def test_type(self):
self.assertIsInstance(
math.lognormexp(torch.rand(1)),
torch.Tensor
)
self.assertIsInstance(
math.lognormexp(np.array([2])),
np.ndarray
)
def test_value(self):
test_input = [1, 2, 3]
temp = np.exp(1) + np.exp(2) + np.exp(3)
test_result = np.log(np.exp(test_input) / temp)
np.testing.assert_allclose(
math.lognormexp(torch.Tensor(test_input)).numpy(),
torch.Tensor(test_result).numpy(),
atol=1e-6
)
np.testing.assert_allclose(
math.lognormexp(np.array(test_input)),
np.array(test_result),
atol=1e-6
)
def encode_particles(self, latent_state):
"""
RNN that encodes the set of particles into one latent vector that can be passed to policy.
"""
batch_size, num_particles, h_dim = latent_state.h.size()
state = torch.cat([latent_state.h, latent_state.phi_z], dim=2)
# latent_state.h [batch, particles, h_dim?]
normalized_log_weights = math.lognormexp(
# inference_result.log_weights[-1],
latent_state.log_weight,
dim=1)
particle_state = torch.cat(
[state, torch.exp(normalized_log_weights).unsqueeze(-1)], dim=2)
if self.num_particles == 1:
# Get rid of particle dimension, particle_gru is just a nn.Linear
particle_state = particle_state.squeeze(1)
encoded_particles = self.particle_gru(particle_state)
# encoded_particles = self.particle_gru_bn(encoded_particles)
return encoded_particles
else:
_, encoded_particles = self.particle_gru(particle_state)
# encoded_particles [num_layers * num_directions, batch, h_dim]
# First dimension: num_layers * num_directions
# Dimension of Output?
return encoded_particles[0]
def test_dimensions(self):
self.assertEqual(
math.lognormexp(
torch.rand(2, 3, 4, 5),
dim=2
).size(),
torch.Size([2, 3, 4, 5])
)
self.assertEqual(
math.lognormexp(torch.rand(3)).size(),
torch.Size([3])
)
self.assertEqual(
math.lognormexp(torch.rand(1)).size(),
torch.Size([1])
)
self.assertEqual(
list(np.shape(math.lognormexp(
np.random.rand(2, 3, 4, 5),
dim=2
))),
[2, 3, 4, 5]
)
self.assertEqual(
list(np.shape(math.lognormexp(np.random.rand(3)))),
[3]
)
self.assertEqual(
list(np.shape(math.lognormexp(np.random.rand(1)))),
[1]
)