Moving a head in 4 different directions with convolutions and einsum

move_heads.py

import torch
import torch.nn.functional as F

movement_filters = torch.Tensor([
    [
        [0, 1, 0],
        [0, 0, 0],
        [0, 0, 0],
    ],
    [
        [0, 0, 0],
        [0, 0, 1],
        [0, 0, 0],
    ],
    [
        [0, 0, 0],
        [0, 0, 0],
        [0, 1, 0],
    ],
    [
        [0, 0, 0],
        [1, 0, 0],
        [0, 0, 0],
    ],
]).unsqueeze(1).float()

heads = torch.zeros((4, 1, 5, 5))
# Place heads in center of 4 5x5 environments
heads[:, 0, 2, 2] = 1

# Each moves in a different cardinal direction
actions_onehot = torch.zeros((4, 4))
actions_onehot[torch.arange(4), torch.arange(4)] = 1

intermediate = F.conv2d(heads, movement_filters, padding=1)

heads = torch.einsum('bchw,bc->bhw', [intermediate, actions_onehot]).unsqueeze(1)

print(heads[:, 0])