def main(unused_argv):
if FLAGS.checkpoint is None or not FLAGS.checkpoint:
raise ValueError(
'Need to provide a path to checkpoint directory.')
wmodel = lib_graph.load_checkpoint(FLAGS.checkpoint)
if FLAGS.eval_logdir is None:
raise ValueError(
'Set flag eval_logdir to specify a path for saving eval statistics.')
else:
eval_logdir = os.path.join(FLAGS.eval_logdir, 'eval_stats')
tf.gfile.MakeDirs(eval_logdir)
evaluator = lib_evaluation.BaseEvaluator.make(
FLAGS.unit, wmodel=wmodel, chronological=FLAGS.chronological)
evaluator = lib_evaluation.EnsemblingEvaluator(evaluator, FLAGS.ensemble_size)
if not FLAGS.sample_npy_path and FLAGS.fold is None:
raise ValueError(
'Either --fold must be specified, or paths of npy files to load must '
'be given, but not both.')
if FLAGS.fold is not None:
evaluate_fold(FLAGS.fold, evaluator, wmodel.hparams, eval_logdir)
if FLAGS.sample_npy_path is not None:
evaluate_paths([FLAGS.sample_npy_path], evaluator, wmodel.hparams,
eval_logdir)
print('Done')
def main(ckpt, evaldir, unit, chronological, ensemble_size, sample_path,
folder, index, data_dir):
checkpoint_dir = ckpt
if not checkpoint_dir:
# If a checkpoint directory is not specified, see if there is only one
# subdir in this folder and use that.
possible_checkpoint_dirs = tf.gfile.ListDirectory(evaldir)
possible_checkpoint_dirs = [
i for i in possible_checkpoint_dirs
if tf.gfile.IsDirectory(os.path.join(evaldir, i))
]
if EVAL_SUBDIR in possible_checkpoint_dirs:
possible_checkpoint_dirs.remove(EVAL_SUBDIR)
if len(possible_checkpoint_dirs) == 1:
checkpoint_dir = os.path.join(evaldir, possible_checkpoint_dirs[0])
tf.logging.info('Using checkpoint dir: %s', checkpoint_dir)
else:
raise ValueError(
'Need to provide a path to checkpoint directory or use an '
'eval_logdir with only 1 checkpoint subdirectory.')
wmodel = lib_graph.load_checkpoint(checkpoint_dir)
if evaldir is None:
raise ValueError(
'Set flag eval_logdir to specify a path for saving eval statistics.'
)
else:
eval_logdir = os.path.join(evaldir, EVAL_SUBDIR)
tf.gfile.MakeDirs(eval_logdir)
evaluator = lib_evaluation.BaseEvaluator.make(unit,
wmodel=wmodel,
chronological=chronological)
evaluator = lib_evaluation.EnsemblingEvaluator(evaluator, ensemble_size)
if not sample_path and folder is None:
raise ValueError(
'Either --fold must be specified, or paths of npy files to load must '
'be given, but not both.')
if folder is not None:
evaluate_fold(folder, evaluator, wmodel.hparams, eval_logdir,
checkpoint_dir, index, unit, ensemble_size,
chronological, data_dir)
if sample_path is not None:
evaluate_paths([sample_path], evaluator, wmodel.hparams, eval_logdir,
unit, ensemble_size, chronological)
tf.logging.info('Done')