def get_topK_parses(img,
k,
score_fn,
configs_per=100,
trials_per=800,
device=None,
seed=3,
**grp_kwargs):
# generate random walks (the "base parses")
base_parses = generate_random_parses(I=img, seed=seed, **grp_kwargs)
# convert strokes to minimal splines
base_parses = [process_parse(parse, device) for parse in base_parses]
# search for best stroke ordering & stroke direction configurations
np.random.seed(seed)
n = len(base_parses)
parses = []
log_probs = []
for i in range(n):
parses_i, log_probs_i = search_parse(base_parses[i], score_fn,
configs_per, trials_per)
parses.extend(parses_i)
log_probs.append(log_probs_i)
log_probs = torch.cat(log_probs)
# refine to unique & sort
log_probs, idx = np.unique(log_probs.cpu().numpy(), return_index=True)
log_probs = torch.from_numpy(log_probs).flip(dims=[0])
idx = torch.from_numpy(idx).flip(dims=[0])
parses = [parses[i] for i in idx]
return parses[:k], log_probs[:k]
def main():
# load image to numpy binary array
img = imageio.imread('./image_H.jpg')
img = np.array(img > 200)
# generate random parses
parses = generate_random_parses(img, seed=3)
# plot parsing results
nparse = len(parses)
n = math.ceil(nparse / 10)
m = 10
fig, axes = plt.subplots(n, m + 1, figsize=(m + 1, n))
# first column
axes[0, 0].imshow(img, cmap=plt.cm.binary)
axes[0, 0].set_xticks([])
axes[0, 0].set_yticks([])
axes[0, 0].set_title('Input')
for i in range(1, n):
axes[i, 0].set_axis_off()
# remaining_columns
for i in range(n):
for j in range(1, m + 1):
ix = i * m + (j - 1)
if ix >= nparse:
axes[i, j].set_axis_off()
continue
plot_parse(axes[i, j], parses[ix])
plt.subplots_adjust(hspace=0., wspace=0.)
plt.show()
def get_topK_parses(img,
k,
score_fn,
configs_per=100,
trials_per=800,
device=None,
seed=3,
pp_kwargs=None,
**grp_kwargs):
""" copied from
gns.inference.parsing.top_k, except allwoing for pp_kwargs, which is useful
if want to skip spline fit optimiziation step
"""
from pybpl.matlab.bottomup import generate_random_parses
from gns.inference.parsing.top_k import search_parse
# generate random walks (the "base parses")
base_parses = generate_random_parses(I=img, seed=seed, **grp_kwargs)
# convert strokes to minimal splines
base_parses = [
process_parse(parse, device, **pp_kwargs) for parse in base_parses
]
# search for best stroke ordering & stroke direction configurations
np.random.seed(seed)
n = len(base_parses)
parses = []
log_probs = []
for i in range(n):
parses_i, log_probs_i = search_parse(base_parses[i], score_fn,
configs_per, trials_per)
parses.extend(parses_i)
log_probs.append(log_probs_i)
log_probs = torch.cat(log_probs)
# refine to unique & sort
log_probs, idx = np.unique(log_probs.cpu().numpy(), return_index=True)
log_probs = torch.from_numpy(log_probs).flip(dims=[0])
idx = torch.from_numpy(idx).flip(dims=[0])
parses = [parses[i] for i in idx]
return parses[:k], log_probs[:k]