Yet-another-efficientdet-pytorch: insufficient shared memory

[HaolyShiit]

Hello, thanks for your great work!
Whatever I change the 'num_workers' or 'batch_size', error occurs as follows:
ERROR: Unexpected bus error encountered in worker. This might be caused by insufficient shared memory (shm).
RuntimeError: DataLoader worker (pid 6230) is killed by signal: Bus error.

I add a line to the code to specify the GPU number:
os.environ['CUDA_VISIBLE_DEVICES'] = '1'

How to solve the problem? Thank you.

[zylo117]

You can't solve it without adding more RAM to your server or reducing num_workers or batchsize

[HaolyShiit]

You can't solve it without adding more RAM to your server or reducing num_workers or batchsize

Thans for your advice.
I tried to set 'batchsize = 4' and 'num_workers = 1', and it worked.
The GPU memory is about 12G, and it only takes 1157 M for training. When I add 'batchsize ', same error occurs.
Is that normal?

[HaolyShiit]

You can't solve it without adding more RAM to your server or reducing num_workers or batchsize

Hello, maybe I solved the problem. I reduce 'num_workers' to 0, and 'batchsize' can be added to 32 or more.
Thans for your great work again!

[merisgong208]

I think there maybe a bug here. For a 6GB gpu can only load one num_workers and batchsize 4 ? It eats too much.

[zylo117]

num workers have nothing to do with gpu mem, batchsize does. But using larger batchsize or num_workers, it consumes more system memory, which in my experience, lots of servers are lack of.