def test_fit_params(teardown):
x_data = iris.data
y_t_data = iris.target
random_state = 123
n_components = 2
sample_weight = y_t_data + 1 # Just weigh the classes differently
fit_params = {"logreg__sample_weight": sample_weight}
# baikal way
x = Input()
y_t = Input()
x_pca = PCA(n_components=n_components, random_state=random_state, name="pca")(x)
y = LogisticRegression(
multi_class="multinomial",
solver="lbfgs",
random_state=random_state,
name="logreg",
)(x_pca, y_t)
model = Model(x, y, y_t)
model.fit(x_data, y_t_data, **fit_params)
# traditional way
pca = PCA(n_components=n_components, random_state=random_state)
logreg = LogisticRegression(
multi_class="multinomial", solver="lbfgs", random_state=random_state
)
pipe = Pipeline([("pca", pca), ("logreg", logreg)])
pipe.fit(x_data, y_t_data, **fit_params)
# Use assert_allclose instead of all equal due to small numerical differences
# between fit_transform(...) and fit(...).transform(...)
assert_allclose(model.get_step("logreg").coef_, pipe.named_steps["logreg"].coef_)
def test_transformed_target(teardown):
x = Input()
y_t = Input()
y_t_mod = Lambda(lambda y: np.log(y))(y_t)
y_p_mod = LinearRegression()(x, y_t_mod)
y_p = Lambda(lambda y: np.exp(y))(y_p_mod)
model = Model(x, y_p, y_t)
x_data = np.arange(4).reshape(-1, 1)
y_t_data = np.exp(2 * x_data).ravel()
model.fit(x_data, y_t_data)
assert_array_equal(model.get_step("LinearRegression_0").coef_, np.array([2.0]))
def test_steps_cache(teardown):
x1_data = iris.data[:, :2]
x2_data = iris.data[:, 2:]
y1_t_data = iris.target
x1 = Input(name="x1")
x2 = Input(name="x2")
y1_t = Input(name="y1_t")
y1 = LogisticRegression(name="LogReg")(x1, y1_t)
y2 = PCA(name="PCA")(x2)
hits, misses = 0, 0
# 1) instantiation always misses
misses += 1
model = Model([x1, x2], [y1, y2], y1_t)
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 2) calling fit for the first time, hence a miss
misses += 1
model.fit([x1_data, x2_data], y1_t_data)
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 3) same as above, just different format, hence a hit
hits += 1
model.fit({x1: x1_data, x2: x2_data}, {y1_t: y1_t_data})
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 4) trainable flags are considered in cache keys, hence a miss
misses += 1
model.get_step("LogReg").trainable = False
model.fit(
[x1_data, x2_data], y1_t_data
) # NOTE: target is superfluous, but it affects caching
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 5) same as above, just different format, hence a hit
hits += 1
model.fit({x1: x1_data, x2: x2_data}, y1_t_data)
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 6) we drop the (superflous) target, hence a miss
misses += 1
model.fit({x1: x1_data, x2: x2_data})
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 7) same as above, hence a hit
hits += 1
model.fit({x1: x1_data, x2: x2_data})
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 8) we restore the flag, becoming the same as 2) and 3), hence a hit
hits += 1
model.get_step("LogReg").trainable = True
model.fit({x1: x1_data, x2: x2_data}, y1_t_data)
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 9) new inputs/targets/outputs signature, hence a miss
misses += 1
model.predict([x1_data, x2_data])
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 10) same inputs/outputs signature as 9), hence a hit
hits += 1
model.predict({"x1": x1_data, "x2": x2_data}, ["PCA:0/0", "LogReg:0/0"])
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 11) new inputs/outputs signature, hence a miss
misses += 1
model.predict({x1: x1_data}, "LogReg:0/0")
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
# 12) same as above, hence a hit
hits += 1
model.predict({x1: x1_data}, "LogReg:0/0")
assert model._nodes_cache.hits == hits and model._nodes_cache.misses == misses
