def create_config(model_dir, hparams=None):
"""Creates config instance."""
tf.logging.info("model_dir = " + model_dir)
assert model_dir
if hparams:
tf.logging.info("Given override cfg:\n%s" % pprint.pformat(hparams))
else:
hparams = dict()
# Build the default config
cfg = model.build_config(model_dir=model_dir, data_path=FLAGS.data_path)
if FLAGS.config_file:
with tf.gfile.GFile(FLAGS.config_file) as f:
file_cfg = json.load(f)
file_cfg = utils.to_config(file_cfg)
tf.logging.info("Loaded config from file:\n%s" % file_cfg)
cfg = utils.merge_fixed_structure(cfg, file_cfg)
# Override from flags
overrides = dict()
if FLAGS.config:
overrides = utils.parse_config_string(FLAGS.config)
tf.logging.info("Parsed config overrides:\n%s" % overrides)
cfg = utils.merge_fixed_structure(cfg, overrides)
if FLAGS.master:
cfg.master = FLAGS.master
cfg = utils.merge_fixed_structure(cfg, utils.unflatten_dict(hparams))
tf.logging.info("Operative config:\n%s" % cfg)
return cfg
def evaluate(override_cfg, model_dir, continuous=True):
"""Run training and evaluation."""
tf.logging.info("model_dir = " + model_dir)
try:
cfg = _load_config(model_dir)
except tf.errors.NotFoundError:
tf.logging.info("Model directory does not exist yet. Creating new config.")
cfg = model.build_config(model_dir=model_dir, data_path=FLAGS.data_path)
tf.logging.info(cfg)
tf.logging.info(override_cfg)
cfg = utils.merge(cfg, override_cfg)
cfg.tpu.enable = False
cfg.dataset.max_length = None
# Construct inputs and estimator
_, eval_input = data.build_dataset(cfg.dataset, is_tpu=cfg.tpu.enable)
estimator = model.get_estimator(**cfg)
if continuous:
checkpoints_iterator = tf.contrib.training.checkpoints_iterator(
cfg.model_dir)
eval_metrics = None
for ckpt_path in checkpoints_iterator:
eval_metrics = estimator.evaluate(
input_fn=eval_input, checkpoint_path=ckpt_path)
tf.logging.info(pprint.pformat(eval_metrics))
return eval_metrics
else:
eval_metrics = estimator.evaluate(input_fn=eval_input)
return eval_metrics