cfg.logger.debug('Schedular patience: {}'.format(cfg.sch_patience))
cfg.logger.debug('Schedular verbose: {}'.format(cfg.sch_verbose))
cfg.logger.debug('Device: {}'.format(cfg.device))
cfg.logger.debug('Embedding model directory: {}'.format(cfg.emb_model_dir))
cfg.logger.debug('Lyrics data directory: {}'.format(cfg.lyrics_dir))
if cfg.pretrained_lm_dir:
cfg.logger.debug('Pre-trained language model: {}'.format(
cfg.pretrained_lm_dir))
else:
cfg.logger.debug('Pre-trained language model: initial training')
# Training
language_model = LanguageModel(wv_dict, cfg.hidden_dim).to(cfg.device)
criterion = nn.NLLLoss()
optimizer = optim.Adam(language_model.parameters(), lr=cfg.lr)
schedular = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='min',
factor=cfg.sch_factor,
patience=cfg.sch_patience,
verbose=cfg.sch_verbose)
if cfg.pretrained_lm_dir:
lm_loading_res = language_model.load_state_dict(
torch.load(cfg.pretrained_lm_dir))
cfg.logger.debug('Loading language model: {}'.format(lm_loading_res))
train_losses, train_accs = [], [] # losses & accuracies to save
if cfg.test_ratio > 0:
test_losses, test_accs = [], []
cfg.logger.info('Training.')
