def create_datasets(config, train_split, test_split, data_format='squad'):
"""Setup datasets.
Args:
config: a config describing the dataset
train_split: name of train split
test_split: name of test split
data_format: 'squad' or 'squad2'
Returns:
A ConfigDict containing 4 keys:
train_fn, test_fn - Input functions
train_dataset, test_dataset - instantiated dataset objects
"""
# Load dataset
train_config = configurable.merge(config, split_name=train_split)
eval_config = configurable.merge(config, split_name=test_split)
dataset_class = configurable.Configurable.load(config)
train = dataset_class.get_input_fn(
mode='train', config=train_config, data_format=data_format)
test = dataset_class.get_input_fn(
mode='eval', config=eval_config, data_format=data_format)
train_dataset = dataset_class(
mode='train', config=train_config, data_format=data_format)
test_dataset = dataset_class(
mode='eval', config=eval_config, data_format=data_format)
return dict(
train_fn=train,
test_fn=test,
train_dataset=train_dataset,
test_dataset=test_dataset)
def test_merge(self):
a = {'a': 1, 'b': {'c': 2}}
b = {'a': 3, 'b': {'c': 4, 'd': 5}}
a_b_result = {'a': 3, 'b': {'c': 4, 'd': 5}}
b_a_result = {'a': 1, 'b': {'c': 2, 'd': 5}}
a_b_merge = configurable.merge(a, b)
b_a_merge = configurable.merge(b, a)
self.assertAllEqual(a_b_result, a_b_merge)
self.assertAllEqual(b_a_result, b_a_merge)
def main(_):
"""Parse config flag and config files then run experiment."""
# Load config
parsed_config = configurable.load_config(FLAGS.config_file, FLAGS.config)
if 'fn' not in parsed_config:
# Fill in the default template
parsed_config = configurable.merge(
parsed_config, fn='ConfigurableExperiment')
configurable.save_config(FLAGS.model_dir, parsed_config)
# TODO(thangluong): consider use FLAGS if we do need to use this often
parsed_config['log_device_placement'] = False
parsed_config['allow_soft_placement'] = True
if 'data_format' not in parsed_config:
parsed_config['data_format'] = 'squad'
assert FLAGS.model_dir
experiment_fn = experiment.create_experiment_fn(
default_config=parsed_config, return_distributed_spec=False)
exp = experiment_fn(hparams=None, model_dir=FLAGS.model_dir)
if FLAGS.mode == 'train':
exp.train()
elif FLAGS.mode == 'eval':
exp.evaluate()
elif FLAGS.mode == 'train_and_eval':
exp.train_and_evaluate()
else:
raise ValueError('Unknown mode %s' % FLAGS.mode)
def _create_config(hparams):
"""Create trial config and save to disk.
Args:
hparams: Nondefault params to merge
Returns:
A configurable object `spec` instantiated from the final config.
spec.config will return the config.
"""
hparams = hparams or contrib_training.HParams()
tuned_config = configurable.unflatten_dict(hparams.values())
pprinter = pprint.PrettyPrinter()
tf.logging.info('Provided extra params:\n%s',
pprinter.pformat(tuned_config))
try:
merged_config = configurable.merge(default_config, tuned_config)
tf.logging.info('Tuned default config:\n%s', merged_config)
except TypeError as e:
tf.logging.info(
'Do not provide same config in both config string and vizier.'
' This may lead to type errors.')
raise e
# Instantiate a ConfigurableExperiment object.
experiment_spec = configurable.Configurable.initialize(merged_config)
tf.logging.info('Final config:\n%s', experiment_spec.config)
return experiment_spec
