def optimize_global_layout(
P,
Z,
a,
b,
alpha,
n_epochs,
verbose=False,
savefig=False,
label=None
):
costs = []
for i in range(n_epochs):
d_squared = np.square(adjacency_matrix(Z))
z_diff = np.expand_dims(Z, axis=1) - np.expand_dims(Z, axis=0)
d_inverse = np.expand_dims(pow(1 + a * d_squared ** b, -1), axis=2)
# Q is the normalized distance in low dimensional space
Q = pow(0.001 + d_squared, -1)
np.fill_diagonal(Q, 0)
Q = np.dot(1 - P, Q)
Q /= np.sum(Q, axis=1, keepdims=True)
# Q /= Q.max()
grad = np.expand_dims(
2 * a * b * P * (1e-12 + d_squared) ** (b - 1) - 2 * b * Q, axis=2
)
dZ = np.sum(grad * z_diff * d_inverse, axis=1)
Z -= alpha * dZ
if verbose:
tmpZ = Z / np.sum(Z, axis=1, keepdims=True)
cost = get_CE(P, tmpZ, d_squared, a, b)
# cost = get_DTM(P, tmpZ, sigma=0.1)
costs.append(cost)
print(
f"[INFO] Current loss: {cost:.6f}, @ iteration: {i+1}/{n_epochs}"
)
if savefig:
if i < 10 or i == 30:
from umato.umato_ import plot_tmptmp
plot_tmptmp(data=Z, label=label, name=f"pic1_global{i}")
return Z
for i in range(len(learning_rate)):
for j in range(len(perplexity)):
# run TSNE
y = TSNE(n_components=args.dim,
perplexity=perplexity[j],
learning_rate=learning_rate[i],
init="pca",
n_jobs=-1,
random_state=0,
verbose=2).fit_transform(x)
# save as csv
path = os.path.join(os.getcwd(), "visualization", "public",
"results", args.data)
save_csv(path,
alg_name=f"tsne_{perplexity[j]}_{learning_rate[i]}",
data=y,
label=label)
else:
y = TSNE(n_components=args.dim,
random_state=0,
verbose=2,
init="pca",
n_jobs=-1).fit_transform(x)
path = os.path.join(os.getcwd(), "visualization", "public", "results",
args.data)
save_csv(path, alg_name="tsne", data=y, label=label)
plot_tmptmp(y, label, "tsne")
def nn_layout_optimize(
head_embedding,
tail_embedding,
head,
tail,
hub_info,
n_epochs,
n_vertices,
epochs_per_sample,
a,
b,
rng_state,
gamma,
learning_rate=1.0,
negative_sample_rate=5.0,
parallel=False,
verbose=False,
k=0,
label=None,
):
(_, dim) = head_embedding.shape
move_other = head_embedding.shape[0] == tail_embedding.shape[0]
alpha = learning_rate
epochs_per_negative_sample = epochs_per_sample / negative_sample_rate
epoch_of_next_negative_sample = epochs_per_negative_sample.copy()
epoch_of_next_sample = epochs_per_sample.copy()
optimize_fn = numba.njit(
_nn_layout_optimize_single_epoch, fastmath=True, parallel=parallel
)
for n in range(n_epochs):
optimize_fn(
head_embedding,
tail_embedding,
head,
tail,
hub_info,
n_vertices,
epochs_per_sample,
a,
b,
rng_state,
gamma,
dim,
move_other,
alpha,
epochs_per_negative_sample,
epoch_of_next_negative_sample,
epoch_of_next_sample,
n,
)
alpha = learning_rate * (1.0 - (float(n) / float(n_epochs)))
if verbose and n % 10 == 0:
from umato.umato_ import plot_tmptmp
plot_tmptmp(data=head_embedding, label=label, name=f"pic3_{k}_local{n}")
print("\tcompleted ", n, " / ", n_epochs, "epochs")
return head_embedding
from umato.utils import init_position
import argparse
parser = argparse.ArgumentParser(description="args for umato")
parser.add_argument("--data", type=str, help="choose data: spheres, mnist, fmnist, kmnist, flow, swissroll, scurve, single-cell", default="allen")
parser.add_argument("--hub_num", type=int, help="choose number of hubs", default=400)
parser.add_argument("--n_samples", type=int, help="choose number of samples", default=1500)
parser.add_argument("--init", type=str, help="choose initialization method", default="pca")
args = parser.parse_args()
if __name__ == "__main__":
x, label = get_data(args.data, n_samples=args.n_samples)
y = umato.UMATO(verbose=True, ll=label, hub_num=args.hub_num).fit_transform(x)
plot_tmptmp(y, label, f"umato")
save_csv('./', alg_name=f"umato", data=y, label=label)
# x = x[np.arange(0, 10000, 50)]
# label = label[np.arange(0, 10000, 50)]
# for epoch in [200, 500, 1000, 2000, 5000]:
# x, label = get_data(args.data, n_samples=args.n_samples) # spheres, mnist, fmnist, kmnist
# y = umato.UMATO(verbose=True, ll=label, hub_num=args.hub_num, global_n_epochs=epoch).fit_transform(x)
# plot_tmptmp(y, label, f"umato_{args.data}_{epoch}")
# save_csv('./', alg_name=f"umato_{args.data}_{epoch}", data=y, label=label)
# # UMTO
# for dt in ['fmnist', 'mnist', 'kmnist']:
# # x = load_digits() # (1797, 64 dim)
# x, label = get_data(dt, n_samples=args.n_samples) # spheres, mnist, fmnist, kmnist
parser = argparse.ArgumentParser(description="UMAP embedding")
parser.add_argument("--data", type=str, help="choose dataset", required=True)
parser.add_argument("--dim", type=str, help="choose embedding dimension", default=2)
parser.add_argument("--init", type=str, help="choose initialization method", default="pca")
parser.add_argument("--hp", type=bool, help="whether to explore hyperparameter settings", default=False)
parser.add_argument("--n_samples", type=int, help="choose number of samples", default=1500)
args = parser.parse_args()
if __name__ == "__main__":
x, label = get_data(args.data, n_samples=args.n_samples)
y = UMAP(n_components=args.dim, verbose=True).fit_transform(x)
plot_tmptmp(y, label, f"umap")
save_csv('./', alg_name=f"umap", data=y, label=label)
# if args.hp:
# # read data
# x, label = get_data(args.data, n_samples=args.n_samples)
# n_neighbor = np.arange(5, 55, 5)
# min_dist = np.arange(0, 1.1, 0.1)
# for i in range(len(n_neighbor)):
# for j in range(len(min_dist)):
# # run UMAP
# y = UMAP(n_components=args.dim, n_neighbors=n_neighbor[i], min_dist=min_dist[j], verbose=True).fit_transform(x)
import os
from .dataset import get_data, save_csv
from umato.umato_ import plot_tmptmp
parser = argparse.ArgumentParser(description="PCA embedding")
parser.add_argument("--data", type=str, help="choose dataset", required=True)
parser.add_argument("--dim",
type=str,
help="choose embedding dimension",
default=2)
parser.add_argument("--n_samples",
type=int,
help="choose number of samples",
default=1500)
args = parser.parse_args()
if __name__ == "__main__":
# read data
x, label = get_data(args.data, n_samples=args.n_samples)
# run PCA
y = PCA(n_components=args.dim).fit_transform(x)
# save as csv
path = os.path.join(os.getcwd(), "visualization", "public", "results",
args.data)
plot_tmptmp(y, label, "pca")
save_csv(path, alg_name="pca", data=y, label=label)