def test_multitask_custom_attachments(self):
"""Attach the task heads at user-specified layers"""
edges = [(0, 1)]
cards = [2, 2]
tg = TaskHierarchy(cards, edges)
em = MTEndModel(
layer_out_dims=[2, 8, 4],
task_graph=tg,
seed=1,
verbose=False,
task_head_layers=[1, 2],
)
self.assertEqual(em.task_map[1][0], 0)
self.assertEqual(em.task_map[2][0], 1)
em.train_model(
(self.Xs[0], self.Ys[0]),
valid_data=(self.Xs[1], self.Ys[1]),
verbose=False,
n_epochs=10,
checkpoint=False,
)
score = em.score((self.Xs[2], self.Ys[2]),
reduce="mean",
verbose=False)
self.assertGreater(score, 0.95)
def __init__(self,
task_graph=None,
input_module=None,
seed=None,
**kwargs):
defaults = recursive_merge_dicts(em_default_config,
mt_em_default_config,
misses='insert')
self.config = recursive_merge_dicts(defaults, kwargs)
# If no task_graph is specified, default to a single binary task
if task_graph is None:
task_graph = TaskHierarchy(edges=[], cardinalities=[2])
self.task_graph = task_graph
self.K_t = self.task_graph.K_t # Cardinalities by task
self.T = self.task_graph.T # Total number of tasks
MTClassifier.__init__(self, cardinalities=self.K_t, seed=seed)
if input_module is None:
input_module = IdentityModule(self.config['layer_output_dims'][0])
self._build(input_module)
# Show network
if self.config['verbose']:
print("\nNetwork architecture:")
self._print()
print()
def LabelLearner(cardinalities=2, dependencies=[], **kwargs):
"""
A generative factory function for data programming (MeTal).
"""
# Use single task label model if cardinality is an integer,
# Otherwise use multi task label model
if isinstance(cardinalities, int):
logger.info("Using MeTaL single task label model")
label_model = LabelModel(k=cardinalities, **kwargs)
else:
logger.info("Using MeTaL multi task label model")
task_graph = TaskHierarchy(cardinalities=cardinalities,
edges=dependencies)
label_model = MTLabelModel(task_graph=task_graph, **kwargs)
return label_model
def test_scoring(self):
edges = [(0, 1)]
cards = [2, 2]
tg = TaskHierarchy(cards, edges)
em = MTEndModel(layer_out_dims=[2, 8, 4],
task_graph=tg,
seed=1,
verbose=False)
em.train_model(
(self.Xs[0], self.Ys[0]),
valid_data=(self.Xs[1], self.Ys[1]),
verbose=False,
n_epochs=3,
checkpoint=False,
validation_task=0,
)
tasks = [0, 1]
for metric in METRICS:
all_scores = em.score((self.Xs[2], self.Ys[2]),
metric=metric,
reduce=None,
verbose=False)
task_specific_scores_score_method = [
em.score(
(self.Xs[2], self.Ys[2]),
metric=metric,
validation_task=task,
verbose=False,
) for task in tasks
]
task_specific_scores_score_task_method = [
em.score_task(self.Xs[2],
self.Ys[2],
t=task,
metric=metric,
verbose=False) for task in tasks
]
for i in range(len(tasks)):
self.assertEqual(
all_scores[i],
task_specific_scores_score_method[i],
task_specific_scores_score_task_method[i],
)
def test_multitask_custom(self):
edges = [(0, 1)]
cards = [2, 2]
tg = TaskHierarchy(edges, cards)
em = MTEndModel(
task_graph=tg,
seed=1,
verbose=False,
dropout=0.0,
layer_output_dims=[2, 8, 4],
task_head_layers=[1, 2],
)
em.train(
self.Xs[0],
self.Ys[0],
self.Xs[1],
self.Ys[1],
verbose=False,
n_epochs=10,
)
score = em.score(self.Xs[2], self.Ys[2], reduce='mean', verbose=False)
self.assertGreater(score, 0.95)