def test_file_io():
temp = utils.tempdir()
file_path = temp.relpath("temp.log")
tsk, target = get_sample_task()
inputs = [
MeasureInput(target, tsk, tsk.config_space.get(i))
for i in range(0, 10)
]
results = [MeasureResult((i, ), 0, 0, 0) for i in range(0, 10)]
invalid_inp = MeasureInput(target, tsk, tsk.config_space.get(10))
invalid_res = MeasureResult((10, ), 0, 0, 0)
# Erase the entity map to test if it will be ignored when loading back.
invalid_inp.config._entity_map = {}
with open(file_path, "w") as fo:
cb = autotvm.callback.log_to_file(fo)
cb(None, inputs, results)
cb(None, [invalid_inp], [invalid_res])
ref = zip(inputs, results)
for x, y in zip(ref, autotvm.record.load_from_file(file_path)):
assert x[1] == y[1]
def test_file_io():
temp = utils.tempdir()
file_path = temp.relpath("temp.log")
tsk, target = get_sample_task()
inputs = [
MeasureInput(target, tsk, tsk.config_space.get(i))
for i in range(0, 10)
]
results = [MeasureResult((i, ), 0, 0, 0) for i in range(0, 10)]
invalid_inp = MeasureInput(target, tsk, tsk.config_space.get(10))
invalid_res = MeasureResult((10, ), 0, 0, 0)
# Erase the entity map to test if it will be ignored when loading back.
invalid_inp.config._entity_map = {}
with open(file_path, "w") as fo:
cb = autotvm.callback.log_to_file(fo)
cb(None, inputs, results)
cb(None, [invalid_inp], [invalid_res])
ref = zip(inputs, results)
for x, y in zip(ref, autotvm.record.load_from_file(file_path)):
assert x[1] == y[1]
# Confirm functionality of multiple file loads
hist_best = ApplyHistoryBest([file_path, file_path])
x = hist_best.query(target, tsk.workload)
assert str(x) == str(inputs[0][2])
def test_update():
task, target = get_sample_task()
tuner = autotvm.tuner.XGBTuner(task)
n_records = 5
records = get_sample_records(n=n_records)
tuner.update([inp for inp, _ in records], [res for _, res in records])
assert len(tuner.xs) == n_records
assert len(tuner.ys) == n_records
assert len(tuner.visited) == n_records
assert all(x in tuner.visited for x in tuner.xs)
def test_tuner():
task, target = get_sample_task()
records = get_sample_records(n=10)
tuner = autotvm.tuner.XGBTuner(task)
tuner.load_history(records, min_seed_records=10)
# Confirm that loading history successfully loaded a
# base_model.
assert tuner.cost_model.base_model is not None
tuner = autotvm.tuner.XGBTuner(task)
tuner.load_history(records, min_seed_records=11)
# Confirm that loading history did not load base_model
# when not enough records according to `min_seed_records`
# are provided
assert tuner.cost_model.base_model is None
def test_apply_history_best():
tsk, target = get_sample_task()
records = [
(MeasureInput(target, tsk, tsk.config_space.get(0)),
MeasureResult((0.1, ), 0, 2.3, 0)),
(MeasureInput(target, tsk, tsk.config_space.get(1)),
MeasureResult((0.3, ), 0, 2.3, 0)),
(MeasureInput(target, tsk, tsk.config_space.get(2)),
MeasureResult((0.01, ), 0, 2.3, 0)),
(MeasureInput(target, tsk, tsk.config_space.get(4)),
MeasureResult((0.4, ), 0, 2.3, 0)),
]
hist_best = ApplyHistoryBest(records)
x = hist_best.query(target, tsk.workload)
assert str(x) == str(tsk.config_space.get(2))
def test_task_tuner_without_measurement():
"""test task and tuner without measurement"""
task, _ = get_sample_task()
measure_option = autotvm.measure_option(builder=autotvm.LocalBuilder(), runner=DummyRunner())
logging.info("%s", task.config_space)
for tuner_class in [
autotvm.tuner.RandomTuner,
autotvm.tuner.GridSearchTuner,
autotvm.tuner.GATuner,
autotvm.tuner.XGBTuner,
]:
tuner = tuner_class(task)
tuner.tune(n_trial=10, measure_option=measure_option)
assert tuner.best_flops > 1
def test_fit():
task, target = get_sample_task()
records = get_sample_records(n=500)
base_model = XGBoostCostModel(task,
feature_type="itervar",
loss_type="rank")
base_model.fit_log(records, plan_size=32)
upper_model = XGBoostCostModel(task,
feature_type="itervar",
loss_type="rank")
upper_model.load_basemodel(base_model)
xs = np.arange(10)
ys = np.arange(10)
upper_model.fit(xs, ys, plan_size=32)
def test_load_dump():
task, target = get_sample_task()
inp = MeasureInput(target, task, task.config_space.get(0))
result = MeasureResult((2.0, 2.23, 0.23, 0.123, 0.234, 0.123),
MeasureErrorNo.NO_ERROR, 2.3, time.time())
for protocol in ["json", "pickle"]:
row = encode(inp, result, protocol=protocol)
inp_2, result_2 = decode(row, protocol=protocol)
assert measure_str_key(inp) == measure_str_key(inp_2), "%s vs %s" % (
measure_str_key(inp),
measure_str_key(inp_2),
)
assert result.costs == result_2.costs
assert result.error_no == result_2.error_no
assert result.timestamp == result_2.timestamp
def test_random_tuner():
"""Test RandomTuner"""
task, _ = get_sample_task()
measure_option = autotvm.measure_option(builder=autotvm.LocalBuilder(),
runner=DummyRunner())
tuner = autotvm.tuner.RandomTuner(task, range_idx=(8, 15))
assert tuner.range_length == 8
assert tuner.index_offset == 8
# Tuner should only focus on the specified range and should visit all indices
tuner.tune(n_trial=8, measure_option=measure_option)
assert tuner.counter == 8
assert not tuner.has_next()
visited = set()
for idx in tuner.visited:
assert idx not in visited
assert 8 <= idx <= 15
def test_gridsearch_tuner():
"""Test GridSearchTuner"""
task, _ = get_sample_task()
measure_option = autotvm.measure_option(builder=autotvm.LocalBuilder(),
runner=DummyRunner())
# When no range index, range_length should be the length of config space
tuner = autotvm.tuner.GridSearchTuner(task)
assert tuner.range_length == len(task.config_space)
assert tuner.index_offset == 0
# With range index, range_length should be the length of the specified range
tuner = autotvm.tuner.GridSearchTuner(task, range_idx=(8, 15))
assert tuner.range_length == 8
assert tuner.index_offset == 8
# Tuner should only focus on the specified range
tuner.tune(n_trial=8, measure_option=measure_option)
assert tuner.counter == 8
assert not tuner.has_next()
def test_fit():
task, target = get_sample_task()
records = get_sample_records(n=500)
base_model = XGBoostCostModel(task,
feature_type="itervar",
loss_type="rank")
base_model.fit_log(records, plan_size=32)
upper_model = XGBoostCostModel(task,
feature_type="itervar",
loss_type="rank")
upper_model.load_basemodel(base_model)
xs = np.arange(10)
ys = np.arange(10)
upper_model.fit(xs, ys, plan_size=32)
# feature lengths are not guaranteed to always be the same
upper_model.predict(np.ones(12))
upper_model.predict(np.ones(8))
