data[:-1])
out = model.log_likelihood(data,
momentum_vecs=momentum_vecs,
interaction_vecs=interaction_vecs)
print(out)
##################### training ############################
num_iters = 10
losses, opt = model.fit(data,
num_iters=num_iters,
lr=0.001,
momentum_vecs=momentum_vecs,
interaction_vecs=interaction_vecs)
##################### sampling ############################
print("start sampling")
sample_z, sample_x = model.sample(30)
#################### inference ###########################
print("inferiring most likely states...")
z = model.most_likely_states(data,
momentum_vecs=momentum_vecs,
interaction_vecs=interaction_vecs)
#print("k step prediction")
#x_predict = k_step_prediction_for_coupled_momentum_model(model, z, data, momentum_vecs=momentum_vecs, features=features)
#x_predict = k_step_prediction(model, z, data, 10)
# TODO: need to revise the k-step prediction, specifically the way to calculate the momentum
import matplotlib.pyplot as plt torch.manual_seed(0) np.random.seed(0) # test fitting K = 3 D = 2 lags= 5 trans1 = LinearTransformation(K=K, D=D, lags=lags) obs1 = ARGaussianObservation(K=K, D=D, transformation=trans1) model1 = HMM(K=K,D=D, observation=obs1) T = 100 sample_z, sample_x = model1.sample(T) model2 = HMM(K=K, D=D, observation='gaussian', observation_kwargs=dict(lags=lags)) lls, opt = model2.fit(sample_x, num_iters=2000, lr=0.001) z_infer = model2.most_likely_states(sample_x) x_predict = k_step_prediction(model2, z_infer, sample_x) plt.figure() plt.plot(x_predict[:,0], label='prediction') plt.plot(sample_x[:,0], label='truth') plt.show()
data = torch.randn(T, D, dtype=torch.float64)
lags = 1
bounds = np.array([[-2, 2], [0, 1], [-2, 2], [0, 1]])
As = np.array([
np.column_stack([np.identity(D),
np.zeros((D, (lags - 1) * D))]) for _ in range(K)
])
torch.manual_seed(0)
np.random.seed(0)
tran = LinearTransformation(K=K, D=D, lags=lags, As=As)
observation = ARTruncatedNormalObservation(K=K,
D=D,
M=0,
transformation=tran,
bounds=bounds)
model = HMM(K=K, D=D, M=0, observation=observation)
lls = model.log_likelihood(data)
print(lls)
#losses_1, optimizer_1 = model_1.fit(data_1, method='adam', num_iters=2000, lr=0.001)
z_1 = model.most_likely_states(data)
x_predict_arr_lag1 = k_step_prediction(model, z_1, data)
momentum_vecs = MomentumTransformation._compute_momentum_vecs(
data[:-1], lags=momentum_lags)
out = model.log_likelihood(data, momentum_vecs=momentum_vecs)
print(out)
##################### training ############################
num_iters = 10
losses, opt = model.fit(data,
num_iters=num_iters,
lr=0.001,
momentum_vecs=momentum_vecs)
##################### sampling ############################
print("start sampling")
sample_z, sample_x = model.sample(30)
#################### inference ###########################
print("inferiring most likely states...")
z = model.most_likely_states(data, momentum_vecs=momentum_vecs)
print("k step prediction")
x_predict = k_step_prediction_for_momentum_model(model,
z,
data,
momentum_vecs=momentum_vecs)
#x_predict = k_step_prediction(model, z, data, 10)
# TODO: need to revise the k-step prediction, specifically the way to calculate the momentum
def test_model():
torch.manual_seed(0)
np.random.seed(0)
T = 100
D = 4
# data = np.array([[1.0, 1.0, 1.0, 6.0], [3.0, 6.0, 8.0, 6.0],
# [4.0, 7.0, 8.0, 5.0], [6.0, 7.0, 5.0, 6.0], [8.0, 2.0, 6.0, 1.0]])
data = np.random.randn(T, D)
data = torch.tensor(data, dtype=torch.float64)
xmax = max(np.max(data[:, 0].numpy()), np.max(data[:, 2].numpy()))
xmin = min(np.min(data[:, 0].numpy()), np.min(data[:, 2].numpy()))
ymax = max(np.max(data[:, 1].numpy()), np.max(data[:, 3].numpy()))
ymin = min(np.min(data[:, 1].numpy()), np.min(data[:, 3].numpy()))
bounds = np.array([[xmin - 1, xmax + 1], [ymin - 1, ymax + 1],
[xmin - 1, xmax + 1], [ymin - 1, ymax + 1]])
def toy_feature_vec_func(s):
"""
:param s: self, (T, 2)
:param o: other, (T, 2)
:return: features, (T, Df, 2)
"""
corners = torch.tensor([[0, 0], [0, 8], [10, 0], [10, 8]],
dtype=torch.float64)
return feature_direction_vec(s, corners)
K = 3
Df = 4
lags = 1
tran = UniLSTMTransformation(K=K,
D=D,
Df=Df,
feature_vec_func=toy_feature_vec_func,
lags=lags,
dh=10)
# observation
obs = ARTruncatedNormalObservation(K=K,
D=D,
lags=lags,
bounds=bounds,
transformation=tran)
# model
model = HMM(K=K, D=D, observation=obs)
print("calculating log likelihood")
feature_vecs_a = toy_feature_vec_func(data[:-1, 0:2])
feature_vecs_b = toy_feature_vec_func(data[:-1, 2:4])
feature_vecs = (feature_vecs_a, feature_vecs_b)
packed_data = get_packed_data((data[:-1]), lags=lags)
model.log_likelihood(data,
feature_vecs=feature_vecs,
packed_data=packed_data)
# fit
losses, _ = model.fit(data,
optimizer=None,
method="adam",
num_iters=50,
feature_vecs=feature_vecs,
packed_data=packed_data)
plt.figure()
plt.plot(losses)
plt.show()
# most-likely-z
print("Most likely z...")
z = model.most_likely_states(data,
feature_vecs=feature_vecs,
packed_data=packed_data)
# prediction
if data.shape[0] <= 1000:
data_to_predict = data
else:
data_to_predict = data[-1000:]
print("0 step prediction")
if data.shape[0] <= 1000:
data_to_predict = data
else:
data_to_predict = data[-1000:]
x_predict = k_step_prediction_for_lstm_model(model,
z,
data_to_predict,
feature_vecs=feature_vecs)
x_predict_err = np.mean(np.abs(x_predict - data_to_predict.numpy()),
axis=0)
print("10 step prediction")
x_predict_2 = k_step_prediction(model, z, data_to_predict, k=10)
x_predict_2_err = np.mean(np.abs(x_predict_2 -
data_to_predict[10:].numpy()),
axis=0)
# samples
print("sampling...")
sample_T = 5
sample_z, sample_x = model.sample(sample_T)
def test_model():
torch.manual_seed(0)
np.random.seed(0)
T = 5
x_grids = np.array([0.0, 5.0, 10.0])
y_grids = np.array([0.0, 4.0, 8.0])
data = np.array([[1.0, 1.0, 1.0, 6.0], [3.0, 6.0, 8.0, 6.0],
[4.0, 7.0, 8.0, 5.0], [6.0, 7.0, 5.0, 6.0],
[8.0, 2.0, 6.0, 1.0]])
data = torch.tensor(data, dtype=torch.float64)
def toy_feature_vec_func(s):
"""
:param s: self, (T, 2)
:param o: other, (T, 2)
:return: features, (T, Df, 2)
"""
corners = torch.tensor([[0, 0], [0, 8], [10, 0], [10, 8]],
dtype=torch.float64)
return feature_direction_vec(s, corners)
K = 3
D = 4
M = 0
Df = 4
bounds = np.array([[0.0, 10.0], [0.0, 8.0], [0.0, 10.0], [0.0, 8.0]])
tran = LinearGridTransformation(K=K,
D=D,
x_grids=x_grids,
y_grids=y_grids,
Df=Df,
feature_vec_func=toy_feature_vec_func)
obs = ARTruncatedNormalObservation(K=K,
D=D,
M=0,
lags=1,
bounds=bounds,
transformation=tran)
model = HMM(K=K, D=D, M=M, transition="stationary", observation=obs)
model.observation.mus_init = data[0] * torch.ones(
K, D, dtype=torch.float64)
# calculate memory
gridpoints_idx_a = tran.get_gridpoints_idx_for_batch(data[:-1, 0:2])
gridpoints_idx_b = tran.get_gridpoints_idx_for_batch(data[:-1, 2:4])
gridpoints_a = tran.get_gridpoints_for_batch(gridpoints_idx_a)
gridpoints_b = tran.get_gridpoints_for_batch(gridpoints_idx_b)
feature_vecs_a = toy_feature_vec_func(data[:-1, 0:2])
feature_vecs_b = toy_feature_vec_func(data[:-1, 2:4])
gridpoints_idx = (gridpoints_idx_a, gridpoints_idx_b)
gridpoints = (gridpoints_a, gridpoints_b)
feature_vecs = (feature_vecs_a, feature_vecs_b)
# fit
losses, opt = model.fit(data,
optimizer=None,
method='adam',
num_iters=100,
lr=0.01,
pbar_update_interval=10,
gridpoints=gridpoints,
gridpoints_idx=gridpoints_idx,
feature_vecs=feature_vecs)
plt.figure()
plt.plot(losses)
plt.show()
# most-likely-z
print("Most likely z...")
z = model.most_likely_states(data,
gridpoints_idx=gridpoints_idx,
feature_vecs=feature_vecs)
# prediction
print("0 step prediction")
if data.shape[0] <= 1000:
data_to_predict = data
else:
data_to_predict = data[-1000:]
x_predict = k_step_prediction_for_lineargrid_model(
model,
z,
data_to_predict,
gridpoints_idx=gridpoints_idx,
feature_vecs=feature_vecs)
x_predict_err = np.mean(np.abs(x_predict - data_to_predict.numpy()),
axis=0)
print("2 step prediction")
x_predict_2 = k_step_prediction(model, z, data_to_predict, k=2)
x_predict_2_err = np.mean(np.abs(x_predict_2 -
data_to_predict[2:].numpy()),
axis=0)
# samples
sample_T = 5
sample_z, sample_x = model.sample(sample_T)
def test_model():
torch.manual_seed(0)
np.random.seed(0)
T = 5
x_grids = np.array([0.0, 10.0])
y_grids = np.array([0.0, 8.0])
bounds = np.array([[0.0, 10.0], [0.0, 8.0], [0.0, 10.0], [0.0, 8.0]])
data = np.array([[1.0, 1.0, 1.0, 6.0], [3.0, 6.0, 8.0, 6.0],
[4.0, 7.0, 8.0, 5.0], [6.0, 7.0, 5.0, 6.0],
[8.0, 2.0, 6.0, 1.0]])
data = torch.tensor(data, dtype=torch.float64)
K = 3
D = 4
M = 0
obs = GPObservation(K=K,
D=D,
x_grids=x_grids,
y_grids=y_grids,
bounds=bounds,
train_rs=True)
correct_kerneldist_gg = torch.tensor(
[[0., 0., 64., 64., 100., 100., 164., 164.],
[0., 0., 64., 64., 100., 100., 164., 164.],
[64., 64., 0., 0., 164., 164., 100., 100.],
[64., 64., 0., 0., 164., 164., 100., 100.],
[100., 100., 164., 164., 0., 0., 64., 64.],
[100., 100., 164., 164., 0., 0., 64., 64.],
[164., 164., 100., 100., 64., 64., 0., 0.],
[164., 164., 100., 100., 64., 64., 0., 0.]],
dtype=torch.float64)
assert torch.all(torch.eq(correct_kerneldist_gg,
obs.kernel_distsq_gg)), obs.kernel_distsq_gg
log_prob_nocache = obs.log_prob(data)
print("log_prob_nocache = {}".format(log_prob_nocache))
kernel_distsq_xg_a = kernel_distsq_doubled(data[:-1, 0:2],
obs.inducing_points)
kernel_distsq_xg_b = kernel_distsq_doubled(data[:-1, 2:4],
obs.inducing_points)
correct_kernel_distsq_xg_a = torch.tensor(
[[2., 2., 50., 50., 82., 82., 130., 130.],
[2., 2., 50., 50., 82., 82., 130., 130.],
[45., 45., 13., 13., 85., 85., 53., 53.],
[45., 45., 13., 13., 85., 85., 53., 53.],
[65., 65., 17., 17., 85., 85., 37., 37.],
[65., 65., 17., 17., 85., 85., 37., 37.],
[85., 85., 37., 37., 65., 65., 17., 17.],
[85., 85., 37., 37., 65., 65., 17., 17.]],
dtype=torch.float64)
assert torch.all(torch.eq(correct_kernel_distsq_xg_a,
kernel_distsq_xg_a)), kernel_distsq_xg_a
memory_kwargs = dict(kernel_distsq_xg_a=kernel_distsq_xg_a,
kernel_distsq_xg_b=kernel_distsq_xg_b)
log_prob = obs.log_prob(data, **memory_kwargs)
print("log_prob = {}".format(log_prob))
assert torch.all(torch.eq(log_prob_nocache, log_prob))
Sigma_a, A_a = obs.get_gp_cache(data[:-1, 0:2], 0, **memory_kwargs)
Sigma_b, A_b = obs.get_gp_cache(data[:-1, 2:4], 1, **memory_kwargs)
memory_kwargs_2 = dict(Sigma_a=Sigma_a, A_a=A_a, Sigma_b=Sigma_b, A_b=A_b)
print("calculating log prob 2...")
log_prob2 = obs.log_prob(data, **memory_kwargs_2)
assert torch.all(torch.eq(log_prob, log_prob2))
model = HMM(K=K, D=D, M=M, transition="stationary", observation=obs)
model.observation.mus_init = data[0] * torch.ones(
K, D, dtype=torch.float64)
# fit
losses, opt = model.fit(data,
optimizer=None,
method='adam',
num_iters=100,
lr=0.01,
pbar_update_interval=10,
**memory_kwargs)
plt.figure()
plt.plot(losses)
plt.show()
# most-likely-z
print("Most likely z...")
z = model.most_likely_states(data, **memory_kwargs)
# prediction
print("0 step prediction")
if data.shape[0] <= 1000:
data_to_predict = data
else:
data_to_predict = data[-1000:]
x_predict = k_step_prediction_for_gpmodel(model, z, data_to_predict,
**memory_kwargs)
x_predict_err = np.mean(np.abs(x_predict - data_to_predict.numpy()),
axis=0)
print("2 step prediction")
x_predict_2 = k_step_prediction(model, z, data_to_predict, k=2)
x_predict_2_err = np.mean(np.abs(x_predict_2 -
data_to_predict[2:].numpy()),
axis=0)
# samples
sample_T = 5
sample_z, sample_x = model.sample(sample_T)
obs = ARTruncatedNormalObservation(K=K, D=D, M=M, lags=momentum_lags, bounds=bounds, transformation=tran)
# model
model = HMM(K=K, D=D, M=M, observation=obs)
log_prob = model.log_likelihood(data, masks=(masks_a, masks_b),
memory_kwargs_a=m_kwargs_a, memory_kwargs_b=m_kwargs_b)
log_prob_2 = model.log_likelihood(data)
assert torch.eq(log_prob, log_prob_2)
print(log_prob)
# training
print("start training...")
num_iters = 10
losses, opt = model.fit(data, num_iters=num_iters, lr=0.001, masks=(masks_a, masks_b),
memory_kwargs_a=m_kwargs_a, memory_kwargs_b=m_kwargs_b)
# sampling
print("start sampling")
sample_z, sample_x = model.sample(T)
# inference
print("inferiring most likely states...")
z = model.most_likely_states(data, masks=(masks_a, masks_b),
memory_kwargs_a=m_kwargs_a, memory_kwargs_b=m_kwargs_b)
print("0 step prediction")
x_predict = k_step_prediction_for_grid_model(model, z, data, memory_kwargs_a=m_kwargs_a, memory_kwargs_b=m_kwargs_b)
print("k step prediction")
x_predict_10 = k_step_prediction(model, z, data, 10)
out = model.log_likelihood(data,
momentum_vecs=momentum_vecs,
features=features)
print(out)
##################### training ############################
num_iters = 10
losses, opt = model.fit(data,
num_iters=num_iters,
lr=0.001,
momentum_vecs=momentum_vecs,
features=features)
##################### sampling ############################
print("start sampling")
sample_z, sample_x = model.sample(30)
#################### inference ###########################
print("inferiring most likely states...")
z = model.most_likely_states(data,
momentum_vecs=momentum_vecs,
features=features)
print("k step prediction")
x_predict = k_step_prediction_for_momentum_feature_model(
model, z, data, momentum_vecs=momentum_vecs, features=features)
#x_predict = k_step_prediction(model, z, data, 10)
# TODO: need to revise the k-step prediction, specifically the way to calculate the momentum
optimizer = torch.optim.Adam(model.params, lr=0.001)
losses = []
for i in np.arange(num_iters):
optimizer.zero_grad()
loss = model.loss(data)
loss.backward(retain_graph=True)
optimizer.step()
loss = loss.detach().numpy()
losses.append(loss)
if i % 10 == 0:
pbar.set_description('iter {} loss {:.2f}'.format(i, loss))
pbar.update(10)
# check reconstruction
x_reconstruct = model.sample_condition_on_zs(z, data[0])
# infer the latent states
infer_z = model.most_likely_states(data)
perm = find_permutation(z.numpy(), infer_z, K1=K, K2=K)
model.permute(perm)
hmm_z = model.most_likely_states(data)
# check prediction
x_predict_cond_z = k_step_prediction(model, z, data)
