def test_process_next_active(self):
data = simpledataset()
sup = simplesupervisor(data[1])
agent = ag.ActiveAgent(0.5,
1,
ag.ObjectsMemory(),
ag.SupervisionMemory(),
simplemodel(),
sup,
bootstrap=2,
max_neigh_check=1,
add_seen_element=ag.add_seen_separate)
output = [
agent.process_next([data[0][itx]], data[2][itx])
for itx in range(len(data[0]))
]
asked_sup = np.array([o[2] for o in output])
for itx in range(len(output)):
with self.subTest(n=itx):
if itx > 0:
self.assertTrue(output[itx][1] < itx)
self.assertTrue(len(agent.sup_mem), asked_sup.sum())
def test_predict(self):
data = simpledataset()
sup = simplesupervisor(data[1])
agent = ag.ActiveAgent(0.5,
1,
ag.ObjectsMemory(),
ag.SupervisionMemory(),
simplemodel(),
sup,
bootstrap=2,
max_neigh_check=1,
add_seen_element=ag.add_seen_separate)
output = [
agent.process_next([data[0][itx]], data[2][itx])
for itx in range(len(data[0]))
]
predictions = [agent.predict([d])[1] for d in data[0]]
is_known = [agent.predict([d])[0] for d in data[0]]
npt.assert_equal(np.concatenate(predictions), data[2])
all_pred = agent.predict(list(data[0]))
npt.assert_equal(np.concatenate(is_known), all_pred[0])
npt.assert_equal(np.concatenate(predictions), all_pred[1])
npt.assert_equal(agent.predict(list(data[0])), all_pred)
def test_refine(self):
data = simpledataset()
sup = simplesupervisor(data[1])
def refine(agent):
optim = torch.optim.sgd()
l = loss.ContrastiveLoss()
e = ag.create_siamese_trainer(agent, optim, l)
sampler = samp.SeadableRandomSampler(agent.sup_mem, 1)
data_loader = torch.utils.data.DataLoader(agent.sup_mem,
sampler=sampler)
e.run(data_loader, max_epochs=2)
agent = ag.Agent(1,
ag.ObjectsMemory(),
ag.SupervisionMemory(),
simplemodel(),
sup,
bootstrap=2,
max_neigh_check=1,
add_seen_element=ag.add_seen_separate,
refine=refine)
for itx in range(len(data[0])):
with self.subTest(n=itx):
out = agent.process_next([data[0][itx]], data[2][itx])
if itx < 2 and itx != 0:
self.assertTrue(out[2])
self.assertEqual(len(agent.obj_mem), itx + 1)
self.assertEqual(len(agent.sup_mem), itx)
def test_process_next_internals(self):
data = simpledataset()
sup = simplesupervisor(data[1])
agent = ag.Agent(1,
ag.ObjectsMemory(),
ag.SupervisionMemory(),
simplemodel(),
sup,
bootstrap=2,
max_neigh_check=1,
add_seen_element=ag.add_seen_separate)
for itx in range(len(data[0])):
with self.subTest(n=itx):
out = agent.process_next([data[0][itx]], data[2][itx])
if itx < 2 and itx != 0:
self.assertTrue(out[2])
self.assertEqual(len(agent.obj_mem), itx + 1)
self.assertEqual(len(agent.sup_mem), itx)
data = simpledataset()
sup = simplesupervisor(data[1])
def test_del_entry(self):
m = ag.SupervisionMemory()
d1, l1 = bogus_data(), ([0], [3.0])
d2, l2 = bogus_data(), ([1], [0.5])
d3, l3 = bogus_data(), ([0], [1.5])
d4, l4 = bogus_data(), ([1], [1.0])
m.add_entry(d1, *l1)
m.add_entry(d2, *l2)
m.add_entry(d3, *l3)
m.add_entry(d4, *l4)
m.del_entry()
npt.assert_equal(m.labels, np.array([1, 1, 0]))
npt.assert_equal(m.distances, np.array([0.5, 1.0, 1.5]))
npt.assert_equal(m.insertion_orders, np.array([1, 3, 2]))
npt.assert_equal(m.couples, [d2, d4, d3])
self.assertEqual(len(m), 3)
m.del_entry(1)
npt.assert_equal(m.labels, np.array([1, 0]))
npt.assert_equal(m.distances, np.array([0.5, 1.5]))
npt.assert_equal(m.insertion_orders, np.array([1, 2]))
npt.assert_equal(m.couples, [d2, d3])
self.assertEqual(len(m), 2)
def af(seed, sup):
torch.manual_seed(seed)
agent = ag.Agent(seed,
ag.ObjectsMemory(),
ag.SupervisionMemory(),
simplemodel(),
sup,
bootstrap=2,
max_neigh_check=1,
add_seen_element=ag.add_seen_separate)
return agent
def test_supervision(self):
data = simpledataset()
sup = simplesupervisor(data[1])
agent = ag.ActiveAgent(1.0,
1,
ag.ObjectsMemory(),
ag.SupervisionMemory(),
simplemodel(),
sup,
bootstrap=2,
max_neigh_check=1,
add_seen_element=ag.add_seen_separate)
output = np.array([
agent.process_next([data[0][itx]], data[2][itx])
for itx in range(len(data[0]))
])
self.assertTrue(output[1:, 2].any())
agent = ag.ActiveAgent(0.01,
1,
ag.ObjectsMemory(),
ag.SupervisionMemory(),
simplemodel(),
sup,
bootstrap=2,
max_neigh_check=1,
add_seen_element=ag.add_seen_separate)
output = np.array([
agent.process_next([data[0][itx]], data[2][itx])
for itx in range(len(data[0]))
])
self.assertFalse(output[1:, 2].all())
def test_getitem(self):
m = ag.SupervisionMemory()
d1, l1 = bogus_data(), ([0], [3.0])
d2, l2 = bogus_data(), ([1], [0.5])
d3, l3 = bogus_data(), ([0], [1.5])
d4, l4 = bogus_data(), ([1], [1.0])
m.add_entry(d1, *l1)
m.add_entry(d2, *l2)
m.add_entry(d3, *l3)
m.add_entry(d4, *l4)
item = m[2]
npt.assert_equal(item, (d3, l3[0][0]))
m.del_entry(1)
item = m[1]
npt.assert_equal(item, (d3, l3[0][0]))
def test_process_next_out(self):
data = simpledataset()
sup = simplesupervisor(data[1])
agent = ag.Agent(1,
ag.ObjectsMemory(),
ag.SupervisionMemory(),
simplemodel(),
sup,
bootstrap=2,
max_neigh_check=1,
add_seen_element=ag.add_seen_separate)
output = [
agent.process_next([data[0][itx]], data[2][itx])
for itx in range(len(data[0]))
]
for itx in range(1, len(output)):
with self.subTest(n=itx):
self.assertTrue(output[itx][1] < itx)
self.assertTrue(output[itx][2])