def create_job(kwargs):
# model arguments
nb_states = kwargs.pop('nb_states')
trans_type = kwargs.pop('trans_type')
obs_prior = kwargs.pop('obs_prior')
# em arguments
obs = kwargs.pop('obs')
act = kwargs.pop('act')
prec = kwargs.pop('prec')
nb_iter = kwargs.pop('nb_iter')
train_obs, train_act = zip(
*random.sample(list(zip(obs, act)), int(0.8 * len(obs))))
dm_obs = train_obs[0].shape[-1]
dm_act = train_act[0].shape[-1]
rarhmm = rARHMM(nb_states,
dm_obs,
dm_act,
trans_type=trans_type,
obs_prior=obs_prior,
trans_kwargs=trans_kwargs)
rarhmm.initialize(train_obs, train_act)
lls = rarhmm.em(train_obs,
train_act,
nb_iter=nb_iter,
prec=prec,
verbose=False)
return rarhmm, lls[-1]
import matplotlib.pyplot as plt
from hips.plotting.colormaps import gradient_cmap
import seaborn as sns
sns.set_style("white")
sns.set_context("talk")
color_names = [
"windows blue", "red", "amber", "faded green", "dusty purple", "orange"
]
colors = sns.xkcd_palette(color_names)
cmap = gradient_cmap(colors)
true_rarhmm = rARHMM(nb_states=3, dm_obs=2, trans_type='recurrent')
# trajectory lengths
T = [1250, 1150, 1025]
true_z, x = true_rarhmm.sample(horizon=T)
true_ll = true_rarhmm.log_probability(x)
obs_prior = {'mu0': 0., 'sigma0': 1e12, 'nu0': 2, 'psi0': 1.}
trans_kwargs = {'degree': 3}
rarhmm = rARHMM(nb_states=3,
dm_obs=2,
trans_type='poly',
obs_prior=obs_prior,
trans_kwargs=trans_kwargs)
rarhmm.initialize(x)
'mean': np.array([0., 0., 0.]),
'std': np.array([np.pi, 8., 2.5])
}
}
trans_mstep_kwargs = {'nb_iter': 100} # 'batch_size': 1024, 'lr': 1e-3}
# trans_type = 'poly'
# trans_kwargs = {'degree': 1,
# 'norm': {'mean': np.array([0., 0., 0.]),
# 'std': np.array([np.pi, 8., 2.5])}}
# trans_mstep_kwargs = {'nb_iter': 100, 'batch_size': 1024, 'lr': 1e-3}
rarhmm = rARHMM(nb_states,
dm_obs,
dm_act,
trans_type=trans_type,
obs_prior=obs_prior,
trans_prior=trans_prior,
trans_kwargs=trans_kwargs)
# rarhmm.initialize(obs, act)
rarhmm.stochastic_em(obs,
act,
nb_epochs=250,
verbose=True,
method='adam',
step_size=5e-4)
lls = rarhmm.em(obs,
act,
nb_iter=10,
import matplotlib.pyplot as plt
from hips.plotting.colormaps import gradient_cmap
import seaborn as sns
sns.set_style("white")
sns.set_context("talk")
color_names = ["windows blue", "red", "amber", "faded green", "dusty purple", "orange"]
colors = sns.xkcd_palette(color_names)
cmap = gradient_cmap(colors)
true_rarhmm = rARHMM(nb_states=3, dm_obs=2, trans_type='neural')
# trajectory lengths
T = [1250, 1150, 1025]
true_z, x = true_rarhmm.sample(horizon=T)
true_ll = true_rarhmm.log_norm(x)
rarhmm = rARHMM(nb_states=3, dm_obs=2, trans_type='neural')
rarhmm.initialize(x)
lls = rarhmm.em(x, nb_iter=100, prec=0., verbose=True)
print("true_ll=", true_ll, "hmm_ll=", lls[-1])
plt.figure(figsize=(5, 5))
plt.plot(np.ones(len(lls)) * true_ll, '-r')