def testModelRunsOnUCIData(self):
model = utils.get_uci_model()
model.compile(optimizer='adam', loss='binary_crossentropy')
data = utils.get_uci_data()
print('DATA:', data)
dataset = utils.df_to_dataset(data, batch_size=64)
# Model can train on UCI data
model.fit(dataset, epochs=1)
# Model can evaluate on UCI data
model.evaluate(dataset)
def testConvertToDataset(self):
expected_vals = range(10) # Arbitrary values
expected_labels = [0, 0, 1, 0, 1, 1, 1, 1, 0, 0] # Arbitrary labels.
data = {
'col': pd.Series(expected_vals),
'target': pd.Series(expected_labels)
}
df = pd.DataFrame(data)
dataset = utils.df_to_dataset(df)
vals, labels = zip(*[(val, label.numpy()) for val, label in dataset])
# Assert values are all dicts with exactly one column.
for val in vals:
self.assertSetEqual(set(val.keys()), set(['col']))
vals = [val['col'].numpy() for val in vals]
self.assertAllClose(vals, expected_vals)
self.assertAllClose(labels, expected_labels)
def testConvertToDatasetWithShuffle(self):
expected_vals = range(10) # Arbitrary values
expected_labels = [0, 0, 1, 0, 1, 1, 1, 1, 0, 0] # Arbitrary labels.
data = {
'col': pd.Series(expected_vals),
'target': pd.Series(expected_labels)
}
df = pd.DataFrame(data)
dataset = utils.df_to_dataset(df, shuffle=True)
vals, labels = zip(*[(val, label.numpy()) for val, label in dataset])
# Assert values are all dicts with exactly one column.
for val in vals:
self.assertSetEqual(set(val.keys()), set(['col']))
vals = [val['col'].numpy() for val in vals]
# These values should *NOT* be close because vals should be out of order
# since we set shuffle=True. Note that it seems like the tests provide a
# consistent seed so we don't have to worry about getting unlucky. If this
# changes then we can look into explicitly setting the a seed.
self.assertNotAllClose(vals, expected_vals)
# Assert that the contents are the same, just reordered
self.assertAllClose(sorted(vals), sorted(expected_vals))
self.assertAllClose(sorted(labels), sorted(expected_labels))