return model
# We need to know the maximum number of latent factors that we'll need to plot.
n_latent_factors = []
for comparison_model_path in comparison_model_paths:
if not os.path.exists(comparison_model_path): continue
model = _get_model(comparison_model_path)
n_latent_factors.append(model.n_latent_factors)
# TODO: Set this from config, or something crazy to keep colours consistent?
J_max = np.max(n_latent_factors)
#J_max = 8
cmap = mpl_utils.discrete_cmap(J_max, base_cmap="Spectral_r")
colors = [cmap(j) for j in range(J_max)][::-1]
# Create comparison figure.
#A_original = original_model.theta_[original_model.parameter_names.index("A")]
D = 17
J_ast = len(config["grouped_elements"])
A_astrophysical = np.zeros((D, J_ast))#np.random.normal(0, 0.1, size=A_est.shape)
for i, tes in enumerate(config["grouped_elements"][:J_ast]):
for j, te in enumerate(tes):
try:
idx = label_names.index("{0}_h".format(te.lower()))
ax.set_xticks(xi)
ax.set_xlabel(r"$\textrm{dimension}$")
ax.set_ylabel(r"$\mathbf{L}$")
ylim = np.ceil(10 * np.abs(ax.get_ylim()).max()) / 10
ax.plot([0, D + 1], [0, 0], ":", c="#000000", zorder=-1, lw=0.5)
ax.set_xlim(0.5, D + 0.5)
ax.set_ylim(-ylim, +ylim)
ax.set_yticks([-ylim, 0, ylim])
fig_factor_loads.tight_layout()
savefig(fig_factor_loads, "factor_loads")
# Plot the latent space.
cmap = mpl_utils.discrete_cmap(data_kwds["n_components"], "Spectral")
label_names = [
f"$\\mathbf{{S}}_{{{i}}}$" for i in range(data_kwds["n_latent_factors"])
]
fig_latent = mpl_utils.plot_latent_space(model,
Y,
cmap=cmap,
label_names=label_names)
for ax in fig_latent.axes:
if ax.get_visible():
if ax.is_last_row():
ax.xaxis.set_major_locator(MaxNLocator(3))
if ax.is_first_col():
ax.yaxis.set_major_locator(MaxNLocator(3))
AL = linalg.cholesky(R_opt.T @ R_opt)
R_opt2 = R_opt @ linalg.solve(AL, np.eye(model.n_latent_factors))
chi1 = np.sum(np.abs(A_est @ R - A_astrophysical))
chi2 = np.sum(np.abs(A_est @ R_opt2 - A_astrophysical))
R = R_opt2 if chi2 < chi1 else R
# Now make it a valid rotation matrix.
model.rotate(R, X=X, ensure_valid_rotation=True)
# Identifiability.
J_max = config["max_n_latent_factors_for_colormap"]
cmap = mpl_utils.discrete_cmap(J_max, base_cmap="Spectral_r")
colors = [cmap(j) for j in range(J_max)][::-1]
fig_fac = mpl_utils.plot_factor_loads_and_contributions(model, X,
label_names=label_names, colors=colors)
savefig(fig_fac, "latent-factors-and-contributions")
fig_fac = mpl_utils.plot_factor_loads_and_contributions(model, X,
label_names=label_names, colors=colors,
target_loads=A_astrophysical)
savefig(fig_fac, "latent-factors-and-contributions-with-targets")
# Plot the specific variances.
xticklabels = [r"$\textrm{{{0}}}$".format(ln.split("_")[0].title()) for ln in label_names]
fig_scatter = mpl_utils.plot_specific_scatter(model,
savefig(fig_mml, f"size-{size}-gridsearch-mml")
except:
None
# F*****g save everything.
with open(output_path, "wb") as fp:
pickle.dump((Js, Ks, gJs, gKs, converged, meta, X_H, label_names), fp)
# Compare model with previously saved model.
model = meta["best_models"][config["adopted_metric"]]
saved_models[size] = model
# Plot clustering in data space.
component_cmap = mpl_utils.discrete_cmap(7, base_cmap="Spectral_r")
fig, axes = plt.subplots(5, 3, figsize=(7.1, 9.0))
axes = np.atleast_1d(axes).flatten()
x = X_H.T[label_names.index("fe_h")]
c = np.argmax(model.tau_, axis=1)
K = model.n_components
y_idx = 0
for i, ax in enumerate(axes):
if label_names[i] == "fe_h":
y_idx += 1
y = X_H.T[y_idx] - x
ax.set_xticks(xi)
ax.set_xlabel(r"$\textrm{dimension}$")
ax.set_ylabel(r"$\mathbf{L}$")
ylim = np.ceil(10 * np.abs(ax.get_ylim()).max()) / 10
ax.plot([0, D + 1], [0, 0], ":", c="#000000", zorder=-1, lw=0.5)
ax.set_xlim(0.5, D + 0.5)
ax.set_ylim(-ylim, +ylim)
ax.set_yticks([-ylim, 0, ylim])
fig_factor_loads.tight_layout()
savefig(fig_factor_loads, "factor_loads")
# Plot the latent space.
cmap = mpl_utils.discrete_cmap(n_components, "Spectral")
label_names = [f"$\\mathbf{{S}}_{{{i}}}$" for i in range(n_latent_factors)]
fig_latent = mpl_utils.plot_latent_space(model,
Y,
cmap=cmap,
label_names=label_names)
for ax in fig_latent.axes:
if ax.get_visible():
if ax.is_last_row():
ax.xaxis.set_major_locator(MaxNLocator(3))
if ax.is_first_col():
ax.yaxis.set_major_locator(MaxNLocator(3))
xlim = np.max(np.abs(ax.get_xlim()))
ylim = np.max(np.abs(ax.get_ylim()))