def main():
for ds_dir in fullpath_list(args.root_dir):
ds_name = os.path.basename(ds_dir)
if args.datasets and ds_name not in args.datasets:
continue
exp_dir = os.path.join(ds_dir, 'flipp_0.0000', loss_fn_str)
width = 6.5 if args.no_ylabel else 7
fig, ax = plt.subplots(figsize=(width, 5))
for man_dir in fullpath_list(exp_dir):
man_name = os.path.basename(man_dir)
if args.manifolds and man_name not in args.manifolds:
continue
factor_names = manifold_factors_from_path_label(man_name)
man_factors = build_manifold(*factor_names)
man_label = manifold_label_for_display(*factor_names)
dim = sum([m.dim for m in man_factors])
if args.dims and dim not in args.dims:
continue
# load the angle ratio samples
samples = load_angle_ratio_samples(man_dir)
# plot them
plot_angle_ratios(
ax,
samples,
label=man_label,
color=get_color_for_manifold(man_name))
# save the figure
configure_and_save_plots(ax, fig, ds_name, loss_fn_str)
def main():
if args.verbose:
logging.getLogger().setLevel(logging.DEBUG)
# use double precision by default
torch.set_default_dtype(torch.float64)
for ds_dir in fullpath_list(args.root_dir):
ds_name = os.path.basename(ds_dir)
if args.datasets and ds_name not in args.datasets:
continue
# Load the dataset graph in order to compute F1 scores.
_, g = load_graph_pdists(os.path.join('../data',
ds_name + '.edges.gz'),
cache_dir='.cached_pdists')
n_nodes = g.number_of_nodes()
with Timer('constructing FastPrecision'):
fp = FastPrecision(g)
nodes_per_layer = fp.nodes_per_layer()[1:]
nodes_per_layer = nodes_per_layer / np.sum(nodes_per_layer)
for flipp_dir in fullpath_list(ds_dir):
flipp = os.path.basename(flipp_dir).split('_')[1]
if args.flip_probabilities and flipp not in args.flip_probabilities:
continue
for loss_fn_dir in fullpath_list(flipp_dir):
loss_fn_str = os.path.basename(loss_fn_dir)
if args.loss_fns and loss_fn_str not in args.loss_fns:
continue
# create one plot per (dataset, flipp, loss_fn) combination
width = 6 if args.leftmost or args.rightmost else 5
fig, ax = plt.subplots(figsize=(width, 5))
plot_id = 0
for man_dir in fullpath_list(loss_fn_dir):
man_name = os.path.basename(man_dir)
if args.manifolds and man_name not in args.manifolds:
continue
factor_names = manifold_factors_from_path_label(man_name)
man_factors = build_manifold(*factor_names)
man_label = manifold_label_for_display(*factor_names)
dim = sum([m.dim for m in man_factors])
if args.dims and dim not in args.dims:
continue
# compute the metric
means, stds = comp_metric(ds_name, n_nodes, fp, flipp,
loss_fn_str, man_dir,
man_factors, man_label)
if means is None:
continue
# add them to the plot
plot_f1_scores(ax, means, stds, plot_id, label=man_label)
plot_id += 1
# save the figure
configure_and_save_plots(ax, fig, ds_name, flipp, loss_fn_str,
nodes_per_layer)
def main():
for dim_dir in fullpath_list(args.root_dir, only_dirs=True):
dim = os.path.basename(dim_dir)
if args.dims and dim not in args.dims:
continue
dim = int(dim)
width = 7 if dim == 2 else 6
fig, ax = plt.subplots(figsize=(width, 5))
bins = np.linspace(xmin, xmax, 100)
i = 0
for man_dir in fullpath_list(dim_dir, only_dirs=True):
man_name = os.path.basename(man_dir)
if args.manifolds and man_name not in args.manifolds:
continue
filename = os.path.join(man_dir, 'angle_ratios.npy')
values = np.load(filename)
plt.hist(
values,
bins,
density=True,
label=manifold_label_for_display(man_name),
color=get_color_for_manifold(man_name),
alpha=0.5)
i += 1
ax.grid(color='lightgray', lw=2, alpha=0.5)
ax.set_axisbelow(True)
ax.set_xlabel('Normalized Sum of Angles')
if dim == 2:
ax.set_ylabel('PDF')
else:
ax.set_yticklabels([])
ax.set_title('Triangles Thickness (n={})'.format(dim), y=1.18)
ax.set_xlim(xmin, xmax)
ax.set_ylim(top=args.ymax)
hs, labels = ax.get_legend_handles_labels()
hs, labels = zip(*sorted(zip(hs, labels), key=lambda t: t[1]))
ax.legend(
hs,
labels,
bbox_to_anchor=(0, 1.02, 1, 0.2),
loc='lower left',
mode='expand',
borderaxespad=0,
ncol=4)
plt.tight_layout()
figpath = os.path.join(args.save_dir, f'angles-{dim}.pdf')
fig.savefig(figpath, bbox_inches='tight')
plt.close()
def main():
torch.set_default_dtype(torch.float64)
with ProcessPoolExecutor(max_workers=args.num_cpus) as pool:
futures = []
for dim_dir in fullpath_list(args.root_dir, only_dirs=True):
for man_dir in fullpath_list(dim_dir, only_dirs=True):
man_name = os.path.basename(man_dir)
if args.manifolds and man_name not in args.manifolds:
continue
f = pool.submit(grid_fn, man_dir, man_name)
futures.append(f)
for f in futures:
f.result(None)
def load_angle_ratio_samples(man_dir):
all_samples = []
for run_dir in fullpath_list(man_dir):
samples = load_samples_for_best_embedding(run_dir)
all_samples.append(samples)
return np.concatenate(all_samples)
def comp_metric(ds_name, n_nodes, fp, flipp, loss_fn_str, man_dir, man_factors,
man_label):
run_dirs = list(fullpath_list(man_dir))
num_pdists = n_nodes * (n_nodes - 1) // 2
n_runs = len(run_dirs)
all_pdists = np.ndarray(shape=(num_pdists * n_runs))
for i, run_dir in enumerate(run_dirs):
# load the embedding
embedding = ManifoldEmbedding(n_nodes, man_factors)
emb_state_dict, _, _ = load_best_embedding(run_dir)
embedding.load_state_dict(emb_state_dict)
# compute the pairwise distances
with Timer('computing pdists'), torch.no_grad():
man_pdists = embedding.compute_dists(None)
man_pdists.sqrt_()
indices = np.arange(i * num_pdists, (i + 1) * num_pdists)
all_pdists[indices] = man_pdists.numpy()
# compute the f1 scores
run_id = make_run_id(dataset=ds_name,
fp=flipp,
loss_fn=loss_fn_str,
manifold=man_label)
logging.info('Computing F1 scores for (%s)', run_id)
with Timer('computing F1 scores'):
means, stds = fp.layer_mean_average_f1_scores(all_pdists, n_runs)
return means[:args.max_layers], stds[:args.max_layers]
def main():
with ProcessPoolExecutor(max_workers=args.num_cpus) as pool:
futures = []
for dim_dir in fullpath_list(args.root_dir, only_dirs=True):
dim = os.path.basename(dim_dir)
if args.dims and dim not in args.dims:
continue
f = pool.submit(grid_fn, dim_dir, int(dim))
futures.append(f)
for f in futures:
f.result(None)
def grid_fn(input_dir, dim):
results = {}
for man_dir in fullpath_list(input_dir, only_dirs=True):
man_name = os.path.basename(man_dir)
results[man_name] = {}
for thresh_dir in fullpath_list(man_dir, only_dirs=True):
thresh = os.path.basename(thresh_dir)
results[man_name][thresh] = {}
for quantity in args.plots:
ret = read_quantity_results(thresh_dir, quantity)
if ret is not None:
results[man_name][thresh][quantity] = ret
quantity_labels = dict(
degrees='Node Degree', seccurvs='Graph Sectional Curvature')
quantity_titles = dict(
degrees='Degree Distributions', seccurvs='Curvature Estimates')
for quantity in args.plots:
ls = ['-', '--', '-.', ':']
ms = ['o', 'v', '*', 'd', 'x', '1']
width = 7 if dim == 2 else 6
height = 5.5 if quantity == 'degrees' else 5
fig, ax = plt.subplots(figsize=(width, height))
for i, (man_name, values) in enumerate(results.items()):
if man_name.startswith('euc') and quantity != 'degrees':
continue
xs = []
ys = []
for k in values.keys():
thresh_values = values[k]
if quantity in thresh_values:
xs.append(float(k))
ys.append(thresh_values[quantity])
xs, ys = zip(*sorted(zip(xs, ys), key=lambda e: e[0]))
p25, p50, p75 = zip(*ys)
line, = plt.plot(
xs,
p50,
label=manifold_label_for_display(man_name),
color=get_color_for_manifold(man_name),
lw=6 - i,
ls=ls[i % 4],
marker=ms[i % 6],
ms=10)
ax.fill_between(xs, p25, p75, facecolor=line.get_color(), alpha=0.3)
ax.set_ylim(bottom=args.ymin, top=args.ymax)
ax.set_xlim(left=0, right=args.xmax)
ax.xaxis.set_major_locator(MaxNLocator(integer=True))
ax.grid(color='lightgray', lw=2, alpha=0.5)
ax.set_xlabel('Distance Threshold')
if dim == 2:
ax.set_ylabel(quantity_labels[quantity])
else:
ax.set_yticklabels([])
ax.set_title(
'{} (n={})'.format(quantity_titles[quantity], dim),
y=1.32 if quantity == 'degrees' else 1.18)
ax.set_axisbelow(True)
hs, labels = ax.get_legend_handles_labels()
hs, labels = zip(*sorted(zip(hs, labels), key=lambda t: t[1]))
ax.legend(
hs,
labels,
bbox_to_anchor=(0, 1.02, 1, 0.2),
loc='lower left',
mode='expand',
borderaxespad=0,
ncol=2 if quantity == 'degrees' else 3)
plt.tight_layout()
fig_name = os.path.join(args.save_dir, f'{quantity}-{dim}.pdf')
fig.savefig(fig_name, bbox_inches='tight')
plt.close()
def main():
for dim_dir in fullpath_list(args.root_dir, only_dirs=True):
dim = os.path.basename(dim_dir)
if args.dims and dim not in args.dims:
continue
dim = int(dim)
width = 7 if dim == 3 else 6
fig, ax = plt.subplots(figsize=(width, 5))
bins = np.linspace(xmin, xmax, 100)
colors = ['tab:orange', 'tab:green']
i = 0
ymax = 0
for man_dir in sorted(fullpath_list(dim_dir, only_dirs=True)):
man_name = os.path.basename(man_dir)
if args.manifolds and man_name not in args.manifolds:
continue
filename = os.path.join(man_dir, 'angle_ratios.npy')
values = np.load(filename)
ret = plt.hist(
values,
bins,
density=True,
label=manifold_label_for_display(man_name),
color=colors[i],
alpha=0.5)
ymax = max(ymax, ret[0].max())
i += 1
ax.grid(color='lightgray', lw=2, alpha=0.5)
ax.set_axisbelow(True)
ax.set_xlabel('Normalized Sum of Angles')
if dim == 3:
ax.set_ylabel('PDF')
else:
ax.annotate(
'cut off (max {:.2f})'.format(ymax),
xy=(-0.47, 19.5),
xytext=(-0.38, 17),
arrowprops=dict(facecolor='k', arrowstyle='-|>'),
bbox=dict(boxstyle='round', facecolor='none'))
ax.set_yticklabels([])
ax.set_title('Triangles Thickness (n={})'.format(dim), y=1.18)
ax.set_xlim(xmin, xmax)
if args.ymax:
ax.set_ylim(top=args.ymax + 0.1)
hs, labels = ax.get_legend_handles_labels()
hs, labels = zip(*sorted(zip(hs, labels), key=lambda t: t[1]))
ax.legend(
hs,
labels,
bbox_to_anchor=(0, 1.02, 1, 0.2),
loc='lower left',
mode='expand',
borderaxespad=0,
ncol=4)
plt.tight_layout()
figpath = os.path.join(args.save_dir, f'angles-{dim}.pdf')
fig.savefig(figpath, bbox_inches='tight')
plt.close()