def get_base_dataset(num_examples=100, **kwargs):
# return tf.data.Dataset.from_tensor_slices(
# get_base_data(num_exmaples=num_examples, **kwargs))
from deep_cloud.problems.modelnet import ModelnetProblem
from deep_cloud.problems.builders import pointnet_builder
problem = ModelnetProblem(builder=pointnet_builder(2), positions_only=False)
with gin.config_scope('train'):
dataset = problem.get_base_dataset('validation').map(augment_cloud)
return dataset
def plot_ks():
builder = pointnet_builder(pointnet_version=2)
means = []
maxs = []
radii = np.linspace(2, 10, 51)
total = 100
for features, _ in tqdm(tfds.as_numpy(
builder.as_dataset(split='train', as_supervised=True).take(total)),
total=total):
ks = get_ks(features['positions'], radii)
means.append(np.mean(ks, axis=1))
maxs.append(np.max(ks, axis=-1))
ax = plt.gca()
ax.plot(radii, np.mean(means, axis=0))
ax.plot(radii, np.mean(maxs, axis=0))
ax.legend(['mean', 'max'])
ax.set_yscale('log')
ax.set_xscale('log')
plt.show()
# print('done')
# exit()
# coords = np.stack([base_coords] * 2, axis=0)
# coords[1, :, 2] = 1
# normals = np.stack([base_normals] * 2, axis=0)
# dataset = tf.data.Dataset.from_tensor_slices((dict(positions=coords,
# normals=normals), [0,
# 0]))
from deep_cloud.problems.modelnet import ModelnetProblem
from deep_cloud.problems.builders import pointnet_builder
import tensorflow_datasets as tfds
from time import time
from tqdm import tqdm
problem = ModelnetProblem(builder=pointnet_builder(2),
positions_only=False)
dataset = problem.get_base_dataset(split='validation')
num_examples = 100
batch_size = 2
def profile(edge_fn, depth=4, num_examples=10):
times = []
for example, _ in tqdm(tfds.as_numpy(dataset.take(num_examples)),
total=num_examples):
start = time()
edge_fn(example['positions'], example['normals'], depth=depth)
times.append(time() - start)
return np.array(times)
# t_base = profile(compute_edges_eager, num_examples=num_examples)
from __future__ import absolute_import from __future__ import division from __future__ import print_function import numpy as np import tensorflow_datasets as tfds from deep_cloud.problems.builders import pointnet_builder from deep_cloud.problems.modelnet import ModelnetProblem from scipy.spatial import cKDTree # pylint: disable=no-name-in-module import tqdm import trimesh builder = pointnet_builder(2, uniform_density=False) split = 'train' problem = ModelnetProblem(builder, positions_only=False) # dataset = builder.as_dataset(split=split, as_supervised=True) dataset = problem.get_base_dataset(split=split) # all_labels = [] # for data, label in tfds.as_numpy(dataset): # all_labels.append(label) # labels = np.array(all_labels) # minval = np.min(labels) # maxval = np.max(labels) # print(minval, maxval) # import matplotlib.pyplot as plt # plt.hist(labels, normed=True, bins=40) # plt.show() k = 10
def profile_multi(radius=4, leafsize=4):
builder = pointnet_builder(pointnet_version=2)
for features, _ in tfds.as_numpy(
builder.as_dataset(split='train', as_supervised=True)):
profile(features['positions'], radius=radius, leafsize=leafsize)
def multi_full_run(num_runs=10):
builder = pointnet_builder(pointnet_version=2)
for features, _ in tfds.as_numpy(
builder.as_dataset(split='train',
as_supervised=True).take(num_runs)):
full_run(features['positions'])
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import tensorflow_datasets as tfds
from deep_cloud.problems.builders import pointnet_builder
from scipy.spatial import cKDTree as KDTree # pylint: disable=no-name-in-module
import matplotlib.pyplot as plt
builder = pointnet_builder(pointnet_version=2)
PACKING_RATIOS = {
2: np.pi * np.sqrt(3) / 6, # https://en.wikipedia.org/wiki/Circle_packing
3:
np.pi / (3 * np.sqrt(2)
) # https://en.wikipedia.org/wiki/Close-packing_of_equal_spheres
}
def approx_max_neighbors(radius, num_dims):
return PACKING_RATIOS[num_dims] * (1 + radius)**num_dims
def closest_neighbors(tree):
if isinstance(tree, np.ndarray):
tree = KDTree(tree)
distances, indices = tree.query(tree.data, 2)
del indices
return np.min(distances[:, 1])
import tensorflow as tf
from deep_cloud.problems.modelnet import ModelnetProblem
from deep_cloud.problems.builders import pointnet_builder
from deep_cloud.models.pointnet3 import multi_scale_group
from deep_cloud.models.pointnet3 import pre_batch_map, post_batch_map
from deep_cloud.ops.np_utils import tree_utils
from deep_cloud.augment import augment_cloud
import functools
from tqdm import tqdm
from time import time
tf.compat.v1.enable_eager_execution()
N = 500
batch_size = 16
problem = ModelnetProblem(builder=pointnet_builder(2), positions_only=False)
map_fn = functools.partial(augment_cloud,
angle_stddev=0.06,
angle_clip=0.18,
uniform_scale_range=(0.8, 1.25),
rotate_scheme='random',
jitter_stddev=1e-2,
jitter_clip=5e-2)
# map2 = pre_batch_map
map2 = functools.partial(pre_batch_map,
radii_lists=((0.1, ), (0.2, )),
max_neighbors_lists=((16, ), (32, )))
with tf.device('/cpu:0'):