目录
报错信息:
解决方案:
整理不易,欢迎一键三连!!!
跑PyTorch 的深度学习的代码,之前跑没有问题,换了一台服务器,遇到这个bug报错:RuntimeError: can't start new thread,原因是测试的时候线程开得太多了,导致软件开始,不再能够被处理,卡死。
[ERROR:0] global /tmp/pip-req-build-cg56sov1/opencv/modules/core/src/parallel_impl.cpp (24
0) WorkerThread 158: Can't spawn new thread: res = 11
[ERROR:0] global /tmp/pip-req-build-cg56sov1/opencv/modules/core/src/parallel_impl.cpp (24
0) WorkerThread 120: Can't spawn new thread: res = 11
[ERROR:0] global /tmp/pip-req-build-cg56sov1/opencv/modules/core/src/parallel_impl.cpp (24
0) WorkerThread 152: Can't spawn new thread: res = 11
[ERROR:0] global /tmp/pip-req-build-cg56sov1/opencv/modules/core/src/parallel_impl.cpp (24
0) WorkerThread 159: Can't spawn new thread: res = 11
[ERROR:0] global /tmp/pip-req-build-cg56sov1/opencv/modules/core/src/parallel_impl.cpp (24
0) WorkerThread 160: Can't spawn new thread: res = 11
...
RuntimeError: can't start new thread
1.cpu线程数降低,可以尝试将num_work调小,即调小mmseg工程里的config文件中workers_per_gpu参数。
data = dict(
samples_per_gpu=4, #batch_size
workers_per_gpu=1, #num_work
train=dict(...)
)
2.数据增广albumentations库中所使用的基于opencv的各种数据变换占用了太多的cpu,因此需要调小opencv的线程数目,在训练最开始的时候加入以下两行。
import cv2
cv2.setNumThreads(1) # 0也可以
送你们一条美丽的--分割线--
⛵⛵⭐⭐