- 博客(10)
- 收藏
- 关注
RSS订阅原创 pytorch dataloader pin_memory persistent_workers prefetch multiprocessing dataloader
如果模型特别大，GPU显存不够，无法将一个显存放在GPU上，需要把网络的不同模块放在不同GPU上，这样可以训练比较大的网络。：将整个模型放在一块GPU里，再复制到每一块GPU上，同时进行正向传播和反向误差传播。相当于加大了batch_size。
2024-10-28 00:41:54
1108
1
原创 20240127_large synthtext_training 1st_work_on_2_A100_4u4g-0036.txt prefetch_factor
2 ls1. Above4. bs = 4==>Tips:
2024-10-27 18:15:22
961
原创 Change model name pth lightning
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。原文链接：https://blog.csdn.net/zzfive/article/details/125726629。
2024-08-22 16:50:09
206
原创 windows auto repair file explorer windows long time to show file explorer and browser , disable 360
https://www.tenforums.com/tutorials/2376-create-bootable-usb-flash-drive-install-windows-10-a.html https://vietbay.com.vn/en/5-ways-to-fix-windows-file-explorer-search-when-its-slow.htmRedirecting How to Fix Windows File Explorer Slow? How to Fix: File Ex
2024-08-08 09:58:11
266
原创 Failed building wheel for mpi4py
Detail: mpi4py installation failed with Python 3.9 and 3.10 · Issue #335 · mpi4py/mpi4py · GitHubSolution: $ conda install mpi4py
2023-11-02 15:41:33
198
原创 uninstall install update cuda from 11.4 to 11.8 in a tensorrt docker
3 df -h12 df -h13 ls18 ls。
2023-09-20 17:17:31
304
1
原创 trtexec fps batch (multiply qps by the batch size)
https://github.com/NVIDIA/TensorRT/issues/1593You need to multiply qps by the batch size. You should also look at the GPU compute time, which should be equivalent to qps if you do (1000/gpu_compute_time(ms)).Inference time is not improving with the increas
2023-08-29 10:19:25
253
1
空空如也
TA创建的收藏夹 TA关注的收藏夹
TA关注的人