Single GPU training error in DER

[htwang14]

Hi,

Thank you for this wonderful code base!

I noticed the current version doesn't support single GPU training. Could you please add this feature?

Thank you!

[G-U-N]

Would you mind sharing the error logs?
Ideally(i.e., without encountering bugs), change the original params in JSON files

"device": ["0","1","2","3"],

to

"device": ["0"]

, and you can run the algos with single GPU.

[htwang14]

I haven't tried other methods, but for DER, something like this might be necessary: htwang14@e15cbe5

[G-U-N]

Thanks, I have fixed the problem.
The error occurs because of the lack of judgment on whether self._network is an instance of module or DataParallel in function train(self). Here, I create a module pointer (i.e., the self._network_module_ptr) which always points to the module instance.

    def train(self):
        self._network.train()
        if len(self._multiple_gpus) > 1 :
            self._network_module_ptr = self._network.module
        else:
            self._network_module_ptr = self._network
        self._network_module_ptr.convnets[-1].train()
        if self._cur_task >= 1:
            for i in range(self._cur_task):
                self._network_module_ptr.convnets[i].eval()

Hope this can help you.