def make_figure(rslds, zs_rslds, x_rslds):
fig = plt.figure(figsize=(10, 10))
gs = gridspec.GridSpec(2, 2)
fp = FontProperties()
fp.set_weight("bold")
ax3 = fig.add_subplot(gs[0, 0])
plot_most_likely_dynamics(rslds.trans_distn,
rslds.dynamics_distns,
colors,
xlim=(-4, 8),
ylim=(-.5, .5),
ax=ax3)
# Overlay a partial trajectory
rplt.plot_trajectory(zs_rslds[-1], x_rslds, ax=ax3, ls="-")
ax3.set_title("Inferred Dynamics (rSLDS)")
# Plot samples of discrete state sequence
ax4 = fig.add_subplot(gs[1, 0])
rplt.plot_z_samples(rslds.num_states,
zs_rslds,
plt_slice=(0, x_rslds.shape[0]),
ax=ax4)
ax4.set_title("Discrete State Samples")
ax5 = fig.add_subplot(gs[0, 1])
plot_input_dynamics(rslds.trans_distn,
rslds.dynamics_distns,
xlim=(-4, 8),
ylim=(-4, 10),
ax=ax5)
ax6 = fig.add_subplot(gs[1, 1])
plot_other_compartments_dynamics(rslds.trans_distn,
rslds.dynamics_distns,
xlim=(-4, 8),
ylim=(-4, 10),
ax=ax6)
plt.tight_layout()
def make_figure(true_model, z_true, x_true, y,
rslds, zs_rslds, x_rslds,
z_rslds_gen, x_rslds_gen,
z_slds_gen, x_slds_gen):
fig = plt.figure(figsize=(6.5,3.5))
gs = gridspec.GridSpec(2,3)
fp = FontProperties()
fp.set_weight("bold")
# True dynamics
ax1 = fig.add_subplot(gs[0,0])
rplt.plot_most_likely_dynamics(true_model.trans_distn,
true_model.dynamics_distns,
xlim=(-3,3), ylim=(-2,2),
ax=ax1)
# Overlay a partial trajectory
rplt.plot_trajectory(z_true[1:1000], x_true[1:1000], ax=ax1, ls="-")
ax1.set_title("True Latent Dynamics")
plt.figtext(.025, 1-.075, '(a)', fontproperties=fp)
# Plot a few output dimensions
ax2 = fig.add_subplot(gs[1, 0])
for n in range(args.D_obs):
rplt.plot_data(z_true[1:1000], y[1:1000, n], ax=ax2, ls="-")
ax2.set_xlabel("Time")
ax2.set_ylabel("$y$")
ax2.set_title("Observed Data")
plt.figtext(.025, .5 - .075, '(b)', fontproperties=fp)
# Plot the inferred dynamics under the rSLDS
ax3 = fig.add_subplot(gs[0, 1])
rplt.plot_most_likely_dynamics(rslds.trans_distn,
rslds.dynamics_distns,
xlim=(-3, 3), ylim=(-2, 2),
ax=ax3)
# Overlay a partial trajectory
rplt.plot_trajectory(zs_rslds[-1][1:1000], x_rslds[1:1000], ax=ax3, ls="-")
ax3.set_title("Inferred Dynamics (rSLDS)")
plt.figtext(.33 + .025, 1. - .075, '(c)', fontproperties=fp)
# Plot samples of discrete state sequence
ax4 = fig.add_subplot(gs[1,1])
rplt.plot_z_samples(args.K, zs_rslds, zref=z_true, plt_slice=(0,1000), ax=ax4)
ax4.set_title("Discrete State Samples")
plt.figtext(.33 + .025, .5 - .075, '(d)', fontproperties=fp)
# Plot simulated SLDS data
ax5 = fig.add_subplot(gs[0, 2])
rplt.plot_trajectory(z_slds_gen[-1000:], x_slds_gen[-1000:], ax=ax5, ls="-")
plt.grid(True)
ax5.set_title("Generated States (SLDS)")
plt.figtext(.66 + .025, 1. - .075, '(e)', fontproperties=fp)
# Plot simulated rSLDS data
ax6 = fig.add_subplot(gs[1, 2])
rplt.plot_trajectory(z_rslds_gen[-1000:], x_rslds_gen[-1000:], ax=ax6, ls="-")
ax6.set_title("Generated States (rSLDS)")
plt.grid(True)
plt.figtext(.66 + .025, .5 - .075, '(f)', fontproperties=fp)
plt.tight_layout()
plt.savefig(os.path.join(args.output_dir, "nascar.png"), dpi=200)
plt.savefig(os.path.join(args.output_dir, "nascar.pdf"))
plt.show()
def make_figure(true_model, z_true, x_true, y, rslds, zs_rslds, x_rslds,
z_rslds_gen, x_rslds_gen, y_rslds_gen, slds, zs_slds, x_slds,
z_slds_gen, x_slds_gen, y_slds_gen):
"""
Show the following:
- True latent dynamics (for most likely state)
- Segment of trajectory in latent space
- A few examples of observations in 10D space
- ARHMM segmentation of factors
- rSLDS segmentation of factors
- ARHMM synthesis
- rSLDS synthesis
"""
# fig = plt.figure(figsize=(6.5,3.5))
fig = plt.figure(figsize=(13, 7))
gs = gridspec.GridSpec(2, 3)
fp = FontProperties()
fp.set_weight("bold")
# True dynamics
ax1 = fig.add_subplot(gs[0, 0])
plot_most_likely_dynamics(true_model.trans_distn,
true_model.dynamics_distns,
xlim=(-3, 3),
ylim=(-2, 2),
ax=ax1)
# Overlay a partial trajectory
plot_trajectory(z_true[1:1000], x_true[1:1000], ax=ax1, ls="-")
ax1.set_title("True Latent Dynamics")
plt.figtext(.025, 1 - .075, '(a)', fontproperties=fp)
# Plot a few output dimensions
ax2 = fig.add_subplot(gs[1, 0])
for n in range(D_obs):
plot_data(z_true[1:1000], y[1:1000, n], ax=ax2, ls="-")
ax2.set_xlabel("Time")
ax2.set_ylabel("$y$")
ax2.set_title("Observed Data")
plt.figtext(.025, .5 - .075, '(b)', fontproperties=fp)
# Plot the inferred dynamics under the rSLDS
ax3 = fig.add_subplot(gs[0, 1])
ax3_lim = 1.05 * abs(x_rslds[1:1000]).max(axis=0)
plot_most_likely_dynamics(rslds.trans_distn,
rslds.dynamics_distns,
xlim=(-ax3_lim[0], ax3_lim[0]),
ylim=(-ax3_lim[1], ax3_lim[1]),
ax=ax3)
# Overlay a partial trajectory
plot_trajectory(zs_rslds[-1][1:1000], x_rslds[1:1000], ax=ax3, ls="-")
ax3.set_title("Inferred Dynamics (rSLDS)")
plt.figtext(.33 + .025, 1. - .075, '(c)', fontproperties=fp)
# Plot something... z samples?
ax4 = fig.add_subplot(gs[1, 1])
plot_z_samples(K, zs_rslds, zref=z_true, plt_slice=(0, 1000), ax=ax4)
ax4.set_title("Discrete State Samples")
plt.figtext(.33 + .025, .5 - .075, '(d)', fontproperties=fp)
# Plot simulated SLDS data
ax5 = fig.add_subplot(gs[0, 2])
# for n, ls in enumerate(["-", ":", "-."]):
# plot_data(z_slds_gen[-1000:], y_slds_gen[-1000:, n], ax=ax5, ls=ls)
plot_trajectory(z_slds_gen[-1000:], x_slds_gen[-1000:], ax=ax5, ls="-")
# ax5.set_xlabel("Time")
# ax5.set_ylabel("$y$")
plt.grid(True)
ax5.set_title("Generated States (SLDS)")
plt.figtext(.66 + .025, 1. - .075, '(e)', fontproperties=fp)
# Plot simulated rSLDS data
ax6 = fig.add_subplot(gs[1, 2])
# for n, ls in enumerate(["-", ":", "-."]):
# plot_data(z_rslds_gen[-1000:], y_rslds_gen[-1000:, n], ax=ax6, ls=ls)
# ax6.set_xlabel("Time")
# ax6.set_ylabel("$y$")
plot_trajectory(z_rslds_gen[-1000:], x_rslds_gen[-1000:], ax=ax6, ls="-")
ax6.set_title("Generated States (rSLDS)")
plt.grid(True)
plt.figtext(.66 + .025, .5 - .075, '(f)', fontproperties=fp)
plt.tight_layout()
plt.savefig(os.path.join(RESULTS_DIR, "nascar.png"), dpi=200)
plt.savefig(os.path.join(RESULTS_DIR, "nascar.pdf"))
plt.show()
plt.xlabel("Iteration")
plt.ylabel("ELBO")
# rslds, rslds_lps, rslds_z_smpls, rslds_x = \
# fit_rslds_vbem(inputs, y, mask,
# initialization="none",
# N_iters=100)
plot_trajectory_and_probs(rslds_z_smpls[-1][1:],
rslds_x[1:],
trans_distn=rslds.trans_distn,
title="Recurrent SLDS")
plot_all_dynamics(rslds.dynamics_distns)
plot_z_samples(K, rslds_z_smpls, plt_slice=(0, 1000))
## Generate from the model
rslds_y_gen, rslds_x_gen, rslds_z_gen = rslds.generate(T=T_gen,
inputs=np.ones(
(T_gen, 1)),
with_noise=True)
slds_y_gen, slds_x_gen, slds_z_gen = slds.generate(T=T_gen,
inputs=np.ones(
(T_gen, 1)))
make_figure(
true_model,
z_true,
x_true,
y,