尝试加载 model 并启用 memory 增长时，Nvidia Xavier Jetson 出现 tensorflow 分段错误

Question

I have a segmentation fault with a very specific code sequence and only on Xavier Jetson:

import os
import requests
import tensorflow as tf
  
# 1    
print('SET MEMORY GROWTH')
physical_devices = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], True)  

# 2
print(f'REQUESTS GET')
requests.get('https://speed.hetzner.de/100MB.bin')

# 3
command = 'ls'
print(f'SYSTEM CALL ({command})')
os.system(command)

# 4 
print('MODEL LOAD') 
model = tf.keras.models.load_model('mnv2_xavier.h5')

If I remove one of these steps the code will run without issues. I don't know if some other code sequences can lead to this same behavior, but I am pretty sure that they exist.

I am trying to figure out what is the reason to have a segmentation fault here but, until now, I have no luck.

I think than can be something related with tensorflow memory growth policy and with the fact of Xavier Jetson having shared memory between CPU and GPU.

I would like to know if there is any way to solve this problem or a workaround and if someone have an explanation to this behavior.

Notes:

Code to create this model:

from tensorflow.keras.applications import MobileNetV2
from tensorflow.keras.models import Model
from tensorflow.keras import Input

x = Input((224,244,3))
y = MobileNetV2()(x)
model = Model(x,y)
model.save('mnv2_xavier.h5')

Versions:

Jetpack 4.4
tensorflow 2.3.0
keras 2.4.0
python 3.6.9

Output:

2021-04-15 16:51:22.031610: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.10.2
SET MEMORY GROWTH
2021-04-15 16:51:25.349940: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcuda.so.1
2021-04-15 16:51:25.374098: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:949] ARM64 does not support NUMA - returning NUMA node zero
2021-04-15 16:51:25.374309: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1716] Found device 0 with properties: 
pciBusID: 0000:00:00.0 name: Xavier computeCapability: 7.2
coreClock: 1.377GHz coreCount: 8 deviceMemorySize: 31.18GiB deviceMemoryBandwidth: 82.08GiB/s
2021-04-15 16:51:25.374437: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.10.2
2021-04-15 16:51:25.377470: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcublas.so.10
2021-04-15 16:51:25.379874: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcufft.so.10
2021-04-15 16:51:25.380541: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcurand.so.10
2021-04-15 16:51:25.383268: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcusolver.so.10
2021-04-15 16:51:25.385455: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcusparse.so.10
2021-04-15 16:51:25.385918: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudnn.so.8
2021-04-15 16:51:25.386201: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:949] ARM64 does not support NUMA - returning NUMA node zero
2021-04-15 16:51:25.386633: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:949] ARM64 does not support NUMA - returning NUMA node zero
2021-04-15 16:51:25.386723: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1858] Adding visible gpu devices: 0
REQUESTS GET
SYSTEM CALL (ls)
code          logs          logs2
bashc.sh      main-log.log  tests
Desktop       Documents     mnv2_xavier.h5
Downloads     model.py      Music
Videos        Pictures      go  
Public        segfault.py 
MODEL LOAD
2021-04-15 16:51:29.542399: W tensorflow/core/platform/profile_utils/cpu_utils.cc:108] Failed to find bogomips or clock in /proc/cpuinfo; cannot determine CPU frequency
2021-04-15 16:51:29.543521: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0xcbba840 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2021-04-15 16:51:29.543595: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
Segmentation fault (core dumped)

Answer 1

this error happens because the system is trying to use more memory than it should. When the system does not allow this, it gives a Segmentation Fault error. First, check the error file as follows.

$gdb python3
(gdb) run pythonfile.py

If the error is libapt-pkg5.0 install the appropriate package for your operating system For unix-based operating systems (Xaiver,Nano,TX2);

$sudo dpkg --purge --force-depends apt apt-utils libapt-inst2.0:arm64 libapt-pkg5.0:arm64

If the error is still not resolved;

$gedit ~/.bashrc

Adding;

export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libgomp.so.1