def setUpClass(cls):
try:
cls.config = Config('config.toml')
engine = create_engine(cls.config.db_uri_test)
SessionMaker = sessionmaker(bind=engine)
cls.session = SessionMaker()
create_tables(engine)
cls.session.query(Close).delete()
cls.session.query(Price).delete()
cls.session.query(PriceSeq).delete()
cls.session.commit()
dir_resources = Path(cls.config.dir_resources)
dir_prices = dir_resources / Path('pseudo-data') / Path('prices')
missing_rics = ['.TEST']
logger = create_logger(Path('test.log'),
is_debug=False,
is_temporary=True)
# insert database
insert_prices(cls.session, dir_prices, missing_rics, dir_resources,
logger)
except: # noqa: E722
raise unittest.SkipTest('Cannot establish connection')
def main() -> None:
args = parse_args()
if not args.is_debug:
warnings.simplefilter(action='ignore', category=FutureWarning)
config = Config(args.dest_config)
device = torch.device(args.device)
now = datetime.today().strftime('reporter-%Y-%m-%d-%H-%M-%S')
dest_dir = config.dir_output / Path(now) \
if args.output_subdir is None \
else config.dir_output / Path(args.output_subdir)
dest_log = dest_dir / Path('reporter.log')
logger = create_logger(dest_log, is_debug=args.is_debug)
config.write_log(logger)
message = 'start main (is_debug: {}, device: {})'.format(args.is_debug, args.device)
logger.info(message)
# === Alignment ===
has_all_alignments = \
reduce(lambda x, y: x and y,
[(config.dir_output / Path('alignment-{}.json'.format(phase.value))).exists()
for phase in list(Phase)])
if not has_all_alignments:
engine = create_engine(config.db_uri)
SessionMaker = sessionmaker(bind=engine)
pg_session = SessionMaker()
create_tables(engine)
prepare_resources(config, pg_session, logger)
for phase in list(Phase):
config.dir_output.mkdir(parents=True, exist_ok=True)
dest_alignments = config.dir_output / Path('alignment-{}.json'.format(phase.value))
alignments = load_alignments_from_db(pg_session, phase, logger)
with dest_alignments.open(mode='w') as f:
writer = jsonlines.Writer(f)
writer.write_all(alignments)
pg_session.close()
# === Dataset ===
(vocab, train, valid, test) = create_dataset(config, device)
vocab_size = len(vocab)
dest_vocab = dest_dir / Path('reporter.vocab')
with dest_vocab.open(mode='wb') as f:
torch.save(vocab, f)
seqtypes = []
attn = setup_attention(config, seqtypes)
encoder = Encoder(config, device)
decoder = Decoder(config, vocab_size, attn, device)
model = EncoderDecoder(encoder, decoder, device)
optimizer = torch.optim.Adam(model.parameters(), lr=config.learning_rate)
criterion = torch.nn.NLLLoss(reduction='elementwise_mean',
ignore_index=vocab.stoi[SpecialToken.Padding.value])
# === Train ===
dest_model = dest_dir / Path('reporter.model')
prev_valid_bleu = 0.0
max_bleu = 0.0
best_epoch = 0
early_stop_counter = 0
for epoch in range(config.n_epochs):
logger.info('start epoch {}'.format(epoch))
train_result = run(train,
vocab,
model,
optimizer,
criterion,
Phase.Train,
logger)
train_bleu = calc_bleu(train_result.gold_sents, train_result.pred_sents)
valid_result = run(valid,
vocab,
model,
optimizer,
criterion,
Phase.Valid,
logger)
valid_bleu = calc_bleu(valid_result.gold_sents, valid_result.pred_sents)
s = ' | '.join(['epoch: {0:4d}'.format(epoch),
'training loss: {:.2f}'.format(train_result.loss),
'training BLEU: {:.4f}'.format(train_bleu),
'validation loss: {:.2f}'.format(valid_result.loss),
'validation BLEU: {:.4f}'.format(valid_bleu)])
logger.info(s)
if max_bleu < valid_bleu:
torch.save(model.state_dict(), str(dest_model))
max_bleu = valid_bleu
best_epoch = epoch
early_stop_counter = early_stop_counter + 1 \
if prev_valid_bleu > valid_bleu \
else 0
if early_stop_counter == config.patience:
logger.info('EARLY STOPPING')
break
prev_valid_bleu = valid_bleu
# === Test ===
with dest_model.open(mode='rb') as f:
model.load_state_dict(torch.load(f))
test_result = run(test,
vocab,
model,
optimizer,
criterion,
Phase.Test,
logger)
test_bleu = calc_bleu(test_result.gold_sents, test_result.pred_sents)
s = ' | '.join(['epoch: {:04d}'.format(best_epoch),
'Test Loss: {:.2f}'.format(test_result.loss),
'Test BLEU: {:.10f}'.format(test_bleu)])
logger.info(s)
export_results_to_csv(dest_dir, test_result)