Skip to content
Snippets Groups Projects
Commit c8e46f2b authored by dmt's avatar dmt
Browse files

Fix bugs in time/sigma and sigma/zeta deconstructions.

parent ac5c8316
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
cml/domain/deconstruction.py
+ 58
− 56
View file @ c8e46f2b
...@@ -329,38 +329,40 @@ class Deconstructor: ...@@ -329,38 +329,40 @@ class Deconstructor:
success = False success = False
if r_model and p_model.tier < self.settings.highest_tier-1: if r_model and p_model.tier < self.settings.highest_tier-1:
# Get learnblock that trained relative model
second_block = r_model.trained_with(self.source) second_block = r_model.trained_with(self.source)
overlapping = second_block.new_block_from(block.get_column_values("T"))
if overlapping.rows >= self.settings.learn_block_minimum:
alpha = self.calculate_reliability(p_model.pre_image_labels,
r_model.pre_image_labels)
alpha_systematic = alpha < 0
alpha_weak_reliability = 0 <= alpha < self.settings.min_reliability
if (self.settings.allow_weak_reliability and
alpha_weak_reliability) or alpha_systematic:
overlapping_b = block.new_block_from(overlapping.get_column_values("T"))
overblock = self.source.new_learnblock(
values=list(zip(
overlapping.get_column_values("Z"),
overlapping_b.get_column_values("Z"),
overlapping.get_column_values("T"),
("\"\"" for _ in range(overlapping.rows)),
("\"\"" for _ in range(overlapping.rows)))),
columns=(p_model.uid, r_model.uid, "T", "Sigma", "Z"),
index=[i for i in range(overlapping.rows)],
origin=[p_model.uid, r_model.uid]
)
# Get samples that have overlapping timestamp
over_block = block.overlapping_rows(second_block, subset=["T"])
# Check rows constraint
if over_block.rows >= self.settings.learn_block_minimum:
# Calculate reliability (which block???)
alpha = self.calc_reliability(r_model, p_model, block)
#alpha_systematic = alpha < 0
#alpha_weak_reliability = 0 <= alpha < self.settings.min_reliability
#if (self.settings.allow_weak_reliability and
# alpha_weak_reliability) or alpha_systematic:
if self.settings.allow_weak_reliability and \
alpha > self.settings.min_reliability:
# Create learnblock from the aim values of the overlapping
# samples # samples
data = list(zip(over_block.get_column_values("Z"), # data = list(zip(over_block.get_column_values("Z"),
over_block.get_column_values("T"), # over_block.get_column_values("T"),
["\"\"" for _ in range(over_block.rows)], # ["\"\"" for _ in range(over_block.rows)],
["\"\"" for _ in range(over_block.rows)])) # ["\"\"" for _ in range(over_block.rows)]))
feature = ".".join(["0", str(tier+1), "1"]) # feature = ".".join(["0", str(tier+1), "1"])
columns = [feature, "T", "Sigma", "Z"] # columns = [feature, "T", "Sigma", "Z"]
source = self.source.new_learnblock( # source = self.source.new_learnblock(
values=data, columns=columns, index=over_block.indexes, # values=data, columns=columns, index=over_block.indexes,
origin=[p_model.uid, r_model.uid]) # origin=[p_model.uid, r_model.uid])
TS_QUEUE.append((tier+1, source)) TS_QUEUE.append((tier+1, overblock))
success = True success = True
if not success: if not success:
...@@ -421,24 +423,17 @@ class Deconstructor: ...@@ -421,24 +423,17 @@ class Deconstructor:
# Get learnblock that trained relative model # Get learnblock that trained relative model
second_block = r_model.trained_with(self.source) second_block = r_model.trained_with(self.source)
# Get samples that have overlapping rows overlapping_block = block.same_features_fusion(second_block)
overlapping_block = block.overlapping_rows(second_block)
# Check constraint # Check constraint
if overlapping_block.rows >= 2: if overlapping_block.n_features >= 2:
# Model fusion # Model fusion
new_model = p_model.fusion( new_model = p_model.fusion(r_model, self.NEXT_MODEL_COUNTER(tier))
r_model, self.NEXT_MODEL_COUNTER(tier))
which_ml_models = new_model.sigma which_ml_models = new_model.sigma
# Get learnblock
train_block = block.fusion(second_block)
try: try:
# Reconstruct model # Reconstruct model
recon_m = self.reconstructor.reconstruct( recon_m = self.reconstructor.reconstruct(
tier, train_block, which_ml_models, new_model) tier, overlapping_block, which_ml_models, new_model)
self.knowledge_database.replace(r_model, recon_m) self.knowledge_database.replace(r_model, recon_m)
success = True success = True
...@@ -516,7 +511,7 @@ class Deconstructor: ...@@ -516,7 +511,7 @@ class Deconstructor:
# Create submodel from TSgima relative samples # Create submodel from TSgima relative samples
second_block = r_model.trained_with(self.source) second_block = r_model.trained_with(self.source)
new_block = block.fusion(second_block) new_block = block.same_features_fusion(second_block)
ts_relatives = self.source.time_sigma_relatives(new_block) ts_relatives = self.source.time_sigma_relatives(new_block)
which_ml_models = p_model.subject + r_model.subject which_ml_models = p_model.subject + r_model.subject
self.reconstructor.reconstruct( self.reconstructor.reconstruct(
...@@ -527,7 +522,7 @@ class Deconstructor: ...@@ -527,7 +522,7 @@ class Deconstructor:
# Create learnblock # Create learnblock
first_block = p_model.trained_with(self.source) first_block = p_model.trained_with(self.source)
second_block = r_model.trained_with(self.source) second_block = r_model.trained_with(self.source)
new_block = first_block.fusion(second_block) new_block = first_block.same_features_fusion(second_block)
which_ml_models = new_model.sigma which_ml_models = new_model.sigma
try: try:
...@@ -583,18 +578,25 @@ class Deconstructor: ...@@ -583,18 +578,25 @@ class Deconstructor:
set(first_block.columns()).intersection(set(second_block.columns())) set(first_block.columns()).intersection(set(second_block.columns()))
) )
def calc_reliability(self, def calculate_reliability(self, predicts_a, predicts_b):
model_a: PragmaticMachineLearningModel, predictions = [predicts_a, predicts_b]
model_b: PragmaticMachineLearningModel,
block):
y_one = model_a.model.predict(block.as_numpy_array())
y_two = model_b.model.predict(block.as_numpy_array())
reliability_data = [y_one, y_two]
if self.reconstructor.category == "conceptual": if self.reconstructor.category == "conceptual":
return krippendorff.alpha(reliability_data, return krippendorff.alpha(predictions, level_of_measurement="nomimal")
level_of_measurement="nominal") elif self.reconstructor.category:
elif self.reconstructor.category == "procedural": return krippendorff.alpha(predictions, level_of_measurement="ration")
return krippendorff.alpha(reliability_data, #
level_of_measurement="ratio") # def calc_reliability(self,
else: # model_a: PragmaticMachineLearningModel,
raise ValueError() # model_b: PragmaticMachineLearningModel,
# block):
# y_one = model_a.model.predict(block.as_numpy_array())
# y_two = model_b.model.predict(block.as_numpy_array())
# reliability_data = [y_one, y_two]
# if self.reconstructor.category == "conceptual":
# return krippendorff.alpha(reliability_data,
# level_of_measurement="nominal")
# elif self.reconstructor.category == "procedural":
# return krippendorff.alpha(reliability_data,
# level_of_measurement="ratio")
# else:
# raise ValueError()
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment