How to predict the gs at next step in the eval stage？

[DingChunQ]

I want to get the future scene prediction in eval stage.
A key problem is that I need data['action'] as a part of dyna_input in the GeneralizableGSEmbedNet class's forward() function.
data['action'] is got from replay_sample['action'] when training, the shape is [bs, 8], e.g. "tensor([[ 2.8866e-01, -1.4778e-01, 1.2483e+00, 6.0417e-01, 7.9686e-01, -1.7191e-04, -1.3396e-04, 0.0000e+00]], device='cuda:1')"

How can I get the action like this in the eval stage?
Thanks!