def test_FeatureReducer(self):
df = self.test_df
target = 'gap expt'
fr = FeatureReducer()
# ultra-basic case: are we reducing at least 1 feature?
df = fr.fit_transform(df, target)
self.assertTrue(df.shape[1] < self.test_df.shape[1])
# ensure metadata is being written correctly
self.assertTrue(target not in fr.retained_features)
self.assertTrue(len(list(fr.removed_features.keys())) == 2)
# ensure other combinations of feature reducers are working
fr = FeatureReducer(reducers=('corr', 'rebate'), n_rebate_features=40)
df = fr.fit_transform(self.test_df, target)
self.assertEqual(df.shape[1], 41) # 40 features + target
self.assertTrue(target in df.columns)
# ensure the same thing works when fraction is used
fr = FeatureReducer(reducers=('rebate', ), n_rebate_features=0.2)
df = fr.fit_transform(self.test_df, target)
self.assertEqual(df.shape[1], 83 + 1)
# test transferability
df2 = self.test_df
df2 = fr.transform(df2, target)
self.assertListEqual(df.columns.tolist(), df2.columns.tolist())
def test_FeatureReducer_classification(self):
# test classification with corr matrix (no errors)
fr = FeatureReducer(reducers=("corr", ))
df_class = self.test_df
df_class[self.target] = [
"semiconductor" if 0.0 < g < 3.0 else "nonmetal"
for g in df_class[self.target]
]
df_class = fr.fit_transform(df_class, self.target)
self.assertEqual(df_class.shape[0], 200)
def test_FeatureReducer_transferability(self):
# ensure the same thing works when fraction is used
fr = FeatureReducer(reducers=("rebate", ), n_rebate_features=0.2)
df = fr.fit_transform(self.test_df, self.target)
self.assertEqual(df.shape[1], 83 + 1)
# test transferability
df2 = deepcopy(self.test_df)
df2 = fr.transform(df2, self.target)
self.assertListEqual(df.columns.tolist(), df2.columns.tolist())
def test_FeatureReducer_basic(self):
fr = FeatureReducer(reducers=("corr", "tree"))
# ultra-basic case: are we reducing at least 1 feature?
df = fr.fit_transform(self.test_df, self.target)
self.assertTrue(df.shape[1] < self.test_df.shape[1])
# ensure metadata is being written correctly
self.assertTrue(self.target not in fr.retained_features)
self.assertTrue(len(list(fr.removed_features.keys())) == 2)
def test_manual_feature_reduction(self):
fr = FeatureReducer(reducers=[], remove_features=["LUMO_element_Th"])
# ultra-basic case: are we reducing at least 1 feature?
df = fr.fit_transform(self.test_df, self.target)
self.assertTrue("LUMO_element_Th" not in df.columns)
self.assertEqual(fr.removed_features["manual"], ["LUMO_element_Th"])
# test removing feature that doesn't exist
fr = FeatureReducer(reducers=[], remove_features=["abcdefg12345!!"])
with self.assertLogs(AMM_LOGGER_BASENAME, level="INFO") as cm:
# should give log warning but not throw an error
fr.fit_transform(self.test_df, self.target)
self.assertTrue("abcdefg12345!!" in " ".join(cm.output))
def test_FeatureReducer_pca(self):
# Case where n_samples < n_features
df = self.test_df.iloc[:10]
fr = FeatureReducer(reducers=("pca", ), n_pca_features='auto')
df_reduced = fr.fit_transform(df, self.target)
self.assertTupleEqual(df_reduced.shape, (10, 11))
# Case where n_samples > n_features
fsubset = [
'HOMO_energy', 'LUMO_energy', 'gap_AO', 'minimum X', 'maximum X',
'range X', 'mean X', 'std_dev X', 'minimum row', 'maximum row',
'range row', 'mean row', 'std_dev row', 'minimum group',
'maximum group', 'range group', 'mean group', 'std_dev group'
]
df = self.test_df[fsubset + [self.target]]
df_reduced = fr.fit_transform(df, self.target)
self.assertEqual(df_reduced.shape[0], 200)
# Manually specified case of n_samples > n_features
fr = FeatureReducer(reducers=('pca', ), n_pca_features=0.5)
df = self.test_df
df_reduced = fr.fit_transform(df, self.target)
self.assertTupleEqual(df_reduced.shape, (200, 201))
def test_FeatureReducer_pca(self):
# Case where n_samples < n_features
df = self.test_df.iloc[:10]
fr = FeatureReducer(reducers=("pca", ), n_pca_features="auto")
df_reduced = fr.fit_transform(df, self.target)
self.assertTupleEqual(df_reduced.shape, (10, 11))
# Case where n_samples > n_features
fsubset = [
"HOMO_energy",
"LUMO_energy",
"gap_AO",
"minimum X",
"maximum X",
"range X",
"mean X",
"std_dev X",
"minimum row",
"maximum row",
"range row",
"mean row",
"std_dev row",
"minimum group",
"maximum group",
"range group",
"mean group",
"std_dev group",
]
df = self.test_df[fsubset + [self.target]]
df_reduced = fr.fit_transform(df, self.target)
self.assertEqual(df_reduced.shape[0], 200)
# Manually specified case of n_samples > n_features
fr = FeatureReducer(reducers=("pca", ), n_pca_features=0.5)
df = self.test_df
df_reduced = fr.fit_transform(df, self.target)
self.assertTupleEqual(df_reduced.shape, (200, 201))
def test_saving_feature_from_removal(self):
fr = FeatureReducer(reducers=("corr", ), keep_features=["maximum X"])
# ultra-basic case: are we reducing at least 1 feature?
df = fr.fit_transform(self.test_df, self.target)
self.assertTrue("maximum X" in df.columns)
def test_FeatureReducer_combinations(self):
df = self.test_df
fr = FeatureReducer(reducers=("pca", "rebate", "tree"))
df_reduced = fr.fit_transform(df, self.target)
self.assertGreater(df.shape[1], df_reduced.shape[1])
def test_FeatureReducer_advanced(self):
# ensure other combinations of feature reducers are working
fr = FeatureReducer(reducers=("corr", "rebate"), n_rebate_features=40)
df = fr.fit_transform(self.test_df, self.target)
self.assertEqual(df.shape[1], 41) # 40 features + self.target
self.assertTrue(self.target in df.columns)