def test_optimizer_by_mock(X: InteractionMatrix, target_epoch: int) -> None:
X_train, X_val = rowwise_train_test_split(X)
evaluator = Evaluator(X_val, 0)
optimizer = MockOptimizer(
X_train,
evaluator,
fixed_params=dict(X_test=X_val.toarray(), target_epoch=target_epoch),
logger=getLogger("IGNORE"),
)
config, _ = optimizer.optimize(n_trials=1, random_seed=42)
assert config["max_epoch"] == target_epoch
def test_optimizer_by_mock(X: sps.csr_matrix) -> None:
X_train, X_val = rowwise_train_test_split(X, test_ratio=0.5, random_seed=0)
evaluator = Evaluator(X_val, 0)
optimizer = MockOptimizer(X_train,
evaluator,
logger=None,
fixed_params=dict(X_test=X_val))
config, _ = optimizer.optimize(n_trials=40, random_seed=42)
assert config["p1"] >= 0.9
assert (config["reg"] >= 0.99) and (config["reg"] <= 1.01)
assert (config["I1"] >= 100) and (config["I1"] <= 102)
assert (config["I2"] >= 500) and (config["I2"] <= 502)
assert config["flag"] in ["foo", "bar"]
def test_cb2cf(X: InteractionMatrix, profile: ProfileMatrix) -> None:
"""Fit IALS & let mlp overfit.
Args:
X (InteractionMatrix): user_item interaction matrix
profile (ProfileMatrix): profile
"""
try:
from irspack.user_cold_start.cb2cf import CB2IALSOptimizer
except:
pytest.skip("Failed to import jax.")
raise
X_cf_train_all, X_val = rowwise_train_test_split(X_cf,
test_ratio=0.5,
random_seed=0)
evaluator = Evaluator(X_val, 0)
optim = CB2IALSOptimizer(
X_cf_train_all,
evaluator,
profile,
)
cb2cfrec, t, mlp_config = optim.search_all(
20,
cf_fixed_params=dict(n_components=5,
alpha=0,
reg=1e-3,
max_cg_steps=30),
random_seed=0,
)
vec_reconstruction = cb2cfrec.mlp.predict(
profile.astype(np.float32).toarray())
vec_target = cb2cfrec.cf_rec.get_user_embedding()
residual = (
(vec_reconstruction - vec_target)**2).sum() / (vec_target**2).sum()
assert residual <= 1e-1
def test_metrics_ColdUser(U: int, I: int, U_test: int) -> None:
rns = np.random.RandomState(42)
uvec = rns.randn(U + U_test, 3)
ivec = rns.randn(I, 3)
true_score = uvec.dot(ivec.T) # + rns.randn(U, I)
X = sps.csr_matrix((true_score > 0).astype(np.float64))
X_train = X[:U]
X_val = X[U:]
X_val_learn, X_val_target = rowwise_train_test_split(X_val, random_seed=0)
X_train_all = sps.vstack([X_train, X_val_learn])
hot_evaluator = Evaluator(
sps.csr_matrix(X_val_target), offset=U, cutoff=I // 2, n_threads=2
)
rec = P3alphaRecommender(X_train_all)
rec.learn()
hot_score = hot_evaluator.get_score(rec)
with pytest.warns(UserWarning):
cold_evaluator = EvaluatorWithColdUser(
X_val_learn.tocsc(), X_val_target, cutoff=I // 2, mb_size=5
) # csc matrix input should raise warning about
# memory ordering, as csc-csc matrix product will be csc,
# hence col-major matrix when made dense.
cold_score = cold_evaluator.get_score(rec)
shuffle_index = np.arange(X_val_learn.shape[0])
rns.shuffle(shuffle_index)
cold_evaluator_shuffled = EvaluatorWithColdUser(
X_val_learn[shuffle_index], X_val_target[shuffle_index], cutoff=I // 2
)
cold_score_shuffled = cold_evaluator_shuffled.get_score(rec)
for key in cold_score:
assert cold_score_shuffled[key] == pytest.approx(cold_score[key])
for key in hot_score:
assert hot_score[key] == pytest.approx(cold_score[key], abs=1e-8)