多GPU训练Yolov3

[saigequn]

两个GPU训练Yolov3

yolov3.yml修改

gpus: "0,1"
mini_batch_size:8

GPU使用情况如下

Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla P100-PCIE...  Off  | 00000000:05:00.0 Off |                    0 |
| N/A   58C    P0   237W / 250W |   8835MiB / 16160MiB |     100%      Default|
+-------------------------------+----------------------+----------------------+
|   1  Tesla P100-PCIE...  Off  | 00000000:89:00.0 Off |                    0 |
| N/A   36C    P0    25W / 250W |     11MiB / 16160MiB |      0%      Default 

其中一个GPU在闲置

[mileistone]

We don't support multi gpus now. It's easy, you can implement it in 10 minutes.

[tonysy]

Could you share how to revise the code to achieve multi-gpu train? I found it difficult to implement.

[yutaizhou]

Can you share a guide on multi gpu implementation? Or give some good resources?

[gongh4]

First, update the cfg set gpus:"0,1"
Then, in _voc_train.py, insert the code after
if self.cuda:
net.cuda()
as
net.net = torch.nn.DataParallel(net.net)

Then, you can train using multi-gpu