[BUG] can't work with cpu

[LXYTSOS]

In references/detection/utils.py line 212: memory=torch.cuda.max_memory_allocated() / MB)). It works fine under the GPU environment, then I train a model with pytorch-cpu, exception occurs: AssertionError: Torch not compiled with CUDA enabled. When I trace the error I found the code in references/detection/utils.py: memory=torch.cuda.max_memory_allocated() / MB)), when you run torch.cuda.max_memory_allocated() in a CPU version it will gives you that error.

[pmeier]

Fixed by @LXYTSOS in #1023.

[fmassa]

Thanks for the fix @LXYTSOS