What coefficient do you use? By the way, in random shifting #5, what padding value do you use?
class RandomShiftsAug(torch.nn.Module):
def __init__(self, pad):
super().__init__()
self.pad = pad # what's the pad value here?
def forward(self, x):
x = x.float()
b, t, c, h, w = x.size()
assert h == w
x = x.view(b*t, c, h, w) # reshape x to [B*T, C, H, W]
padding = tuple([self.pad] * 4)
...
Hi @StarCycle , the coefficients for arm loss, gripper loss, and video loss is 1, 0.01, 0.01. For the padding value, we use 10 for static rgbs and 4 for hand rgbs.
Hi @bdrhtw @hongtaowu67,
In the paper, it seems that the coefficient for every term is 1:
But when I train the policy, it seems that using a higher coefficient on L_arm improves performance.
My loss function looks like this. If it's wrong please let me know...
What coefficient do you use? By the way, in random shifting #5, what padding value do you use?