Round1 vs Round2 PyTorch Model Difference

[Majurski, Michael Paul (Fed)]

When moving from Round1 to Round2 dataset models we noticed that the Round2 models are saved as being on the CPU. They needed an explicit copy of the model weights to the GPU.

Loading the model with 'model = torch.load(filepath).cuda()' was enough to get the weights loaded correctly. 

This example trojan 'detector' correctly loads and inferences the Round2 models.

https://github.com/usnistgov/trojai-example

-----------------------------

Michael Majurski
Computer Vision Scientist

NIST | ITL | SSD | ISG

301-975-3942