def test_load_grammar(self):
grammar = grammar_utils.load_grammar(
grammar_path='third_party/google_research/google_research/'
'neural_guided_symbolic_regression/grammar/'
'univariate_one_constant_grammar.txt')
grammar_production_rule_strings = [
str(rule) for rule in grammar.prod_rules
]
self.assertIsInstance(grammar, arithmetic_grammar.Grammar)
self.assertListEqual(
grammar_production_rule_strings,
[
# pylint: disable=g-inconsistent-quotes
"Nothing -> None", # Padding at 0-th index.
"O -> S", # Unique starting production rule at 1-st index.
"S -> S '+' T",
"S -> S '-' T",
"S -> S '*' T",
"S -> S '/' T",
"S -> T",
"T -> '(' S ')'",
"T -> 'x'",
"T -> '1'",
# pylint: enable=g-inconsistent-quotes
])
def run():
"""Runs train_and_evaluate."""
hparams_filename = os.path.join(FLAGS.model_dir, 'hparams.json')
if FLAGS.is_chief:
gfile.MakeDirs(FLAGS.model_dir)
hparams = core.read_hparams(FLAGS.hparams, get_hparams())
core.write_hparams(hparams, hparams_filename)
# Always load HParams from model_dir.
hparams = core.wait_for_hparams(hparams_filename, get_hparams())
grammar = grammar_utils.load_grammar(grammar_path=hparams.grammar_path)
estimator = tf.estimator.Estimator(
model_fn=functools.partial(model_fn, grammar=grammar),
params=hparams,
config=tf.estimator.RunConfig(
save_checkpoints_secs=hparams.save_checkpoints_secs,
keep_checkpoint_max=hparams.keep_checkpoint_max))
train_spec = tf.estimator.TrainSpec(input_fn=functools.partial(
input_ops.input_fn,
input_pattern=hparams.train_pattern,
grammar=grammar),
max_steps=hparams.train_steps)
# NOTE(leeley): The SavedModel will be stored under the
# tf.saved_model.tag_constants.SERVING tag.
latest_exporter = tf.estimator.LatestExporter(
name='latest_exported_model',
serving_input_receiver_fn=functools.partial(
input_ops.serving_input_receiver_fn,
params=hparams,
num_production_rules=grammar.num_production_rules),
exports_to_keep=hparams.exports_to_keep)
eval_hooks = []
if hparams.num_expressions_per_condition > 0:
eval_hooks.append(
metrics.GenerationWithLeadingPowersHook(
generation_leading_powers_abs_sums=core.hparams_list_value(
hparams.generation_leading_powers_abs_sums),
num_expressions_per_condition=hparams.
num_expressions_per_condition,
max_length=hparams.max_length,
grammar=grammar))
eval_spec = tf.estimator.EvalSpec(
input_fn=functools.partial(input_ops.input_fn,
input_pattern=hparams.tune_pattern,
grammar=grammar),
steps=hparams.eval_steps,
exporters=latest_exporter,
start_delay_secs=hparams.start_delay_secs,
throttle_secs=hparams.throttle_secs,
hooks=eval_hooks)
tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
def test_get_number_valid_next_step(self, prod_rules_sequence_indices,
expected):
grammar = grammar_utils.load_grammar(
grammar_path='third_party/google_research/google_research/'
'neural_guided_symbolic_regression/grammar/'
'univariate_one_constant_grammar.txt')
number_valid_next_step = (
generate_empirical_distribution_df.get_number_valid_next_step(
prod_rules_sequence_indices, grammar))
self.assertEqual(number_valid_next_step, expected)
def setUp(self):
super(GenerateEmpiricalDistributionDfMainTest, self).setUp()
# Production rule sequence of ( 1 ) is 1,6,7,6,9.
# Production rule sequence of ( x ) is 1,6,7,6,8.
self.expression_df = pd.DataFrame(
{'expression_string': ['( 1 )', '( x )'],
'leading_at_0': [0, 1],
'leading_at_inf': [0, 1]})
self.grammar = grammar_utils.load_grammar(
grammar_path='third_party/google_research/google_research/'
'neural_guided_symbolic_regression/grammar/'
'univariate_one_constant_grammar.txt')
self.max_length = 11
def setUp(self):
super(GenerateExpressionTest, self).setUp()
# The grammar contains:
# 0: Nothing -> None
# 1: O -> S
# 2: S -> S '+' T
# 3: S -> S '-' T
# 4: S -> S '*' T
# 5: S -> S '/' T
# 6: S -> T
# 7: T -> '(' S ')'
# 8: T -> 'x'
# 9: T -> '1'
self.grammar = grammar_utils.load_grammar(
grammar_path='third_party/google_research/google_research/'
'neural_guided_symbolic_regression/grammar/'
'univariate_one_constant_grammar.txt')
self.empirical_distribution_df = pd.DataFrame(
np.array([[-4, -4], [-3, -3], [-3, 0], [-3, 1]]),
columns=['leading_at_0', 'leading_at_inf'])
self.empirical_distribution_df['partial_sequence_indices'] = '1'
self.empirical_distribution_df.set_index(
['partial_sequence_indices', 'leading_at_0', 'leading_at_inf'],
inplace=True)
self.empirical_distribution_df = pd.DataFrame(
[[0, 0, 0, 0, 0, 1, 0, 0, 0, 0],
[0, 0, 0, 0, 0.05, 0.9, 0.05, 0, 0, 0],
[0, 0, 0.25, 0.25, 0, 0.5, 0, 0, 0, 0],
[0, 0, 0.25, 0.25, 0, 0.5, 0, 0, 0, 0]],
index=self.empirical_distribution_df.index)
self.limited_history_empirical_distribution_df = (
self.empirical_distribution_df.copy())
self.limited_history_empirical_distribution_df.index.names = (
['tail_partial_sequence_indices'] +
self.limited_history_empirical_distribution_df.index.names[1:])
self.empirical_distribution_df_without_condition = (
self.empirical_distribution_df.iloc[[0], :].reset_index(
level=['leading_at_0', 'leading_at_inf'], drop=True))
self.limited_history_empirical_distribution_df_without_condition = (
self.limited_history_empirical_distribution_df.iloc[
[0], :].reset_index(level=['leading_at_0', 'leading_at_inf'],
drop=True))
self.next_production_rule_mask = [0, 0, 1, 1, 1, 1, 1, 0, 0, 0]
