def plot_latents(df,
X_trans,
Z_mu,
labels,
distances,
latent_titles=None,
suptitle=None,
path=None):
'''
plot latents
'''
util_funcs.plt_setup()
fig, axes = plt.subplots(3, 8, figsize=(12, 8))
# plot correlations
for latent_idx, ax in enumerate(axes[0]):
heatmap_corrs = plot_funcs.correlate_heatmap(len(labels),
len(distances),
X_trans,
Z_mu[:, latent_idx])
plot_funcs.plot_heatmap(ax,
heatmap=heatmap_corrs,
row_labels=labels,
col_labels=distances,
set_row_labels=latent_idx == 0,
cbar=True)
if latent_titles is not None:
t = latent_titles[latent_idx]
ax.set_title(t)
# plot maps
for city_ax_row, city_name, x, y in zip([axes[1], axes[2]],
['Cambridge', 'Milton Keynes'],
(545700, 485970),
(258980, 236920)):
x_extents = (x - 2500, x + 2500)
y_extents = (y - 4500, y + 4500)
for latent_idx, ax in enumerate(city_ax_row):
plot_funcs.plot_scatter(ax,
df.x,
df.y,
Z_mu[:, latent_idx],
x_extents=x_extents, # 10000
y_extents=y_extents, # 10000
relative_extents=False,
s_min=0,
s_max=0.8,
c_exp=5,
s_exp=3.5)
ax.set_xlabel(f'{city_name} #{latent_idx}')
if suptitle is not None:
plt.suptitle(suptitle)
if path is not None:
plt.savefig(path, dpi=300)
else:
plt.show()
np.exp(Z_log_var.numpy().astype('float64')))
# average kl divergence per latent
kl_vec = np.mean(kl_loss, axis=0)
# plot
util_funcs.plt_setup()
fig, axes = plt.subplots(2, 6, figsize=(12, 8))
for ax_row, inverse in zip(axes, [False, True]):
L = np.copy(Z_mu)
if inverse:
L *= -1
for latent_idx in range(latent_dim):
ax = ax_row[latent_idx]
plot_funcs.plot_scatter(ax,
df_20.x,
df_20.y,
L[:, latent_idx],
s_min=0,
s_max=0.8,
c_exp=3,
s_exp=5)
ax.set_title(f'Dim. {latent_idx + 1} - ' +
r'$D_{KL}=' +
f'{kl_vec[latent_idx]:.1f}$')
if inverse:
ax.set_title(f'Dim. {latent_idx + 1} - Inverse')
plt.suptitle(
r"Geographic mappings (and inverses) for the latents of the VAE model"
)
path = f'../phd-doc/doc/images/signatures/latents_map_{beta}_{cap}_{seed}_{theme_base}.pdf'
plt.savefig(path, dpi=300)
for x, y, town_name in [(485970, 236920, 'Milton Keynes'),
(545700, 258980, 'Cambridge'),
(635633, 170829, 'Margate')]:
x_extents = (x - 2500, x + 2500)
y_extents = (y - 5000, y + 5000)
util_funcs.plt_setup()
fig, ax = plt.subplots(1, 1, figsize=(8, 8))
plot_funcs.plot_scatter(
ax,
df_20.x,
df_20.y,
# color range is from 0.0 to 1.0
# don't manipulate in this context
c=y_col,
cmap=util_funcs.cityseer_cmap(),
s=1,
vmin=0.0, # in some cases plt doesn't correct infer extents
vmax=1.0,
x_extents=x_extents,
y_extents=y_extents,
relative_extents=False)
plt.suptitle(f"{town_name} hyperlocal predictions")
path = f'../phd-doc/doc/images/predictive/stage_2_DNN_{town_name}_{clf.theme}_yc{y_conf}.pdf'
plt.savefig(path, dpi=300)
# %%
# create columns for percentages
for pop_id, town_row in selected_data.iterrows():
pred = y_plot[X_raw.city_pop_id == pop_id]
selected_data.loc[pop_id,
for theme_set, theme_labels, t_meta, p_meta, weighted in zip(
theme_sets, theme_label_sets, theme_metas, path_metas, theme_wt):
print(f'processing metas: {t_meta}')
util_funcs.plt_setup()
fig, axes = plt.subplots(3, 2, figsize=(7, 10))
for t_idx, (t, label) in enumerate(zip(theme_set, theme_labels)):
for d_idx, (dist,
beta) in enumerate(zip([200, 800], [r'-0.02', r'-0.005'])):
ax = axes[t_idx][d_idx]
theme = t.format(dist=dist)
data = df_20[theme]
plot_funcs.plot_scatter(fig,
ax,
df_20.x,
df_20.y,
km_per_inch=3.5,
vals=data,
s_exp=4)
ax.set_xlabel(f'{label} ' + r'$d_{max}=' + f'{dist}m$')
plt.suptitle(f'Plots of {t_meta} mixed-use measures for inner London')
path = f'../phd-doc/doc/images/diversity/diversity_comparisons_{p_meta}.pdf'
plt.savefig(path, dpi=300)
# %% prepare PCA
# use of bands gives slightly more defined delineations for latent dimensions
pca_columns_dist = [
'ac_accommodation_{dist}', 'ac_eating_{dist}', 'ac_drinking_{dist}',
'ac_commercial_{dist}', 'ac_tourism_{dist}', 'ac_entertainment_{dist}',
'ac_government_{dist}', 'ac_manufacturing_{dist}', 'ac_retail_food_{dist}',
'ac_retail_other_{dist}', 'ac_transport_{dist}', 'ac_health_{dist}',
util_funcs.plt_setup()
fig, axes = plt.subplots(4, 4, figsize=(7, 10))
# iterate each town for a given set
for ax_col_idx, (centre, town_name, town_pop_id) in enumerate(set):
# map 1) Boundary designation
# targets will be 1 vs. 0 for new town like or not
c = cmap(float(selected_data.loc[town_pop_id, 'targets']))
ax = axes[0, ax_col_idx]
ax.set_xlabel(f'{town_name}')
if ax_col_idx == 0:
ax.set_ylabel(f'Predictions by boundary')
plot_funcs.plot_scatter(fig,
ax,
df_20.x,
df_20.y,
c=c,
cmap=cmap,
s=s,
centre=centre,
km_per_inch=3)
# map 2) classifier
ax = axes[1, ax_col_idx]
ax.set_xlabel(f'{town_name}')
if ax_col_idx == 0:
ax.set_ylabel(f'DNN local predictions')
plot_funcs.plot_scatter(
fig,
ax,
df_20.x,
df_20.y,
# color range is from 0.0 to 1.0
theme_label_sets,
theme_metas,
theme_wts)):
print(f'processing set: {n}')
util_funcs.plt_setup()
fig, axes = plt.subplots(3, 2, figsize=(7, 10))
for t_idx, (t, label) in enumerate(zip(theme_set, theme_labels)):
for d_idx, (dist, beta) in enumerate(zip([400, 1600],
[r'0.01', r'0.0025'])):
ax = axes[t_idx][d_idx]
theme = t.format(dist=dist)
data = df_20[theme]
plot_funcs.plot_scatter(fig,
ax,
df_20.x,
df_20.y,
vals=data,
km_per_inch=3.5,
s_exp=theme_wt)
ax.set_xlabel(f'{label} ' + '$d_{max}=' + f'{dist}m$')
plt.suptitle(f'Plots of centrality measures for inner London (set {n + 1})')
path = f'../phd-doc/doc/images/centrality/primal_centrality_comparisons_{theme_meta}.pdf'
plt.savefig(path, dpi=300)
# %%
'''
demonstrate the typical distributions
the contention is that low values preclude mixed-uses but high values don't guarantee them
the spike in the distributions is an artefact of search distances vis-a-vis straight roads with no intersections,
in which case a certain number of adjacent nodes is reached, resulting in a stable preponderance of a certain value
this would not be present in gravity-weighted forms of centrality
(409352, 425145, 'Halifax'), (497561, 371383, 'Lincoln'),
(635633, 170829, 'Margate'), (439743, 557041, 'Sunderland'),
(623184, 308525, 'Norwich'), (353876, 429315, 'Preston'),
(330891, 436268, 'Blackpool')
]:
x_extents = (x - 2500, x + 2500)
y_extents = (y - 5000, y + 5000)
util_funcs.plt_setup()
fig, ax = plt.subplots(1, 1, figsize=(8, 8))
mappable = plot_funcs.plot_scatter(
ax,
df_20.x,
df_20.y,
# color range is from 0.0 to 1.0
# don't manipulate in this context
c=m2_pred_nt,
cmap=util_funcs.cityseer_cmap(),
s=1,
x_extents=x_extents,
y_extents=y_extents,
relative_extents=False)
ax.set_axis_off()
divider = make_axes_locatable(ax)
cax = divider.append_axes('right', size='2%', pad=0.05)
plt.colorbar(mappable, cax=cax, aspect=80)
plt.suptitle(
f'{town_name} - 20m resolution "New Town" predicted probabilities')
path = f'../phd-doc/doc/images/predictive/nt_hyperlocal_pred_{town_name}_{m2.theme}_{clf.theme}.pdf'
plt.savefig(path, dpi=300)
# %%
np.nanmax(y_pred),
abs(np.nanmin(y_all_idx)),
np.nanmax(y_all_idx)
])
vmin, vmax = 0, mm
cmap = LinearSegmentedColormap.from_list(
'reds', ['#ffffff', '#ff6659', '#d32f2f'])
pow = 1.25
c = vals**pow
s = 0.01 + (vals**pow / vmax * 1.5)
im = plot_funcs.plot_scatter(fig,
ax,
df_20.x[select_idx],
df_20.y[select_idx],
c=c,
s=s,
vmin=vmin,
vmax=vmax,
cmap=cmap,
centre=centre,
km_per_inch=km_per_inch)
ax.set_xlabel(title)
cbar = fig.colorbar(im, ax=ax, aspect=100, pad=0.05)
plt.suptitle(
f'Observed, predicted, differenced eateries at 400m threshold. Test set accuracy: {y_score_r2:.1%} r2'
)
path = f'../phd-doc/doc/images/predictive/eatery_predictions_400_{reg.theme}.pdf'
plt.savefig(path, dpi=300)
# %%
util_funcs.plt_setup()
cluster_assign_VaDE = np.argmax(cluster_probs_VaDE, axis=1)
# %%
'''
Overall cluster max plot
'''
util_funcs.plt_setup()
fig, ax = plt.subplots(1, 1, figsize=(7, 7))
sample_colours, sample_sizes, mu_means, mu_means_cols, sorted_cluster_idx = \
plot_funcs.map_diversity_colour(X_raw,
cluster_assign_VaDE,
n_components=n_components)
plot_funcs.plot_scatter(fig,
ax,
table.x,
table.y,
c=sample_colours,
s=sample_sizes,
km_per_inch=3.25,
rasterized=True)
# custom legend
mu_means_sorted = mu_means[sorted_cluster_idx]
mu_means_cols_sorted = mu_means_cols[sorted_cluster_idx]
legend_elements = []
for cluster_idx, mu_mean, mu_mean_col in zip(sorted_cluster_idx,
mu_means_sorted,
mu_means_cols_sorted):
legend_elements.append(
Patch(facecolor=mu_mean_col,
edgecolor='w',
label=f'#{cluster_idx + 1}: {mu_mean:.2f}'))
leg = ax.legend(handles=legend_elements,
test_indices=test_idx)
trainer.train()
else:
vade.load_weights(str(dir_path / 'weights'))
# %% classify
cluster_probs_VaDE = vade.classify(X_trans).numpy()
cluster_assign_VaDE = np.argmax(cluster_probs_VaDE, axis=1)
# %% plot cluster matrix
plot_funcs.plot_prob_clusters(
X_raw, cluster_probs_VaDE, n_components,
f'{vade.theme}_d{latent_dim}_e{epochs}_comp{n_components}_clust_ind',
table.x, table.y)
# plot combined map
util_funcs.plt_setup()
fig, ax = plt.subplots(1, 1, figsize=(12, 8))
colours, sizes = plot_funcs.map_diversity_colour(X_raw,
cluster_assign_VaDE,
n_components=n_components)
plot_funcs.plot_scatter(ax,
table.x,
table.y,
c=colours,
s=sizes,
x_extents=(-7000, 11500),
y_extents=(-4500, 7500),
rasterized=True)
path = f'../phd-doc/doc/images/signatures/{vade.theme}_d{latent_dim}_e{epochs}_comp{n_components}_clust_comb.pdf'
plt.savefig(path, dpi=300)
plot latent maps
'''
# plot
util_funcs.plt_setup()
fig, axes = plt.subplots(2, 6, figsize=(7, 9))
for ax_row, inverse in zip(axes, [False, True]):
L = np.copy(Z_mu)
if inverse:
L *= -1
for latent_idx in range(latent_dim):
ax = ax_row[latent_idx]
plot_funcs.plot_scatter(fig,
ax,
table.x,
table.y,
vals=L[:, latent_idx],
s_min=0,
s_max=1,
c_exp=2,
s_exp=3,
rasterized=True)
if not inverse:
ax.set_xlabel(f'#{latent_idx + 1} - ' + r'$D_{KL}=' +
f'{kl_vec[latent_idx]:.2f}$')
else:
ax.set_xlabel(f'#{latent_idx + 1} - Inverse')
plt.suptitle(
r"Geographic mappings (and inverses) for the latents of the VAE model")
path = f'../phd-doc/doc/images/signatures/latents_map_inv_{beta}_{str(cap).replace(".", "_")}.pdf'
plt.savefig(path, dpi=300)
# %%