def testFromMultipleDataValidation(self):
# Non-MapData raises an error
with self.assertRaisesRegex(ValueError, "Non-MapData in inputs."):
data_elt_A = Data(df=pd.DataFrame([{
"arm_name": "0_1",
"mean": 3.7,
"sem": 0.5,
"metric_name": "b",
}]), )
data_elt_B = Data(df=pd.DataFrame([{
"arm_name": "0_1",
"mean": 3.7,
"sem": 0.5,
"metric_name": "b",
}]), )
MapData.from_multiple_data([data_elt_A, data_elt_B])
# Inconsistent keys raise an error
with self.assertRaisesRegex(
ValueError, "Inconsistent map_keys found in data iterable."):
data_elt_A = MapData(
df=pd.DataFrame([{
"arm_name": "0_1",
"mean": 3.7,
"sem": 0.5,
"metric_name": "b",
"epoch": 0,
}]),
map_keys=["epoch"],
)
data_elt_B = MapData(
df=pd.DataFrame([{
"arm_name": "0_1",
"mean": 3.7,
"sem": 0.5,
"metric_name": "b",
"iteration": 1,
}]),
map_keys=["iteration"],
)
MapData.from_multiple_data([data_elt_A, data_elt_B])
def testFetchTrialsData(self):
exp = self._setupBraninExperiment(n=5)
batch_0 = exp.trials[0]
batch_1 = exp.trials[1]
batch_0.mark_completed()
batch_1.mark_completed()
batch_0_data = exp.fetch_trials_data(trial_indices=[0])
self.assertEqual(set(batch_0_data.df["trial_index"].values), {0})
self.assertEqual(set(batch_0_data.df["arm_name"].values),
{a.name
for a in batch_0.arms})
batch_1_data = exp.fetch_trials_data(trial_indices=[1])
self.assertEqual(set(batch_1_data.df["trial_index"].values), {1})
self.assertEqual(set(batch_1_data.df["arm_name"].values),
{a.name
for a in batch_1.arms})
self.assertEqual(
exp.fetch_trials_data(trial_indices=[0, 1]),
MapData.from_multiple_data([batch_0_data, batch_1_data]),
)
# Since NoisyFunctionMap metric has overwrite_existing_data = True,
# we should only have one df per trial now
self.assertEqual(len(exp.data_by_trial[0]), 1)
with self.assertRaisesRegex(ValueError, ".* not associated .*"):
exp.fetch_trials_data(trial_indices=[2])
# Try to fetch data when there are only metrics and no attached data.
exp.remove_tracking_metric(
metric_name="b") # Remove implemented metric.
exp.add_tracking_metric(
MapMetric(name="b")) # Add unimplemented metric.
self.assertEqual(len(exp.fetch_trials_data(trial_indices=[0]).map_df),
30)
# Try fetching attached data.
exp.attach_data(batch_0_data)
exp.attach_data(batch_1_data)
self.assertEqual(exp.fetch_trials_data(trial_indices=[0]),
batch_0_data)
self.assertEqual(exp.fetch_trials_data(trial_indices=[1]),
batch_1_data)
self.assertEqual(set(batch_0_data.df["trial_index"].values), {0})
self.assertEqual(set(batch_0_data.df["arm_name"].values),
{a.name
for a in batch_0.arms})
def testFromMultipleData(self):
data = [
MapData(
df=pd.DataFrame(
[
{
"arm_name": "0_1",
"mean": 3.7,
"sem": 0.5,
"metric_name": "b",
"epoch": 0,
},
]
),
map_keys=["epoch"],
),
MapData(
df=pd.DataFrame(
[
{
"arm_name": "0_1",
"mean": 3.7,
"sem": 0.5,
"metric_name": "b",
"epoch": 0,
},
{
"arm_name": "0_1",
"mean": 3.7,
"sem": 0.5,
"metric_name": "b",
"epoch": 1,
},
]
),
map_keys=["epoch"],
),
]
merged_data = MapData.from_multiple_data(data)
self.assertIsInstance(merged_data, MapData)
self.assertEqual(3, merged_data.df.shape[0])
def testFetchTrialsData(self):
exp = self._setupBraninExperiment(n=5)
batch_0 = exp.trials[0]
batch_1 = exp.trials[1]
batch_0.mark_completed()
batch_1.mark_completed()
batch_0_data = exp.fetch_trials_data(trial_indices=[0])
self.assertEqual(set(batch_0_data.df["trial_index"].values), {0})
self.assertEqual(set(batch_0_data.df["arm_name"].values),
{a.name
for a in batch_0.arms})
batch_1_data = exp.fetch_trials_data(trial_indices=[1])
self.assertEqual(set(batch_1_data.df["trial_index"].values), {1})
self.assertEqual(set(batch_1_data.df["arm_name"].values),
{a.name
for a in batch_1.arms})
self.assertEqual(
exp.fetch_trials_data(trial_indices=[0, 1]),
MapData.from_multiple_data([batch_0_data, batch_1_data]),
)
with self.assertRaisesRegex(ValueError, ".* not associated .*"):
exp.fetch_trials_data(trial_indices=[2])
# Try to fetch data when there are only metrics and no attached data.
exp.remove_tracking_metric(
metric_name="b") # Remove implemented metric.
exp.add_tracking_metric(
MapMetric(name="b")) # Add unimplemented metric.
self.assertTrue(exp.fetch_trials_data(trial_indices=[0]).df.empty)
# Try fetching attached data.
exp.attach_data(batch_0_data)
exp.attach_data(batch_1_data)
self.assertEqual(exp.fetch_trials_data(trial_indices=[0]),
batch_0_data)
self.assertEqual(exp.fetch_trials_data(trial_indices=[1]),
batch_1_data)
self.assertEqual(set(batch_0_data.df["trial_index"].values), {0})
self.assertEqual(set(batch_0_data.df["arm_name"].values),
{a.name
for a in batch_0.arms})
def test_combine(self):
mmd_double = MapData.from_multiple_map_data([self.mmd, self.mmd])
self.assertEqual(mmd_double.map_df.size, 2 * self.mmd.map_df.size)
self.assertEqual(mmd_double.map_key_infos, self.mmd.map_key_infos)
different_map_df = pd.DataFrame([
{
"arm_name": "0_3",
"timestamp": 11,
"mean": 2.0,
"sem": 0.2,
"trial_index": 1,
"metric_name": "a",
},
{
"arm_name": "0_3",
"timestamp": 18,
"mean": 1.8,
"sem": 0.3,
"trial_index": 1,
"metric_name": "b",
},
])
different_map_key_infos = [
MapKeyInfo(key="timestamp", default_value=0.0)
]
different_mmd = MapData(df=different_map_df,
map_key_infos=different_map_key_infos)
combined = MapData.from_multiple_map_data([self.mmd, different_mmd])
self.assertEqual(len(combined.map_df),
len(self.mmd.map_df) + len(different_mmd.map_df))
self.assertEqual(combined.map_df.columns.size,
self.mmd.map_df.columns.size + 1)
self.assertEqual(combined.map_key_infos,
self.map_key_infos + different_map_key_infos)
combined_subset = MapData.from_multiple_map_data(
[self.mmd, different_mmd], ["a"])
self.assertTrue((combined_subset.map_df["metric_name"] == "a").all())
data_df = pd.DataFrame([
{
"arm_name": "0_4",
"mean": 2.0,
"sem": 0.2,
"trial_index": 1,
"metric_name": "a",
},
{
"arm_name": "0_4",
"mean": 1.8,
"sem": 0.3,
"trial_index": 1,
"metric_name": "b",
},
])
data = Data(df=data_df)
downcast_combined = MapData.from_multiple_data([self.mmd, data])
self.assertEqual(len(downcast_combined.map_df),
len(self.mmd.map_df) + len(data.df))
self.assertEqual(downcast_combined.map_df.columns.size,
self.mmd.map_df.columns.size)
self.assertEqual(downcast_combined.map_key_infos, self.map_key_infos)
# Check that the Data's rows' epoch cell has the correct default value
self.assertTrue((downcast_combined.map_df[
downcast_combined.map_df["arm_name"] == "0_4"]["epoch"] ==
self.mmd.map_key_infos[0].default_value).all())
