def testWassersteinNetwork(self):
batch_n = 5
holder = forgetting_nuisance.DatasetHolder(self.dummy_df)
holder.add_shufflers(["compound", "batch"])
input_dim = holder.input_dim
feature_dim = input_dim
network = forgetting_nuisance.WassersteinNetwork(
holder, feature_dim, batch_n, 0, 1)
self.assertEqual(set(network._unique_targets),
set([("drug", ), ("control", )]))
self.assertEqual(set(network._unique_nuisances),
set([("batch1", ), ("batch2", )]))
keys_for_targets = {
("drug", ): [(("drug", "batch1"), None),
(("drug", "batch2"), None)],
("control", ): [(("control", "batch1"), None),
(("control", "batch2"), None)]
}
self.assertSameElements(list(network._keys_for_targets.keys()),
list(keys_for_targets.keys()))
for key in keys_for_targets:
self.assertSameElements(keys_for_targets[key],
network._keys_for_targets[key])
with tf.Session() as sess:
tf.global_variables_initializer().run()
feed_dict = {}
for key, shuffler in holder.data_shufflers.items():
input_mini, _ = shuffler.next_batch(batch_n)
feed_dict[network._x_vals[key]] = input_mini
f_val, x_vals = sess.run([network._features, network._x_vals],
feed_dict=feed_dict)
## make sure each input used in every batch came from the actual inputs
for key, vals in x_vals.items():
for row in vals: # iterate over every element
## identify distance from closest element in inputs
differences = [
la.norm(np.array(row) - np.array(candidates))
for candidates in holder.data_shufflers[key].inputs
]
self.assertAlmostEqual(min(differences), 0.0, places=5)
self.assertEqual(list(f_val.values())[0].shape, (batch_n, input_dim))
self.assertSameElements(list(network.wass_loss_target.keys()),
[("drug", ), ("control", )])
def testMakeHolder2(self):
## make a holder in which the dataframe is formatted by "one_hot".
## In addition batch information is provided in the inputs.
holder = forgetting_nuisance.DatasetHolder(
self.dummy_df,
input_category_level=["batch"],
batch_input_info="one_hot")
## add input/output shufflers for each batch, with compound label outputs.
holder.add_shufflers(["batch"], ["compound"], None)
## Pick the shuffler from batch1 to compounds.
test_shuffler = holder.data_shufflers[("batch1", "compound")]
inputs, outputs = (forgetting_nuisance.get_dense_arr(
test_shuffler.inputs),
forgetting_nuisance.get_dense_arr(
test_shuffler.outputs))
## number of elements shoud be num_compounds * self.num.
## input dim should be (dim + 1) * num_batches for "multiplexed" format.
## output dim should be num_labels.
self.assertEqual(inputs.shape, (2 * self.num, 4))
self.assertEqual(outputs.shape, (2 * self.num, 2))
## Confirm the format is correct.
zeros_chunck = np.zeros((self.num))
ones_chunck = np.ones((self.num))
np.testing.assert_array_equal(inputs[:self.num, 2], ones_chunck)
np.testing.assert_array_equal(inputs[:self.num, 3], zeros_chunck)
np.testing.assert_array_equal(outputs[:self.num, 0], ones_chunck)
np.testing.assert_array_equal(outputs[:self.num, 1], zeros_chunck)
np.testing.assert_array_equal(inputs[self.num:, 2], ones_chunck)
np.testing.assert_array_equal(inputs[self.num:, 3], zeros_chunck)
np.testing.assert_array_equal(outputs[self.num:, 0], zeros_chunck)
np.testing.assert_array_equal(outputs[self.num:, 1], ones_chunck)
## add input/output shufflers for each compound, with batch label outputs.
holder.add_shufflers(["compound"], ["batch"], None)
test_shuffler = holder.data_shufflers[("control", "batch")]
inputs, outputs = (forgetting_nuisance.get_dense_arr(
test_shuffler.inputs),
forgetting_nuisance.get_dense_arr(
test_shuffler.outputs))
self.assertEqual(inputs.shape, (2 * self.num, 4))
self.assertEqual(outputs.shape, (2 * self.num, 2))
## Confirm the format is correct.
zeros_chunck = np.zeros((self.num))
ones_chunck = np.ones((self.num))
np.testing.assert_array_equal(inputs[:self.num, 2], ones_chunck)
np.testing.assert_array_equal(inputs[:self.num, 3], zeros_chunck)
np.testing.assert_array_equal(outputs[:self.num, 0], ones_chunck)
np.testing.assert_array_equal(outputs[:self.num, 1], zeros_chunck)
np.testing.assert_array_equal(inputs[self.num:, 2], zeros_chunck)
np.testing.assert_array_equal(inputs[self.num:, 3], ones_chunck)
np.testing.assert_array_equal(outputs[self.num:, 0], zeros_chunck)
np.testing.assert_array_equal(outputs[self.num:, 1], ones_chunck)
## add input/output shufflers from groups specified by compound and batch
## to output labels specified by batch.
## Allow only shufflers that have control compounds only.
holder.add_shufflers(["compound", "batch"], ["batch"],
[("control", "batch1"), ("control", "batch2")])
test_shuffler = holder.data_shufflers[("control", "batch1"), "batch"]
inputs, outputs = (forgetting_nuisance.get_dense_arr(
test_shuffler.inputs),
forgetting_nuisance.get_dense_arr(
test_shuffler.outputs))
## There should be 10 points in each set.
self.assertEqual(inputs.shape, (self.num, 4))
## There is only one possible output label for this set since both batch and
## compound have been fixed.
self.assertEqual(outputs.shape, (self.num, 1))
