def __getitem__(self, item):
graph = self._gen_graph(item)
n = graph.nr_nodes
if self._is_mnist_colors:
m = self.mnist.__len__()
digits = []
colors = []
for i in range(n):
x = random.randint(m)
digit, color = self.mnist.__getitem__(x)
digits.append(np.array(digit)[np.newaxis])
colors.append(color)
digits, colors = np.array(digits), np.array(colors)
else:
colors = random.randint(self._nr_colors, size=n)
states = np.zeros((n, self._nr_colors))
adjacent = np.zeros((n, self._nr_colors))
# The goal is to predict whether there is a node with desired color
# as adjacent node for each node x.
for i in range(n):
states[i, colors[i]] = 1
adjacent[i, colors[i]] = 1
for j in range(n):
if graph.has_edge(i, j):
adjacent[i, colors[j]] = 1
if self._is_mnist_colors:
states = digits
return dict(
n=n,
relations=np.expand_dims(graph.get_edges(), axis=-1),
states=states,
colors=colors,
target=adjacent,
)
def _restart(self):
self.start_world = randomly_generate_world(self.nr_blocks,
random_order=False)
self.final_world = randomly_generate_world(self.nr_blocks,
random_order=False)
self.world = self.start_world
if self.random_order:
n = self.world.size
# Ground is fixed as index 0 if fix_ground is True
ground_ind = 0 if self.fix_ground else random.randint(n)
def get_order():
raw_order = random.permutation(n - 1)
order = []
for i in range(n - 1):
if i == ground_ind:
order.append(0)
order.append(raw_order[i] + 1)
if ground_ind == n - 1:
order.append(0)
return order
self.start_world.blocks.set_random_order(get_order())
self.final_world.blocks.set_random_order(get_order())
self._prepare_worlds()
self.start_state = decorate(self._get_coordinates(self.start_world),
self.nr_objects, 0)
self.final_state = decorate(self._get_coordinates(self.final_world),
self.nr_objects, 1)
self.is_over = False
self.cached_result = self._get_result()
def _gen_data_filter_exist(self, record):
if self.balance_attribute:
attr = random.choice_list(gdef.all_attributes)
if self.balance_answer:
answer = bool(random.randint(0, 2))
if answer:
concept = record['object'][attr]
else:
rest_concepts = gdef.attribute_concepts[attr].copy()
rest_concepts.remove(record['object'][attr])
concept = random.choice_list(rest_concepts)
else:
concept = random.choice_list(gdef.attribute_concepts[attr])
answer = record['object'][attr] == concept
else:
# TODO(Jiayuan Mao @ 04/10): implement.
raise NotImplementedError(
'Currently not supporting balance_attribute = False')
if attr == 'shape':
question = 'any ' + concept
else:
question = 'any ' + concept + ' object'
program = [
dict(op='scene', inputs=[]),
dict(op='filter', concept=[concept], inputs=[0]),
dict(op='exist', inputs=[1]),
]
return question, 'yes' if answer else 'no', program
def _gen(self):
"""Sample the starting node and the destination according to the distance."""
dist_matrix = self._graph.get_shortest()
st, ed = np.where(dist_matrix == self.dist)
if len(st) == 0:
return None
ind = random.randint(len(st))
return st[ind], ed[ind]
def _sample(self, data, num, label):
"""Sample a batch of size $num from the data, with already determined label."""
# assert num <= len(data)
states, actions = [], []
for _ in range(num):
ind = random.randint(len(data))
state, action = data[ind]
states.append(state)
actions.append([action])
return np.array(states), np.array(actions), np.ones((num, )) * label
def randomly_generate_graph_dnc(n, p=None, directed=False):
"""Random graph generation method as in DNC.
As described in Differentiable neural computers (DNC),
(https://www.nature.com/articles/nature20101.epdf?author_access_token=ImTXBI8aWbYxYQ51Plys8NRgN0jAjWel9jnR3ZoTv0MggmpDmwljGswxVdeocYSurJ3hxupzWuRNeGvvXnoO8o4jTJcnAyhGuZzXJ1GEaD-Z7E6X_a9R-xqJ9TfJWBqz)
Sample $n nodes in a unit square. Then sample out-degree (m) of each nodes,
connect to $m nearest neighbors (Euclidean distance) in the unit square.
Args:
n: The number of nodes in the graph.
p: Control the sampling of the out-degree.
If p=None, the default range is [1, n // 3].
If p is float, the range is [1, int(n * p)].
If p is int, the range is [1, p].
If p is tuple. the range is [p[0], p[1]].
directed: Directed or Undirected graph. Default: False (undirected)
Returns:
A Graph class representing randomly generated graph.
"""
edges = np.zeros((n, n), dtype='float')
pos = random.rand(n, 2)
def dist(x, y):
return ((x - y)**2).mean()
if isinstance(p, tuple):
lower, upper = p
else:
lower = 1
if p is None:
upper = n // 3
elif isinstance(p, int):
upper = p
elif isinstance(p, float):
upper = int(n * p)
else:
assert False, 'Unknown argument type: {}'.format(type(p))
upper = max(upper, 1)
lower = max(lower, 1)
upper = min(upper, n - 1)
for i in range(n):
d = []
k = random.randint(upper - lower + 1) + lower
for j in range(n):
if i != j:
d.append((dist(pos[i], pos[j]), j))
d.sort()
for j in range(k):
edges[i, d[j][1]] = 1
if not directed:
edges = np.maximum(edges, edges.T)
return Graph(n, edges, pos)
def random_size_crop(img,
target_shape,
area_range,
aspect_ratio=None,
contiguous_ar=False,
*,
nr_trial=10):
"""random size crop used for Facebook ImageNet data augmentation
see https://github.com/facebook/fb.resnet.torch/blob/master/datasets/imagenet.lua
"""
target_shape = get_2dshape(target_shape)
h, w = img.shape[:2]
area = h * w
area_range = area_range if isinstance(
area, collections.Iterable) else (area_range, 1)
if aspect_ratio is None:
assert contiguous_ar == False
aspect_ratio = [h / w]
for i in range(nr_trial):
target_area = random.uniform(area_range[0], area_range[1]) * area
target_ar = random.choice(aspect_ratio)
nw = int(round((target_area * target_ar)**0.5))
nh = int(round((target_area / target_ar)**0.5))
if random.rand() < 0.5:
nh, nw = nw, nh
if nh <= h and nw <= w:
sx, sy = random.randint(w - nw + 1), random.randint(h - nh + 1)
img = img[sy:sy + nh, sx:sx + nw]
return imgproc.resize(img, target_shape)
scale = min(*target_shape) / min(h, w)
return imgproc.center_crop(imgproc.resize_scale(img, scale), target_shape)
def test(index, all_objs, all_preds, meter):
obj = all_objs[index]
nr_descriptors = random.randint(1, 3)
desc = random.choice_list(get_desc(obj), size=nr_descriptors)
if isinstance(desc, six.string_types):
desc = [desc]
filtered_objs = np.array(
[i for i, o in enumerate(all_objs) if not run_desc_obj(o, desc)],
dtype=int)
all_scores = run_desc_pred(all_preds, desc)
rank = (all_scores[filtered_objs] > all_scores[index]).sum()
meter.update('r@01', rank <= 1)
meter.update('r@02', rank <= 2)
meter.update('r@03', rank <= 3)
meter.update('r@04', rank <= 4)
meter.update('r@05', rank <= 5)
def test(index, all_objs, all_preds, meter):
obj = all_objs[index]
nr_descriptors = random.randint(2, 5)
desc = random.choice_list(get_desc(obj), size=nr_descriptors)
filtered_objs = [
i for i, o in enumerate(all_objs) if not run_desc_obj(o, desc)
]
all_scores = run_desc_pred(all_preds, desc)
rank = (all_scores[filtered_objs] > all_scores[index]).sum()
# print(desc)
# print(all_scores)
# print(all_scores[index])
meter.update('r@01', rank <= 1)
meter.update('r@05', rank <= 5)
meter.update('r@10', rank <= 10)
meter.update('r@50', rank <= 50)
def _sample_dist(self):
"""Sample the distance between the starting node and the destination."""
lower, upper = self._dist_range
upper = min(upper, self._nr_nodes - 1)
return random.randint(upper - lower + 1) + lower
def randomly_generate_family(n, p_marriage=0.8, verbose=False):
"""Randomly generate family trees.
Mimic the process of families growing using a timeline. Each time a new person
is created, randomly sample the gender and parents (could be none, indicating
not included in the family tree) of the person. Also maintain lists of singles
of each gender. With probability $p_marrige, randomly pick two from each list
to be married. Finally randomly permute the order of people.
Args:
n: The number of people in the family tree.
p_marriage: The probability of marriage happens each time.
verbose: print the marriage and child born process if verbose=True.
Returns:
A family tree instance of $n people.
"""
assert n > 0
ids = list(random.permutation(n))
single_m = []
single_w = []
couples = [None]
# The relations are: husband, wife, father, mother, son, daughter
rel = np.zeros((n, n, 6))
fathers = [None for i in range(n)]
mothers = [None for i in range(n)]
def add_couple(man, woman):
"""Add a couple relation among (man, woman)."""
couples.append((man, woman))
rel[woman, man, 0] = 1 # husband
rel[man, woman, 1] = 1 # wife
if verbose:
print('couple', man, woman)
def add_child(parents, child, gender):
"""Add a child relation between parents and the child according to gender."""
father, mother = parents
fathers[child] = father
mothers[child] = mother
rel[child, father, 2] = 1 # father
rel[child, mother, 3] = 1 # mother
if gender == 0: # son
rel[father, child, 4] = 1
rel[mother, child, 4] = 1
else: # daughter
rel[father, child, 5] = 1
rel[mother, child, 5] = 1
if verbose:
print('child', father, mother, child, gender)
def check_relations(man, woman):
"""Disable marriage between cousins."""
if fathers[man] is None or fathers[woman] is None:
return True
if fathers[man] == fathers[woman]:
return False
def same_parent(x, y):
return fathers[x] is not None and fathers[
y] is not None and fathers[x] == fathers[y]
for x in [fathers[man], mothers[man]]:
for y in [fathers[woman], mothers[woman]]:
if same_parent(man, y) or same_parent(woman, x) or same_parent(
x, y):
return False
return True
while ids:
x = ids.pop()
gender = random.randint(2)
parents = random.choice(couples)
if gender == 0:
single_m.append(x)
else:
single_w.append(x)
if parents is not None:
add_child(parents, x, gender)
if random.rand() < p_marriage and len(single_m) > 0 and len(
single_w) > 0:
mi = random.randint(len(single_m))
wi = random.randint(len(single_w))
man = single_m[mi]
woman = single_w[wi]
if check_relations(man, woman):
add_couple(man, woman)
del single_m[mi]
del single_w[wi]
return Family(n, rel)