def test_dumbnet_predict(self):
gi = TwoPlayerGuessIt()
nn = NeuralNetwork(gi, DumbNet)
init = gi.get_initial_state()
p, v = nn.predict(init)
self.assertEqual(len(p), 4)
self.assertTrue((p == .25).all())
self.assertEqual(v, 0.0)
template = np.zeros_like(gi.get_available_actions(init))
template[0, 1] = 1
s = gi.take_action(init, template)
p, v = nn.predict(s)
self.assertEqual(len(p), 3)
self.assertTrue((p == 0.3333333).all())
self.assertEqual(v, 0.0)
def test_uniform_get_valid_dist(self):
gi = TwoPlayerGuessIt()
nn = NeuralNetwork(gi, DumbNet)
init = gi.get_initial_state()
logits = torch.ones((2, 2))
self.assertEqual(list(nn.get_valid_dist(init, logits)),
[.25, .25, .25, .25])
self.assertEqual(
list(nn.get_valid_dist(init, logits, log_softmax=True)),
[-1.38629436] * 4)
template = np.zeros_like(gi.get_available_actions(init))
template[0, 1] = 1
s = gi.take_action(init, template)
self.assertEqual(list(nn.get_valid_dist(s, logits)), [0.3333333] * 3)
self.assertEqual(list(nn.get_valid_dist(s, logits, log_softmax=True)),
[-1.09861228] * 3)
def test_prior_get_valid_dist(self):
gi = TwoPlayerGuessIt()
nn = NeuralNetwork(gi, DumbNet)
init = gi.get_initial_state()
logits = torch.ones((2, 2))
logits[:, -1] = 2
dist = [float(x) for x in list(nn.get_valid_dist(init, logits))]
self.assertEqual(dist, [
0.13447073101997375, 0.3655293583869934, 0.13447073101997375,
0.3655293583869934
])
template = np.zeros_like(gi.get_available_actions(init))
template[0, 1] = 1
s = gi.take_action(init, template)
dist = [float(x) for x in list(nn.get_valid_dist(s, logits))]
self.assertEqual(
dist,
[0.21194154024124146, 0.21194154024124146, 0.5761168599128723])
def test_combined_loss(self):
gi = TwoPlayerGuessIt()
nn = NeuralNetwork(gi, DumbNet)
init = gi.get_initial_state()
template = np.zeros_like(gi.get_available_actions(init))
template[0, 1] = 1
s = gi.take_action(init, template)
states = np.array([init, s])
p_gt_top = np.ones(4) * .1
p_gt_top[-1] = .7
p_gt_bot = np.ones(3) * .2
p_gt_bot[-1] = .6
p_gt = np.array([p_gt_top, p_gt_bot], dtype=np.object)
p_pred = np.ones((2, 2, 2), dtype=np.float32)
p_pred[:, 1, 1] = 2
p_pred = torch.from_numpy(p_pred)
v_gt, v_pred = torch.Tensor([1., .5]), torch.Tensor([-.35, -.2])
l = nn.loss(states, (p_pred, v_pred),
(p_gt, v_gt)).detach().numpy().reshape(-1)[0]
self.assertAlmostEqual((.49 + .9514) + (1.8225 + 1.0437), l, places=4)
def test_masked_loss(self):
gi = TwoPlayerGuessIt()
nn = NeuralNetwork(gi, DumbNet)
s = gi.get_initial_state()
template = np.zeros_like(gi.get_available_actions(s))
template[0, 1] = 1
s = gi.take_action(s, template)
states = np.array([s])
p_gt = np.ones(3) * .2
p_gt[-1] = .6
p_gt = np.array([p_gt])
p_pred = np.ones((1, 2, 2), dtype=np.float32)
p_pred[0, 1, 1] = 2
p_pred = torch.from_numpy(p_pred)
v_gt, v_pred = torch.Tensor([.5]), torch.Tensor([-.2])
l = nn.loss(states, (p_pred, v_pred),
(p_gt, v_gt)).detach().numpy().reshape(-1)[0]
mse_loss = .49
ce_loss = .9514
self.assertAlmostEqual(mse_loss + ce_loss, l, places=4)