What does the label in dataset mean? about vin_pytorch_visdom HOT 4 CLOSED

@VoVAllen The labels correspond to the actions, generated by A*-planner, i.e. imitation learning.

from vin_pytorch_visdom.

Still confused. It seems one grid game has only one labeled action. What's the meaning of this one action?

from vin_pytorch_visdom.

@VoVAllen

For simplification assume the batch_size is 1, or in other words we only pass in a single map at a time:

The input then contains only a gridworld and the current state (S1, S2). The output of the VIN is then the best action to take (from the current state).

So yes, the output is only a single action because after the network predicts the action it should take from the given state, you would update the current state and feed the new state back in to get the next action.

from vin_pytorch_visdom.

@kentsommer Thanks for your detailed answer. It's helpful to me.

from vin_pytorch_visdom.