1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
|
# ----------------------------------------------------------------------------
# - Open3D: www.open3d.org -
# ----------------------------------------------------------------------------
# The MIT License (MIT)
#
# Copyright (c) 2018-2021 www.open3d.org
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
# IN THE SOFTWARE.
# ----------------------------------------------------------------------------
"""This test simulates a network training by running some ops in graph mode for
tensorflow to catch a bug observed when linking the open3d main lib.
The error is not deterministic. The most frequent message is:
2020-11-21 23:07:56.653976: E tensorflow/stream_executor/cuda/cuda_event.cc:29] Error polling for event status: failed to query event: CUDA_ERROR_LAUNCH_FAILED: unspecified launch failure
2020-11-21 23:07:56.654028: F tensorflow/core/common_runtime/gpu/gpu_event_mgr.cc:220] Unexpected Event status: 1
followed by segfault.
We found that the bug can be reproduced with 125458ad when linking the main lib
and using cmake 3.13.2. The child commit 14c4815d does not show the problem when
linking the main lib. For cmake 3.18.2 and cmake 3.19 we cannot reproduce the
bug with 125458ad . Further the diff between both commits does not show changes
related to the problem. Since we know that the problem can be resolved by using
cmake >= 3.18.2, we think that the way cmake generates the link command may
cause the problem.
Some more info about the systems on which the problem was discovered:
Python 3.7.4
Tensorflow 2.3.0
CUDA 10.1 and 10.2
CMake 3.13.2 and 3.12.4
"""
import open3d as o3d
import numpy as np
np.set_printoptions(linewidth=600)
np.set_printoptions(threshold=np.inf)
import pytest
import mltest
@mltest.parametrize.ml
def test_training_graph_mode(ml):
# the problem is specific to tensorflow
if ml.module.__name__ != 'tensorflow':
return
# the problem is specific to CUDA
if not 'GPU' in ml.device:
return
tf = ml.module
rng = np.random.RandomState(123)
from particle_network_tf import MyParticleNetwork
model = MyParticleNetwork()
optimizer = tf.keras.optimizers.Adam(learning_rate=0.001, epsilon=1e-6)
batch_size = 16
def euclidean_distance(a, b, epsilon=1e-9):
return tf.sqrt(tf.reduce_sum((a - b)**2, axis=-1) + epsilon)
def loss_fn(pr_pos, gt_pos, num_fluid_neighbors):
gamma = 0.5
neighbor_scale = 1 / 40
importance = tf.exp(-neighbor_scale * num_fluid_neighbors)
return tf.reduce_mean(importance *
euclidean_distance(pr_pos, gt_pos)**gamma)
@tf.function(experimental_relax_shapes=True)
def train(model, batch):
with tf.GradientTape() as tape:
losses = []
for batch_i in range(batch_size):
inputs = ([
batch[batch_i]['pos0'], batch[batch_i]['vel0'], None,
batch[batch_i]['box'], batch[batch_i]['box_normals']
])
pr_pos1, pr_vel1 = model(inputs)
l = 0.5 * loss_fn(pr_pos1, batch[batch_i]['pos1'],
model.num_fluid_neighbors)
inputs = (pr_pos1, pr_vel1, None, batch[batch_i]['box'],
batch[batch_i]['box_normals'])
pr_pos2, pr_vel2 = model(inputs)
l += 0.5 * loss_fn(pr_pos2, batch[batch_i]['pos2'],
model.num_fluid_neighbors)
losses.append(l)
losses.extend(model.losses)
total_loss = 128 * tf.add_n(losses) / batch_size
grads = tape.gradient(total_loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
return total_loss
for iteration in range(100):
batch = []
for batch_i in range(batch_size):
num_particles = rng.randint(1000, 2000)
num_box_particles = rng.randint(3000, 4000)
batch.append({
'pos0':
tf.convert_to_tensor(
rng.uniform(size=(num_particles, 3)).astype(np.float32)
),
'vel0':
tf.convert_to_tensor(
rng.uniform(size=(num_particles, 3)).astype(np.float32)
),
'pos1':
tf.convert_to_tensor(
rng.uniform(size=(num_particles, 3)).astype(np.float32)
),
'pos2':
tf.convert_to_tensor(
rng.uniform(size=(num_particles, 3)).astype(np.float32)
),
'box':
tf.convert_to_tensor(
rng.uniform(size=(num_box_particles,
3)).astype(np.float32)),
'box_normals':
tf.convert_to_tensor(
rng.uniform(size=(num_box_particles,
3)).astype(np.float32)),
})
current_loss = train(model, batch)
assert (True) # The test is successful if this line is reached
|
