object_detector default handler does not seem to be using multiple GPUs

[fbbradheintz]

The other default handlers seem to work on many GPUs without difficulty, but CloudWatch is telling me that only GPU 0 is being used by object_detector.

[harshbafna]

@fbbradheintz : This is a issue with the current stable release of torchvision and we already have #104 open for tracking this.

To keep the default handler working, we have deliberately kept the GPU device to cuda:0.

[mycpuorg]

@harshbafna thanks for confirming.

@fbbradheintz I'm removing the launch blocker tag.

[lokeshgupta1975]

@mycpuorg You mentioned in your previous comment that you will remove the launch_blocker tag, but you did not. Should we?

[fbbradheintz]

This is not, in fact, a problem with TorchVision, but a bug in object_detector.py. Putting up a fix shortly.

[fbbradheintz]

Covered by @harshbafna in #265, requested changes pending.

[mycpuorg]

Closing this because #265

[mycpuorg]

Reopening because the changes have not made it to master yet