def compute_wc_offline(dataset,
data_loader,
batch_size,
method_args,
name='',
verfication=False):
print('Compute Witness Complex Pairings {name}'.format(name=name))
dist_X_all = torch.ones((len(data_loader), batch_size, batch_size))
pair_mask_X_all = torch.ones((len(data_loader), batch_size, batch_size))
for batch, (img, label) in enumerate(data_loader):
witness_complex = WitnessComplex(img, dataset[:][:][0])
if method_args['n_jobs'] > 1:
witness_complex.compute_simplicial_complex_parallel(
d_max=1,
r_max=method_args['r_max'],
create_simplex_tree=False,
create_metric=True,
n_jobs=method_args['n_jobs'])
else:
witness_complex.compute_simplicial_complex(
d_max=1,
r_max=method_args['r_max'],
create_simplex_tree=False,
create_metric=True)
if witness_complex.check_distance_matrix:
pass
else:
print('WARNING: choose higher r_max')
landmarks_dist = torch.tensor(witness_complex.landmarks_dist)
sorted, indices = torch.sort(landmarks_dist)
kNN_mask = torch.zeros(
(batch_size, batch_size),
device='cpu').scatter(1, indices[:, 1:(method_args['k'] + 1)], 1)
dist_X_all[batch, :, :] = landmarks_dist
pair_mask_X_all[batch, :, :] = kNN_mask
if method_args['match_edges'] == 'verification' and verfication:
ind_X = np.where(pair_mask_X_all[batch, :, :] == 1)
ind_X = np.column_stack((ind_X[0], ind_X[1]))
make_plot(img, ind_X, label, 'name', path_root=None, knn=False)
return dist_X_all, pair_mask_X_all
ys = []
for i, (X_batch, label_batch) in enumerate(dataloader):
if i == 0:
X0 = X_batch
Xs.append(X_batch)
ys.append(label_batch)
data = torch.cat(Xs, dim=0)
labels = torch.cat(ys, dim=0)
model = UMAP(n_neighbors=NN, min_dist=m_dist)
data_l, labels_l = model.get_latent_train(data.numpy(), labels.numpy())
ind_plot = random.sample(range(data.shape[0]), 15000)
make_plot(data_l[ind_plot, :],
None,
labels_l[ind_plot].astype(int).astype(str),
True,
path_to_save=os.path.join(root_to_save, 'all_base.pdf'))
print(torch.all(torch.from_numpy(labels_l).eq(labels)))
ks = [1, 2, 3, 4]
for batch_i, (X_batch, label_batch) in enumerate(dataloader):
if batch_i == 7 or batch_i == 33 or batch_i == 66:
print('compute')
try:
os.mkdir(os.path.join(root_to_save, '{}'.format(batch_i)))
except:
pass
#X,y = model.get_latent_test(X_batch.numpy(), label_batch.numpy())
X, y = data_l[(bs * batch_i):bs *
(batch_i + 1), :], labels_l[(bs * batch_i):bs *
witnesses_tensor = torch.from_numpy(witnesses)
landmarks_tensor = torch.from_numpy(landmarks)
witness_complex = WitnessComplex(landmarks_tensor, witnesses_tensor)
witness_complex.compute_simplicial_complex(d_max=1,
r_max=10,
create_simplex_tree=False,
create_metric=True)
for k in ks:
print('{} out of {}'.format(counter, ntot))
landmarks_dist = torch.tensor(witness_complex.landmarks_dist)
sorted, indices = torch.sort(landmarks_dist)
kNN_mask = torch.zeros((n_samples, n_samples), device='cpu').scatter(1, indices[:, 1:(k+1)], 1)
pairings_i = np.where(kNN_mask.numpy() == 1)
pairings = np.column_stack((pairings_i[0], pairings_i[1]))
name = 'wc{nw}_k{k}_seed{seed}'.format(nw = N_WITNESSES,k = k, seed = seed)
make_plot(landmarks, pairings, color,name = name, path_root = path_to_save, knn = False, show = True, dpi = 400, cmap = plt.cm.viridis)
counter += 1
for seed in [30]:
print('{} out of {}'.format(progress_count, tot_count))
progress_count += 1
name = 'vr_ns{}_seed{}'.format(n_samples, seed)
data, color = dataset_sampler.sample(n_samples, seed=seed)
data, pairings, color = make_data(data, color, name=name)
# path_pairings = '{}pairings_{}.npy'.format(PATH_ROOT_SWISSROLL, name)
# path_data = '{}data_{}.npy'.format(PATH_ROOT_SWISSROLL, name)
# path_color = '{}color_{}.npy'.format(PATH_ROOT_SWISSROLL, name)
# pairings, data, color = np.load(path_pairings), np.load(path_data), np.load(path_color)
#
make_plot(data,
pairings,
color,
name=name,
path_root=path_to_save,
cmap=plt.cm.viridis)
# name = '512_1'
# data, color = dataset_sampler.sample(512)
# make_data(data, color, name = name)
#
# path_pairings = '{}pairings_{}.npy'.format(PATH_ROOT, name)
# path_data = '{}data_{}.npy'.format(PATH_ROOT, name)
# path_color = '{}color_{}.npy'.format(PATH_ROOT, name)
# pairings, data, color = np.load(path_pairings), np.load(path_data), np.load(path_color)
#
# #
# make_plot(data, pairings, color, name = name)
#