def _instantiate_object(
self,
template_object: SituationTemplateObject,
template: SimpleSituationTemplate,
chooser: SequenceChooser,
) -> SituationObject:
object_supertype = template.objects_to_ontology_types[template_object]
required_properties = template.objects_to_required_properties[
template_object]
compatible_ontology_types = self._ontology.nodes_with_properties(
root_node=object_supertype,
required_properties=required_properties)
if compatible_ontology_types:
ontology_node = chooser.choice(compatible_ontology_types)
return SituationObject.instantiate_ontology_node(
ontology_node,
properties=self._ontology.properties_for_node(ontology_node),
debug_handle=template_object.handle,
ontology=GAILA_PHASE_1_ONTOLOGY,
)
else:
raise RuntimeError(
f"When attempting to instantiate object {template_object} in"
f" {template}: no node at or under {object_supertype} with "
f"properties {required_properties} exists in the ontology "
f"{self._ontology}")
def test_recognizes_ontology_objects(object_type, language_mode):
situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=[
SituationObject.instantiate_ontology_node(
ontology_node=object_type, ontology=GAILA_PHASE_1_ONTOLOGY)
],
)
perception_generator = HighLevelSemanticsSituationToDevelopmentalPrimitivePerceptionGenerator(
GAILA_PHASE_1_ONTOLOGY)
perception = perception_generator.generate_perception(
situation, chooser=RandomChooser.for_seed(0), include_ground=False)
learner = IntegratedTemplateLearner(
object_learner=LANGUAGE_MODE_TO_TEMPLATE_LEARNER_OBJECT_RECOGNIZER[
language_mode])
descriptions = learner.describe(perception)
assert descriptions
if language_mode == LanguageMode.ENGLISH:
assert object_type.handle in one(
descriptions.items())[0].as_token_sequence()
else:
mappings = (
GAILA_PHASE_1_CHINESE_LEXICON._ontology_node_to_word # pylint:disable=protected-access
)
for k, v in mappings.items():
if k.handle == object_type.handle:
assert v.base_form in one(
descriptions.items())[0].as_token_sequence()
def _instantiate_object(
self,
object_var: TemplateObjectVariable,
variable_assignment: "TemplateVariableAssignment",
*,
has_addressee: bool,
default_addressee_node: OntologyNode,
) -> SituationObject:
object_type = variable_assignment.object_variables_to_fillers[
object_var]
asserted_properties = object_var.asserted_properties
if object_type == default_addressee_node and not has_addressee:
asserted_properties = immutableset(
object_var.asserted_properties.union({IS_ADDRESSEE}))
return SituationObject.instantiate_ontology_node(
ontology_node=object_type,
properties=[
# instantiate any property variables associated with this object
variable_assignment.
property_variables_to_fillers[asserted_property] if isinstance(
asserted_property,
TemplatePropertyVariable) else asserted_property
for asserted_property in asserted_properties
],
ontology=self.ontology,
)
def test_trivial_dynamic_situation_with_schemaless_object(language_mode):
dad_situation_object = SituationObject.instantiate_ontology_node(
ontology_node=DAD, ontology=GAILA_PHASE_1_ONTOLOGY)
situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=[dad_situation_object])
perception_generator = HighLevelSemanticsSituationToDevelopmentalPrimitivePerceptionGenerator(
GAILA_PHASE_1_ONTOLOGY)
# We explicitly exclude ground in perception generation
# this generates a static perception...
perception = perception_generator.generate_perception(
situation, chooser=RandomChooser.for_seed(0), include_ground=False)
# so we need to construct a dynamic one by hand from two identical scenes
dynamic_perception = PerceptualRepresentation(
frames=[perception.frames[0], perception.frames[0]])
perception_graph = PerceptionGraph.from_dynamic_perceptual_representation(
dynamic_perception)
perception_semantic_alignment = PerceptionSemanticAlignment.create_unaligned(
perception_graph)
(_, description_to_matched_semantic_node
) = LANGUAGE_MODE_TO_OBJECT_RECOGNIZER[language_mode].match_objects(
perception_semantic_alignment)
assert len(description_to_matched_semantic_node) == 1
assert (language_mode == LanguageMode.ENGLISH and
("Dad", ) in description_to_matched_semantic_node) or (
language_mode == LanguageMode.CHINESE and
("ba4 ba4", ) in description_to_matched_semantic_node)
def build_object_multiples_situations(
ontology: Ontology,
*,
samples_per_object: int = 3,
chooser: RandomChooser) -> Iterable[HighLevelSemanticsSituation]:
for object_type in PHASE_1_CURRICULUM_OBJECTS:
# Exclude slow objects for now
if object_type.handle in ["bird", "dog", "truck"]:
continue
is_liquid = ontology.has_all_properties(object_type, [LIQUID])
# don't want multiples of named people
if not is_recognized_particular(ontology,
object_type) and not is_liquid:
for _ in range(samples_per_object):
num_objects = chooser.choice(range(2, 4))
yield HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=[
SituationObject.instantiate_ontology_node(
ontology_node=object_type,
debug_handle=object_type.handle + f"_{idx}",
ontology=GAILA_PHASE_1_ONTOLOGY,
) for idx in range(num_objects)
],
axis_info=AxesInfo(),
)
def situation_object(
ontology_node: OntologyNode,
*,
debug_handle: Optional[str] = None,
properties: Iterable[OntologyNode] = tuple(),
) -> SituationObject:
return SituationObject.instantiate_ontology_node(
ontology_node=ontology_node,
ontology=GAILA_PHASE_1_ONTOLOGY,
debug_handle=debug_handle,
properties=properties,
)
def test_simple_experiment():
language_generator = SingleObjectLanguageGenerator(
GAILA_PHASE_1_ENGLISH_LEXICON)
perception_generator = DummyVisualPerceptionGenerator()
only_show_truck = GeneratedFromSituationsInstanceGroup(
name="only-ball",
situations=[
LocatedObjectSituation([(
SituationObject.instantiate_ontology_node(
BALL, ontology=GAILA_PHASE_1_ONTOLOGY),
Point(0.0, 0.0, 0.0),
)])
],
language_generator=language_generator,
perception_generator=perception_generator,
chooser=RandomChooser.for_seed(0),
)
pre_acc = CandidateAccuracyObserver("pre")
post_acc = CandidateAccuracyObserver("post")
test_acc = CandidateAccuracyObserver("test")
experiment = Experiment(
name="simple",
training_stages=[only_show_truck],
learner_factory=MemorizingLanguageLearner,
pre_example_training_observers=[
TopChoiceExactMatchObserver("pre"), pre_acc
],
post_example_training_observers=[
TopChoiceExactMatchObserver("post"), post_acc
],
warm_up_test_instance_groups=[only_show_truck],
test_instance_groups=[only_show_truck],
test_observers=[TopChoiceExactMatchObserver("test"), test_acc],
sequence_chooser=RandomChooser.for_seed(0),
)
execute_experiment(experiment)
assert pre_acc.accuracy() == 0.0
assert post_acc.accuracy() == 1.0
assert test_acc.accuracy() == 1.0
def test_with_object_recognizer(language_mode):
integrated_learner = IntegratedTemplateLearner(
object_learner=LANGUAGE_MODE_TO_TEMPLATE_LEARNER_OBJECT_RECOGNIZER[language_mode],
attribute_learner=None,
relation_learner=None,
action_learner=None,
)
dad_situation_object = SituationObject.instantiate_ontology_node(
ontology_node=DAD, ontology=GAILA_PHASE_1_ONTOLOGY
)
situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY, salient_objects=[dad_situation_object]
)
perception_generator = HighLevelSemanticsSituationToDevelopmentalPrimitivePerceptionGenerator(
GAILA_PHASE_1_ONTOLOGY
)
# We explicitly exclude ground in perception generation
# this generates a static perception...
perception = perception_generator.generate_perception(
situation, chooser=RandomChooser.for_seed(0), include_ground=False
)
# so we need to construct a dynamic one by hand from two identical scenes
dynamic_perception = PerceptualRepresentation(
frames=[perception.frames[0], perception.frames[0]]
)
descriptions = integrated_learner.describe(dynamic_perception)
assert len(descriptions) == 1
assert (
language_mode == LanguageMode.ENGLISH
and one(descriptions.keys()).as_token_sequence() == ("Dad",)
) or (
language_mode == LanguageMode.CHINESE
and one(descriptions.keys()).as_token_sequence() == ("ba4 ba4",)
)
def _instantiate_objects(
self,
template: Phase1SituationTemplate,
variable_assignment: "TemplateVariableAssignment",
*,
default_addressee_node: OntologyNode,
):
has_addressee = any(IS_ADDRESSEE in object_.asserted_properties
for object_ in template.all_object_variables)
object_var_to_instantiations_mutable: List[Tuple[
TemplateObjectVariable, SituationObject]] = [(
obj_var,
self._instantiate_object(
obj_var,
variable_assignment,
has_addressee=has_addressee,
default_addressee_node=default_addressee_node,
),
) for obj_var in template.all_object_variables]
if (default_addressee_node not in immutableset(
object_.ontology_node
for (_, object_) in object_var_to_instantiations_mutable)
and not has_addressee):
object_var_to_instantiations_mutable.append((
object_variable(default_addressee_node.handle,
default_addressee_node),
SituationObject.instantiate_ontology_node(
default_addressee_node,
properties=[IS_ADDRESSEE],
debug_handle=default_addressee_node.handle +
"_default_addressee",
ontology=self.ontology,
),
))
return immutabledict(object_var_to_instantiations_mutable)
def build_object_learner_experiment_curriculum_train(
num_samples: Optional[int],
num_noise_objects: Optional[int],
language_generator: LanguageGenerator[
HighLevelSemanticsSituation, LinearizedDependencyTree
],
*,
params: Parameters = Parameters.empty(),
) -> Sequence[Phase1InstanceGroup]:
situations = make_multiple_object_situation(
num_samples, num_noise_objects, language_generator
)
accurate_language_chance = params.floating_point(
"accurate_language_percentage", default=0.5
)
output_situations = []
random.seed(params.integer("random_seed", default=0))
rng = RandomChooser.for_seed(params.integer("language_random_seed", default=0))
for (situation, language, perception) in situations.instances():
if random.random() <= accurate_language_chance:
output_language = language
else:
# Make Invalid Language
if situation and isinstance(situation, HighLevelSemanticsSituation):
# First, gather all OntologyNodes which aren't already present in the situation
present_ontology_nodes = [
_object.ontology_node for _object in situation.all_objects
]
valid_other_objects = [
node
for node in PHASE_1_CURRICULUM_OBJECTS
if node not in present_ontology_nodes
]
# Then choose one at random
chosen_ontology_node = rng.choice(valid_other_objects)
# Make a fake situation with just this object in it, ignoring colors
wrong_situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_2_ONTOLOGY,
salient_objects=[
SituationObject.instantiate_ontology_node(
chosen_ontology_node, ontology=GAILA_PHASE_2_ONTOLOGY
)
],
syntax_hints=[IGNORE_COLORS],
)
# Generate the language as if it came from this fake situation rather than the original one
fake_language = only(
language_generator.generate_language(wrong_situation, chooser=rng)
)
output_language = LinearizedDependencyTree(
dependency_tree=fake_language.dependency_tree,
surface_token_order=fake_language.surface_token_order,
accurate=False,
)
else:
raise RuntimeError(
f"Unable to make invalid language without a situation of type HighlevelSemanticsSituation. Got situation: {situation}"
)
output_situations.append((situation, output_language, perception))
return [
AblatedLanguageSituationsInstanceGroup(
name=f"{situations.name()}_ablated", instances=output_situations
)
]
def run_experiment(learner, curricula, experiment_id):
# Teach each pretraining curriculum
for curriculum in curricula:
print("Teaching", curriculum.name(), "curriculum")
for (
_,
linguistic_description,
perceptual_representation,
) in curriculum.instances():
# Get the object matches first - prepositison learner can't learn without already recognized objects
# print('Observation: ',' '.join(linguistic_description.as_token_sequence()))
learner.observe(
LearningExample(perceptual_representation,
linguistic_description))
# Teach each kind member
empty_situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_2_ONTOLOGY,
salient_objects=immutableset([
SituationObject.instantiate_ontology_node(
ontology_node=GROUND,
debug_handle=GROUND.handle,
ontology=GAILA_PHASE_1_ONTOLOGY,
)
]),
)
empty_perception = GAILA_PHASE_2_PERCEPTION_GENERATOR.generate_perception(
empty_situation, PHASE1_CHOOSER_FACTORY())
pseudoword_to_kind = {"wug": "animal", "vonk": "food", "snarp": "people"}
print("Teaching new objects in known categories")
for word, kind in pseudoword_to_kind.items():
print("Observation: ", word, "s", "are", kind, "s")
learner.observe(
LearningExample(
empty_perception,
TokenSequenceLinguisticDescription(tokens=(word, "s", "are",
kind, "s"))
if kind != "people" else TokenSequenceLinguisticDescription(
tokens=(word, "s", "are", kind, "s")),
))
semantics_manager: SemanticsManager = SemanticsManager(
semantics_graph=learner.semantics_graph)
complete_results = []
print("\nResults for ", experiment_id)
for word, _ in pseudoword_to_kind.items():
results = [(kind,
semantics_manager.evaluate_kind_membership(word, kind))
for kind in pseudoword_to_kind.values()]
complete_results.append(results)
results_df = pd.DataFrame(
[[np.asscalar(i[1]) for i in l] for l in complete_results],
columns=["Animal", "Food", "People"],
)
results_df.insert(0, "Words", pseudoword_to_kind.keys())
# print(results_df.to_csv(index=False))
print(tabulate(results_df, headers="keys", tablefmt="psql"))
learner.log_hypotheses(Path(f"./renders/{experiment_id}"))
learner.render_semantics_to_file(
graph=learner.semantics_graph,
graph_name="semantics",
output_file=Path(f"./renders/{experiment_id}/semantics.png"),
)
def test_subset_learner_subobject():
mom = SituationObject.instantiate_ontology_node(
ontology_node=MOM, ontology=GAILA_PHASE_1_ONTOLOGY)
head = SituationObject.instantiate_ontology_node(
ontology_node=HEAD, ontology=GAILA_PHASE_1_ONTOLOGY)
hand = SituationObject.instantiate_ontology_node(
ontology_node=HAND, ontology=GAILA_PHASE_1_ONTOLOGY)
ball = SituationObject.instantiate_ontology_node(
ontology_node=BALL, ontology=GAILA_PHASE_1_ONTOLOGY)
house = SituationObject.instantiate_ontology_node(
ontology_node=HOUSE, ontology=GAILA_PHASE_1_ONTOLOGY)
ground = SituationObject.instantiate_ontology_node(
ontology_node=GROUND, ontology=GAILA_PHASE_1_ONTOLOGY)
mom_situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY, salient_objects=immutableset([mom]))
floating_head_situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=immutableset([head]),
other_objects=immutableset([ground]),
always_relations=flatten_relations(negate(on(head, ground))),
)
# Need to include some extra situations so that the learner will prune its semantics for 'a'
# away and not recognize it as an object.
floating_hand_situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=immutableset([hand]),
other_objects=immutableset([ground]),
always_relations=flatten_relations(negate(on(hand, ground))),
)
floating_ball_situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=immutableset([ball]),
other_objects=immutableset([ground]),
always_relations=flatten_relations(negate(on(ball, ground))),
)
floating_house_situation = HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=immutableset([house]),
other_objects=immutableset([ground]),
always_relations=flatten_relations(negate(on(house, ground))),
)
object_learner = SubsetObjectLearnerNew(ontology=GAILA_PHASE_1_ONTOLOGY,
beam_size=5,
language_mode=LanguageMode.ENGLISH)
for situation in [
mom_situation,
floating_head_situation,
floating_hand_situation,
floating_ball_situation,
floating_house_situation,
]:
perceptual_representation = GAILA_PHASE_1_PERCEPTION_GENERATOR.generate_perception(
situation, chooser=RandomChooser.for_seed(0))
for linguistic_description in GAILA_PHASE_1_LANGUAGE_GENERATOR.generate_language(
situation, chooser=RandomChooser.for_seed(0)):
perception_graph = PerceptionGraph.from_frame(
perceptual_representation.frames[0])
object_learner.learn_from(
LanguagePerceptionSemanticAlignment(
language_concept_alignment=LanguageConceptAlignment.
create_unaligned(language=linguistic_description),
perception_semantic_alignment=PerceptionSemanticAlignment(
perception_graph=perception_graph, semantic_nodes=[]),
))
mom_perceptual_representation = GAILA_PHASE_1_PERCEPTION_GENERATOR.generate_perception(
mom_situation, chooser=RandomChooser.for_seed(0))
perception_graph = PerceptionGraph.from_frame(
mom_perceptual_representation.frames[0])
enriched = object_learner.enrich_during_description(
PerceptionSemanticAlignment.create_unaligned(perception_graph))
semantic_node_types_and_debug_strings = {
(type(semantic_node), semantic_node.concept.debug_string)
for semantic_node in enriched.semantic_nodes
}
assert (ObjectSemanticNode, "Mom") in semantic_node_types_and_debug_strings
assert (ObjectSemanticNode,
"head") in semantic_node_types_and_debug_strings
assert (ObjectSemanticNode,
"hand") in semantic_node_types_and_debug_strings
def make_spin_tall_short_curriculum( # pylint: disable=unused-argument
num_samples: Optional[int],
noise_objects: Optional[int],
language_generator: LanguageGenerator[HighLevelSemanticsSituation,
LinearizedDependencyTree],
) -> Phase1InstanceGroup:
# "Mom spins a tall chair"
# We generate situations directly since templates fail to generate plurals.
learner = SituationObject.instantiate_ontology_node(
ontology_node=LEARNER,
debug_handle=LEARNER.handle,
ontology=GAILA_PHASE_1_ONTOLOGY,
)
situations = []
for agent_ontology_node in [MOM, DAD, BABY, DOG]:
agent = SituationObject.instantiate_ontology_node(
ontology_node=agent_ontology_node,
debug_handle=agent_ontology_node.handle,
ontology=GAILA_PHASE_1_ONTOLOGY,
)
for _object in [CHAIR, TABLE]:
theme = SituationObject.instantiate_ontology_node(
ontology_node=_object,
debug_handle=_object.handle,
ontology=GAILA_PHASE_1_ONTOLOGY,
)
other_objs = [
SituationObject.instantiate_ontology_node(
ontology_node=_object,
debug_handle=_object.handle + f"_{i}",
ontology=GAILA_PHASE_1_ONTOLOGY,
) for i in range(3)
]
computed_background = [learner]
computed_background.extend(other_objs)
# Tall and short
for relation_list in [
[bigger_than(learner, theme),
bigger_than(other_objs, theme)],
[bigger_than(theme, learner),
bigger_than(theme, other_objs)],
]:
situations.append(
HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=[agent, theme],
other_objects=computed_background,
actions=[
Action(
SPIN,
argument_roles_to_fillers=[
(AGENT, agent),
(THEME, theme),
],
)
],
always_relations=relation_list,
syntax_hints=[USE_VERTICAL_MODIFIERS],
))
return phase1_instances("Tall - Short Curriculum",
situations,
language_generator=language_generator)
def make_eat_big_small_curriculum( # pylint: disable=unused-argument
num_samples: Optional[int],
noise_objects: Optional[int],
language_generator: LanguageGenerator[HighLevelSemanticsSituation,
LinearizedDependencyTree],
) -> Phase1InstanceGroup:
# "Mom eats a big cookie"
# We generate situations directly since templates fail to generate plurals.
learner = SituationObject.instantiate_ontology_node(
ontology_node=LEARNER,
debug_handle=LEARNER.handle,
ontology=GAILA_PHASE_1_ONTOLOGY,
)
situations = []
for eater_ontology_node in [MOM, DAD, BABY, DOG]:
eater = SituationObject.instantiate_ontology_node(
ontology_node=eater_ontology_node,
debug_handle=eater_ontology_node.handle,
ontology=GAILA_PHASE_1_ONTOLOGY,
)
for _object in [COOKIE, WATERMELON]:
object_to_eat = SituationObject.instantiate_ontology_node(
ontology_node=_object,
debug_handle=_object.handle,
ontology=GAILA_PHASE_1_ONTOLOGY,
)
other_edibles = [
SituationObject.instantiate_ontology_node(
ontology_node=_object,
debug_handle=_object.handle + f"_{i}",
ontology=GAILA_PHASE_1_ONTOLOGY,
) for i in range(3)
]
computed_background = [learner]
computed_background.extend(other_edibles)
# Big
for relation_list in [
[
bigger_than(object_to_eat, learner),
bigger_than(object_to_eat, other_edibles),
],
[
bigger_than(learner, object_to_eat),
bigger_than(other_edibles, object_to_eat),
],
]:
situations.append(
HighLevelSemanticsSituation(
ontology=GAILA_PHASE_1_ONTOLOGY,
salient_objects=[eater, object_to_eat],
other_objects=computed_background,
actions=[
Action(
EAT,
argument_roles_to_fillers=[
(AGENT, eater),
(PATIENT, object_to_eat),
],
)
],
always_relations=relation_list,
))
return phase1_instances("Big - Small Curriculum",
situations,
language_generator=language_generator)
