logger.info("%s -> %.6f" % (k, v))
if params.is_master:
logger.info("__log__:%s" % json.dumps(scores))
# end of epoch
if params.validation_metrics != '':
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch(scores)
if __name__ == '__main__':
# generate parser / parse parameters
parser = get_parser()
params = parser.parse_args()
# debug mode
if params.debug:
params.exp_name = 'debug'
params.exp_id = 'debug_%08i' % random.randint(0, 100000000)
params.debug_slurm = True
params.debug_train = True
# check parameters
check_data_params(params)
check_model_params(params)
# run experiment
main(params)
def main(params):
# check_data_params(params)
check_model_params(params)
# initialize the experiment
logger = initialize_exp(params)
# load data
data = load_data(params)
# check_vocab(data)
# build model
if params.encoder_only:
model = build_model(params, data['source_dico'])
else:
encoder, decoder = build_model(params, data['source_dico'], data['target_dico'])
# build trainer, reload potential checkpoints / build evaluator
if params.encoder_only:
trainer = SingleTrainer(model, data, params)
evaluator = SingleEvaluator(trainer, data, params)
else:
trainer = EncDecTrainer(encoder, decoder, data, params)
evaluator = EncDecEvaluator(trainer, data, params)
# evaluation
if params.eval_only:
scores = evaluator.run_all_evals(trainer)
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
exit()
# language model training
for _ in range(params.max_epoch):
logger.info("============ Starting epoch %i ... ============" % trainer.epoch)
trainer.n_iter = 0
while trainer.n_iter < trainer.epoch_size:
if params.cs_step:
trainer.content_selection_step(params.lambda_cs)
if params.sm_step:
trainer.summarization_step(params.lambda_sm)
if params.lm_step:
trainer.clm_step(params.lambda_lm)
trainer.iter()
logger.info("============ End of epoch %i ============" % trainer.epoch)
# evaluate perplexity
scores = evaluator.run_all_evals(trainer)
# print / JSON log
for k, v in scores.items():
logger.info("%s -> %.6f" % (k, v))
logger.info("__log__:%s" % json.dumps(scores))
# end of epoch
trainer.save_best_model(scores)
trainer.save_periodic()
trainer.end_epoch()
