "n_filters"):
f.write("\nFilters : " + str(model.n_filters))
else:
f.write("\nFilters : _ ")
f.write("\n*******\n")
print("-" * 80 + "\n")
test_loss = 0
test_acc = torch.zeros(len(topk))
lr = optimizer.state_dict()["param_groups"][0]["lr"]
for epoch in range(1, args.epochs + 1):
def init_log(self, folder=''):
self.logfile = utils.ensure_path(os.path.join(self.folder, 'log.txt'))
sys.stdout = utils.Logger(self.logfile)