Pytorch-yolov3: multiple gpus?

[aliborji]

Is it possible to train the model (this code) on multiple gpus?

[wilkice]

Not support. You can change the code to support. reference: PyTorch MULTI-GPU EXAMPLES

[kaikaizhu]

@aliborji ,@wilkice
I want to use multi-GPU training. Did you add it successfully? I made a mistake when I added it.

RuntimeError: CUDA error: device-side assert triggered.

[H-YunHui]

@kaikaizhu
I have the same question, and do you solve it now?

[MichaelCong]

I have the same question, and do you solve it now?

[YIYIZH]

@aliborji ,@wilkice
I want to use multi-GPU training. Did you add it successfully? I made a mistake when I added it.

RuntimeError: CUDA error: device-side assert triggered.

I met the same error, and I think the problem comes from here.
loss, outputs = model(imgs, targets)
while i use 4 gpus, I found the imgs and targets are divided equally into 4 gpus. The problem is that the targets should not be equally splitted since each img have different number of targets(bboxes).

One of the solution is to do like this:
outputs = model(imgs)
loss = compute_loss(outputs, targets) # you have to write a new function to compute the loss

Another is to find the size of the longest targets of imgs in one batch, and pad zeros to each target into the same length. such that targets can correspond to the imgs.

If anyone has better solution, pls tell me. Thx.

[GangHu1993]

I have the same question,

@aliborji ,@wilkice
I want to use multi-GPU training. Did you add it successfully? I made a mistake when I added it.
RuntimeError: CUDA error: device-side assert triggered.

I met the same error, and I think the problem comes from here.
loss, outputs = model(imgs, targets)
while i use 4 gpus, I found the imgs and targets are divided equally into 4 gpus. The problem is that the targets should not be equally splitted since each img have different number of targets(bboxes).

One of the solution is to do like this:
outputs = model(imgs)
loss = compute_loss(outputs, targets) # you have to write a new function to compute the loss

Another is to find the size of the longest targets of imgs in one batch, and pad zeros to each target into the same length. such that targets can correspond to the imgs.

If anyone has better solution, pls tell me. Thx.

I have same probliem, do you have solved this ?

[yangzhaonan18]

多GPU加载模型和数据官网有教程，加载权重怎么弄，求个学习链接，谢谢
Multi-GPU loading model and data official website have tutorials, how to load weights, ask for a learning link, thank you

[ywatanabe1989]

https://github.com/ywatanabe1989/PyTorch-gaussian-YOLOv3-1D/
Although there might be mistakes (at least, the print log output is broken now.) and it's not sophisticated at all, I tried to implement the multi-GPU functioning.

Please give me some feedback. :)

[ujsyehao]

@YIYIZH Hi, I solve the problem using your proposed method(Another is to find the size of the longest targets of imgs in one batch, and pad zeros to each target into the same length).

[ujsyehao]

You can refer to https://github.com/ujsyehao/yolov3-multigpu

[sakurasakura1996]

@aliborji ,@wilkice
I want to use multi-GPU training. Did you add it successfully? I made a mistake when I added it.
RuntimeError: CUDA error: device-side assert triggered.

I met the same error, and I think the problem comes from here.
loss, outputs = model(imgs, targets)
while i use 4 gpus, I found the imgs and targets are divided equally into 4 gpus. The problem is that the targets should not be equally splitted since each img have different number of targets(bboxes).

One of the solution is to do like this:
outputs = model(imgs)
loss = compute_loss(outputs, targets) # you have to write a new function to compute the loss

Another is to find the size of the longest targets of imgs in one batch, and pad zeros to each target into the same length. such that targets can correspond to the imgs.

If anyone has better solution, pls tell me. Thx.
Did you solve the problem,I want to run the code on my custom data with one GPU,but I found I am training on CPU

[Flova]

Duplicate of #265 #507 #290