def test_stick_transition():
K = 3
D = 2
M = 0
model = HMM(K=K, D=D, M=M, transition='sticky', observation='gaussian', transition_kwargs=dict(alpha=1, kappa=100))
print("alpha = {}, kappa = {}".format(model.transition.alpha, model.transition.kappa))
data = torch.tensor([[2,3], [4,5], [6, 7]], dtype=torch.float64)
log_prob = model.log_likelihood(data)
print("log_prob", log_prob)
samples_z, samples_x = model.sample(10)
samples_log_prob = model.log_probability(samples_x)
print("sample log_prob", samples_log_prob)
print("model transition\n", model.transition.stationary_transition_matrix)
model1 = HMM(K=K, D=D, M=M, transition='sticky', observation='gaussian', transition_kwargs=dict(alpha=1, kappa=0.1))
print("Before fit")
print("model1 transition\n", model1.transition.stationary_transition_matrix)
losses, opt = model1.fit(samples_x)
print("After fit")
print("model1 transition\n", model1.transition.stationary_transition_matrix)
model2 = HMM(K=K, D=D, M=M, transition='sticky', observation='gaussian', transition_kwargs=dict(alpha=1, kappa=20))
print("Before fit")
print("model2 transition\n", model2.transition.stationary_transition_matrix)
losses, opt = model2.fit(samples_x)
print("After fit")
print("model2 transition\n", model2.transition.stationary_transition_matrix)
model3 = HMM(K=K, D=D, M=M, transition='sticky', observation='gaussian', transition_kwargs=dict(alpha=1, kappa=100))
print("Before fit")
print("model3 transition\n", model3.transition.stationary_transition_matrix)
losses, opt = model3.fit(samples_x)
print("After fit")
print("model3 transition\n", model3.transition.stationary_transition_matrix)
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
from hips.plotting.colormaps import gradient_cmap, white_to_color_cmap
color_names = [
"windows blue", "red", "amber", "faded green", "dusty purple", "orange"
]
colors = sns.xkcd_palette(color_names)
cmap = gradient_cmap(colors)
npr.seed(0)
torch.manual_seed(0)
K = 3
D = 2
M = 1
T = 100
# Create an exogenous input
inpt = np.sin(2 * np.pi * np.arange(T) / 50)[:, None] + 1e-1 * npr.randn(T, M)
inpt = torch.tensor(inpt, dtype=torch.float64)
true_model = HMM(K=K,
D=D,
M=M,
transition='inputdriven',
observation='gaussian')
z, data = true_model.sample(T, input=inpt, return_np=True)
lls = true_model.log_likelihood(data, inpt)
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()
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)
pbar.update(10)
x_reconstruct = model.sample_condition_on_zs(z, data[0])
"""
# test momentum_lags
true_observation = ARGaussianObservation(K=K,
D=D,
M=0,
lags=5,
transformation='linear')
true_model = HMM(K=K, D=D, M=0, observation=true_observation)
z, data = true_model.sample(T, return_np=True)
# fit to a model
observation = ARGaussianObservation(K=K,
D=D,
M=0,
lags=5,
transformation='linear')
model = HMM(K=K, D=D, M=0, observation=observation)
num_iters = 10000
pbar = tqdm(total=num_iters, file=sys.stdout)
optimizer = torch.optim.Adam(model.params, lr=0.0001)
bounds = np.array([[0, 20], [0, 40]])
thetas = np.linspace(0, 2 * np.pi, K, endpoint=False)
mus_init = 3 * np.column_stack((np.cos(thetas), np.sin(thetas))) + 5
true_tran = LinearTransformation(K=K, D=D, lags=lags)
true_observation = ARTruncatedNormalObservation(K=K,
D=D,
M=0,
lags=lags,
transformation=true_tran,
bounds=bounds,
mus_init=mus_init,
train_sigma=False)
true_model = HMM(K=K, D=D, M=0, observation=true_observation)
z, x = true_model.sample(T, return_np=False)
true_ll = true_model.log_likelihood(x)
print(true_ll)
print("\n # model parameters: \n", len(true_model.params))
print("\n # model trainable parameters: \n", len(true_model.trainable_params))
sample_z, sample_x = true_model.sample(T)
"""
# learning
tran = LinearTransformation(K=K, D=D, momentum_lags=1, As=As)
observation = ARTruncatedNormalObservation(K=K, D=D, M=0, momentum_lags=1, with_noise=tran, bounds=bounds, mus_init=mus_init)
model = HMM(K=K, D=D, M=0, observation=true_observation)
# model
model = HMM(K=K, D=D, M=M, observation=obs)
# log like
log_prob = model.log_probability(data)
print("log probability = ", log_prob)
# training
print("start training...")
num_iters = 10
losses, opt = model.fit(data, num_iters=num_iters, lr=0.001)
# sampling
samplt_T = T
print("start sampling")
sample_z, sample_x = model.sample(samplt_T)
print("start sampling based on with_noise")
sample_z2, sample_x2 = model.sample(T, with_noise=True)
# inference
print("inferiring most likely states...")
z = model.most_likely_states(data)
print("0 step prediction")
x_predict = k_step_prediction(model, z, data)
print("k step prediction")
x_predict_10 = k_step_prediction(model, z, data, 2)
D = 2
T = 100
As = [random_rotation(D) for _ in range(K)]
true_tran = LinearTransformation(K=K, d_in=D, D=D, As=As)
bounds = np.array([[0, 20], [-5, 25]])
true_observation = ARLogitNormalObservation(K=K,
D=D,
M=0,
transformation=true_tran,
bounds=bounds)
true_model = HMM(K=K, D=D, M=0, observation=true_observation)
z, data = true_model.sample(T, return_np=False)
# Define a model to fit the data
tran = LinearTransformation(K=K, d_in=D, D=D)
observation = ARLogitNormalObservation(K=K,
D=D,
M=0,
transformation=tran,
bounds=bounds)
model = HMM(K=K, D=D, M=0, observation=observation)
# Model fitting
num_iters = 10 #9000
M=M,
lags=1,
bounds=bounds,
transformation=tran)
# model
model = HMM(K=K, D=D, M=M, observation=obs)
#model.observation.mus_init = data[0] * torch.ones(K, D, dtype=torch.float64)
params = joblib.load(
"/Users/leah/Columbia/courses/19summer/SocialBehavior/SocialBehaviorptc/project_notebooks/gridmodel/model_k2"
)
model.params = params
obs.log_sigmas = torch.log(torch.ones((2, 4), dtype=torch.float64) * 0.01)
center_z = torch.tensor([0], dtype=torch.int)
center_x = torch.tensor([[150, 190, 200, 200]], dtype=torch.float64)
sample_z_center, sample_x_center = model.sample(20,
prefix=(center_z, center_x))
diff = np.diff(sample_x_center, axis=0)
print("diff in dim 0")
print(diff[:, 0])
print("\ndiff in dim 1")
print(diff[:, 1])
