Question

[HuaZheLei]

Thanks for your nice work. I can not figure out 'MixSoftmaxCrossEntropyLoss' in your code, which is the default loss function. Could you explain it for me?

[Tramac]

Hi，the default loss function is cross entropy loss，MixSoftmaxCrossEntropyLoss means auxiliary loss （deep supervised）+ cross entropy loss.

[HuaZheLei]

Thanks for your reply. When I use multi-gpu to evaluate my trained model, I get an error.
Traceback (most recent call last): File "eval.py", line 108, in <module> evaluator = Evaluator(args) File "eval.py", line 49, in __init__ self.model = self.model.module File "/opt/conda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 539, in __getattr__ type(self).__name__, name)) AttributeError: 'FCN32s' object has no attribute 'module'
My conmand is
export NGPUS=4 python -m torch.distributed.launch --nproc_per_node=$NGPUS eval.py --model fcn32s --backbone vgg16 --dataset coco_voc

[roclark]

Hey @HuaZheLei, I'm running into this issue as well (for ENet specifically, but the same principle applies). Did you find a resolution?

[roclark]

Hello again @HuaZheLei, not sure if you are still looking at this or not, but I just created a pull request (#50) which now fully supports multi-GPU evaluation. I tested with the following command:

export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS eval.py --model enet --dataset citys

Feel free to give it a shot and let me know if it solves your issue!