search

charlesq34 / frustum-pointnets

Frustum PointNets for 3D Object Detection from RGB-D Data
Apache License 2.0
1.57k stars 538 forks source link

Segmentation fault in Tensorflow version 1.14.0 #103

Closed witignite closed 4 years ago

witignite commented 4 years ago

Type: Question OS: Ubuntu 16.04 TensorFlow: tensorflow-gpu-1.14.0, installed with pip CUDA: 10.0

I have compiled the custom operators in /models/tf_ops/ folder successfully.

When I tried training with tensorflow-gpu-1.14.0 installed with pip, the custom TensorFlow operators didn't seem to work. I stepped through the code and it breaks at the point where the code tries to call the custom op sampling_module.farthest_point_sample(inp, npoint) here.

It results in a segmentation fault:

python train/train.py --gpu 0 --model frustum_pointnets_v2 --log_dir train/log_v2 --num_point 1024 --max_epoch 201 --batch_size 1 --decay_step 800000 --decay_rate 0.5
pid: 20105
Segmentation fault (core dumped)

However, when I tried with tensorflow-1.12.0, it seems to work fine.

I have completely no idea why it doesn't work with tensorflow-gpu-1.14.0, and I'm not very experienced with writing TensorFlow custom op.

Dear author, @charlesq34 , do you know what might be a possible reason? Or does anyone have a clue why it doesn't work with 1.14.0 version?

witignite commented 4 years ago

Probably it is because of mismatch compiler.

anjingde commented 3 years ago

When I eval the val ,I meet the same problem. How to mdeal with this problem?

Going to eval ground for class: pedestrian Segmentation fault (core dumped)