def _time_step(time, output_ta_t, states):
"""Take a time step of the dynamic RNN.
Args:
time: int32 scalar Tensor.
output_ta_t: List of `TensorArray`s that represent the output.
state: nested tuple of vector tensors that represent the state.
Returns:
The tuple (time + 1, output_ta_t with updated flow, new_state).
"""
if in_graph_mode:
input_t = tuple(ta.read(time) for ta in input_ta)
# Restore some shape information
for input_, shape in zip(input_t, inputs_got_shape):
input_.set_shape(shape[1:])
else:
input_t = tuple(ta[time.numpy()] for ta in input_ta)
input_t = nest.pack_sequence_as(structure=inputs,
flat_sequence=input_t)
row_time = time % width
states_1 = control_flow_ops.cond(math_ops.less(time, width),
lambda: states.read(time_steps),
lambda: states.read(time - width))
states_2 = control_flow_ops.cond(math_ops.equal(row_time, 0),
lambda: states.read(time_steps),
lambda: states.read(time - 1))
call_cell = lambda: cell(input_t, tuple((states_1, states_2)))
if sequence_length is not None:
(output,
new_state) = _rnn_step(time=row_time,
sequence_length=sequence_length,
min_sequence_length=min_sequence_length,
max_sequence_length=max_sequence_length,
zero_output=zero_output,
state=state,
call_cell=call_cell,
state_size=state_size,
skip_conditionals=True)
else:
(output, new_state) = call_cell()
# Pack state if using state tuples
output = nest.flatten(output)
if in_graph_mode:
output_ta_t = tuple(
ta.write(time, out) for ta, out in zip(output_ta_t, output))
states = states.write(time, new_state)
else:
for ta, out in zip(output_ta_t, output):
ta[time.numpy()] = out
return (time + 1, output_ta_t, states)
def visualize(time, true_y, pred_y, odefunc, itr, loss):
if args.viz:
ax_traj.cla()
ax_traj.set_title('Trajectories')
ax_traj.set_xlabel('t')
ax_traj.set_ylabel('x')
ax_traj.plot(time.numpy(), pred_y.numpy()[:, 0, 0], 'r--', time.numpy(), true_y.numpy()[:, 0, 0], 'k-')
ax_traj.set_xlim(time.min(), time.max())
ax_traj.plot(time.numpy(), pred_y.numpy()[:, 0, 1], 'r--', time.numpy(), true_y.numpy()[:, 0, 1], 'k-')
ax_traj.legend()
axy_phase.cla()
axy_phase.set_title('Phase Portrait')
axy_phase.set_xlabel('x')
axy_phase.set_ylabel('y')
axy_phase.plot(true_y.numpy()[:, 0, 0], true_y.numpy()[:, 0, 1], 'r--')
axy_phase.plot(pred_y.numpy()[:, 0, 0], pred_y.numpy()[:, 0, 1], 'k-')
if args.ode_nums == 3:
ay_traj.cla()
ay_traj.set_title('Trajectories')
ay_traj.set_xlabel('t')
ay_traj.set_ylabel('y')
ay_traj.plot(time.numpy(), pred_y.numpy()[:, 0, 1], 'r--', time.numpy(), true_y.numpy()[:, 0, 1], 'k-')
ay_traj.set_xlim(time.min(), time.max())
# ax_traj.set_ylim(-2, 2)
ay_traj.legend()
az_traj.cla()
az_traj.set_title('Trajectories')
az_traj.set_xlabel('t')
az_traj.set_ylabel('z')
az_traj.plot(time.numpy(), pred_y.numpy()[:, 0, 2], 'r--', time.numpy(), true_y.numpy()[:, 0, 2], 'k-')
az_traj.set_xlim(time.min(), time.max())
# ax_traj.set_ylim(-2, 2)
az_traj.legend()
axz_phase.cla()
axz_phase.set_title('Phase Portrait')
axz_phase.set_xlabel('x')
axz_phase.set_ylabel('z')
axz_phase.plot(true_y.numpy()[:, 0, 0], true_y.numpy()[:, 0, 2], 'r--')
axz_phase.plot(pred_y.numpy()[:, 0, 0], pred_y.numpy()[:, 0, 2], 'k-')
ayz_phase.cla()
ayz_phase.set_title('Phase Portrait')
ayz_phase.set_xlabel('y')
ayz_phase.set_ylabel('z')
ayz_phase.plot(true_y.numpy()[:, 0, 1], true_y.numpy()[:, 0, 2], 'r--')
ayz_phase.plot(pred_y.numpy()[:, 0, 1], pred_y.numpy()[:, 0, 2], 'k-')
fig.tight_layout()
plt.savefig('png{}/{}.jpg'.format(args.CASE,itr))
def ILP_R_stat(self, head, relation, tail, time, Type):
head_embedding = self.Transmit.entity_embeddings(torch.tensor([head]))
relation_embedding = self.Transmit.relation_embeddings(
torch.tensor([relation]))
tail_embedding = self.Transmit.entity_embeddings(torch.tensor([tail]))
target_loss = torch.norm(
head_embedding + relation_embedding - tail_embedding,
self.norm).repeat(self.numOfRelation, 1)
C_1 = self.overlap_constrain(head, relation, tail, time.numpy())
C_2_pre = self.order_constrain(head, relation, tail, time.numpy(),
"pre")
C_2_aft = self.order_constrain(head, relation, tail, time.numpy(),
"aft")
if C_1 * C_2_pre * C_2_aft == 0 or (float(target_loss[0])) > 11:
target_loss = torch.FloatTensor([1000
]).repeat(self.numOfRelation, 1)
head_embedding = head_embedding.repeat(self.numOfRelation, 1)
tail_embedding = tail_embedding.repeat(self.numOfRelation, 1)
tmpRelationLoss = torch.norm(
head_embedding + self.Transmit.relation_embeddings.weight.data -
tail_embedding, self.norm, 1).view(-1, 1)
ILP_list_r = list()
for j in range(self.numOfRelation):
C_1 = self.overlap_constrain(head, int(j), tail, time.numpy())
C_2_pre = self.order_constrain(head, int(j), tail, time.numpy(),
"pre")
C_2_aft = self.order_constrain(head, int(j), tail, time.numpy(),
"aft")
if C_1 * C_2_pre * C_2_aft == 0 or (float(
tmpRelationLoss[j])) > 11:
ILP_list_r.append([10000])
else:
ILP_list_r.append([0])
ILP_tensor_r = torch.FloatTensor(ILP_list_r)
tmpRelationLoss = tmpRelationLoss + ILP_tensor_r
wrongRelation = torch.nonzero(
nn.functional.relu(target_loss - tmpRelationLoss))
print(wrongRelation.size()[0])
for tmp in wrongRelation:
print(tmp[0])
def visualize(vt, true_y, pred_y, odefunc, itr, loss):
if args.viz:
ax_traj.cla()
ax_traj.set_title('Trajectories')
ax_traj.set_xlabel('t')
ax_traj.set_ylabel('x')
for i in range(len(indexv)):
if i == 0:
ax_traj.plot(vt[i].numpy(),
pred_y.numpy()[:indexv[i] + 1, 0, 0], 'r--',
vt[i].numpy(),
true_y.numpy()[:indexv[i] + 1, 0, 0], 'k-')
else:
ax_traj.plot(
vt[i].numpy(),
pred_y.numpy()[indexv[i - 1] + 1:indexv[i] + 1, 0,
0], 'r--', vt[i].numpy(),
true_y.numpy()[indexv[i - 1] + 1:indexv[i] + 1, 0, 0],
'k-')
ax_traj.legend()
ay_traj.cla()
ay_traj.set_title('Trajectories')
ay_traj.set_xlabel('t')
ay_traj.set_ylabel('y')
for i in range(len(indexv)):
if i == 0:
ay_traj.plot(vt[i].numpy(),
pred_y.numpy()[:indexv[i] + 1, 0, 1], 'r--',
vt[i].numpy(),
true_y.numpy()[:indexv[i] + 1, 0, 1], 'k-')
else:
ay_traj.plot(
vt[i].numpy(),
pred_y.numpy()[indexv[i - 1] + 1:indexv[i] + 1, 0,
1], 'r--', vt[i].numpy(),
true_y.numpy()[indexv[i - 1] + 1:indexv[i] + 1, 0, 1],
'k-')
ay_traj.legend()
if args.variable_nums == 3:
ay_traj.cla()
ay_traj.set_title('Trajectories')
ay_traj.set_xlabel('t')
ay_traj.set_ylabel('y')
ay_traj.plot(time.numpy(),
pred_y.numpy()[:, 0, 1], 'r--', time.numpy(),
true_y.numpy()[:, 0, 1], 'k-')
ay_traj.set_xlim(time.min(), time.max())
# ax_traj.set_ylim(-2, 2)
ay_traj.legend()
az_traj.cla()
az_traj.set_title('Trajectories')
az_traj.set_xlabel('t')
az_traj.set_ylabel('z')
az_traj.plot(time.numpy(),
pred_y.numpy()[:, 0, 2], 'r--', time.numpy(),
true_y.numpy()[:, 0, 2], 'k-')
az_traj.set_xlim(time.min(), time.max())
# ax_traj.set_ylim(-2, 2)
az_traj.legend()
axz_phase.cla()
axz_phase.set_title('Phase Portrait')
axz_phase.set_xlabel('x')
axz_phase.set_ylabel('z')
axz_phase.plot(true_y.numpy()[:, 0, 0],
true_y.numpy()[:, 0, 2], 'r--')
axz_phase.plot(pred_y.numpy()[:, 0, 0],
pred_y.numpy()[:, 0, 2], 'k-')
ayz_phase.cla()
ayz_phase.set_title('Phase Portrait')
ayz_phase.set_xlabel('y')
ayz_phase.set_ylabel('z')
ayz_phase.plot(true_y.numpy()[:, 0, 1],
true_y.numpy()[:, 0, 2], 'r--')
ayz_phase.plot(pred_y.numpy()[:, 0, 1],
pred_y.numpy()[:, 0, 2], 'k-')
fig.tight_layout()
plt.savefig('png{}/{}.jpg'.format(args.CASE, itr))
def sample(self, sim, indices=None, samples=1000):
if sim.N_real < 4:
raise AttributeError(
"SPOCK Error: SPOCK only works for systems with 3 or more planets"
)
if indices:
if len(indices) != 3:
raise AttributeError(
"SPOCK Error: indices must be a list of 3 particle indices"
)
trios = [indices] # always make it into a list of trios to test
else:
trios = [[i, i + 1, i + 2] for i in range(1, sim.N_real - 2)
] # list of adjacent trios
ic(trios)
kwargs = OrderedDict()
kwargs['Norbits'] = int(1e4)
kwargs['Nout'] = 1000
kwargs['trios'] = trios
args = list(kwargs.values())
# These are the .npy.
# In the other file, we concatenate (restseries, orbtseries, mass_array)
tseries, stable = get_extended_tseries(sim, args)
if not stable:
return None
tseries = np.array(tseries)
simt = sim.copy()
alltime = []
for i, trio in enumerate(trios):
sim = simt.copy()
# These are the .npy.
# In the other file, we concatenate (restseries, orbtseries, mass_array)
cur_tseries = tseries[None, i, ::10]
mass_array = np.array(
[sim.particles[j].m / sim.particles[0].m for j in trio])
X = data_setup_kernel(mass_array, cur_tseries)
X = self.ssX.transform(X.reshape(-1, X.shape[-1])).reshape(X.shape)
X = torch.tensor(X).float()
if self.cuda:
X = X.cuda()
time = torch.cat(
[self.sample_full_swag(X)[None] for _ in range(samples)],
dim=0).detach()
if self.cuda:
time = time.cpu()
time = time.numpy()
alltime.append(time)
out = np.array(alltime)[..., 0, :]
mu = out[..., 0]
std = out[..., 1]
#Old:
# a = (4 - mu)/std
# b = np.inf
# try:
# samples = truncnorm.rvs(a, b)*std + mu
# except ValueError:
# return None
# first_inst = np.min(samples, 0)
# return first_inst
#HACK - should this be inf at the top?
# a, b = (4 - out[..., 0]) / out[..., 1], np.inf #(12 - out[..., 0]) / out[..., 1]
# try:
# samples = truncnorm.rvs(a, b, loc=out[..., 0], scale=out[..., 1])
# except ValueError:
# return None
# return np.min(samples, 0)
return mu, std