Fix: Load a model for eval and test

[SinDongHwan]

Issue : When load a model for evaluating and testing by default, not get a device information from args.gpu_idx, but read it from the pre-train model(.pth). In a yolov4 case, the device inforamtion is written "cuda:2". So, if have one graphic card, cause error.

Fix : When call torch.load(), I used a parameter(map_location). map_location : a function, torch.device, string or a dict specifying how to remap storage locations.

[maudzung]

This is similar to the PR #18 Can you please resolve the conflict?