Incremental Learning

from vyntri import Vyntri
from vyntri.data import split

s = split("./initial_data", train=0.7, test=0.2, seed=42)
model = Vyntri()
model.fit(train=s.train, val=s.val)

# Add more images for existing classes
result = model.update("./more_data")
print(f"Samples added: {result.n_samples}")
print(f"New classes: {result.new_classes}")  # Empty — all classes existed

from vyntri import Vyntri
from vyntri.data import split

s = split("./initial_data", train=0.7, test=0.2, seed=42)
model = Vyntri()
model.fit(train=s.train, val=s.val)
print(f"Classes: {list(model.classes_)}")  # ['cats', 'dogs']

# Add new classes — detected automatically
result = model.update("./new_classes_data")
print(f"New classes: {result.new_classes}")  # ['birds']
print(f"Classes now: {result.classes_after}")  # ['birds', 'cats', 'dogs']

from vyntri import Vyntri
from vyntri.data import split

s = split("./data_v1", train=0.7, test=0.2, seed=42)
model = Vyntri()
model.fit(train=s.train, val=s.val)
print(f"After v1: {list(model.classes_)}")

model.update("./data_v2")
print(f"After v2: {list(model.classes_)}")

model.update("./data_v3")
print(f"After v3: {list(model.classes_)}")

# Evaluate at any point
result = model.evaluate("./test_data")
print(f"Accuracy: {result.accuracy:.1%}")

from vyntri import Vyntri
from vyntri.data import split

# Approach 1: Sequential updates
model = Vyntri(seed=42)
s1 = split("./data_part1", train=0.7, test=0.2, seed=42)
model.fit(train=s1.train, val=s1.val)
model.update("./data_part2")
model.update("./data_part3")
seq_acc = model.evaluate("./test").accuracy

# Approach 2: Joint refit on all data
joint = Vyntri(seed=42)
joint.fit("./all_data")
joint_acc = joint.evaluate("./test").accuracy

print(f"Sequential: {seq_acc:.3f}")
print(f"Joint:      {joint_acc:.3f}")
# These match — sufficient statistics produce the same result