The parameter 'validate' is not work. about mmdetection HOT 4 CLOSED

We haven't implemented the evaluation hook for non-distributed training and will add it later. Currently you can use distributed training, which is strongly recommended.

from mmdetection.

Is there any tutorials to use distributed training? My lab has two machines, both of them have 4 GPUs.
Thank you! @hellock

from mmdetection.

Distributed training does not require multiple machines. If your two machines are not connected with an InfiniBand, it will be slow. In fact, we run all the experiments with 8 GPUs on a single machine.

To run on a single machine with distributed training, please refer to the readme.

If you do want to run on multiple machines, you can modify the commands in tools/dist_train.sh according to the guide.

from mmdetection.

OK I will try this. Thanks a lot @hellock .

from mmdetection.