In compute_madds.py, for compute_Linear_madd, the number of additions computed is computed this way :
add = num_in_features - 1.
I guess it comes from the additions of the dot product between the input vector and the weights vector. But if the bias parameter of the torch.nn.Linear layer is True, an additional addition will be performed with that bias.So the correct formula should be
add = num_in_features, in case there is a bias.
There is the same problem with compute_Linear_flops in compute_flops,
inp.size()[1] * out.size()[1]
should be replaced by
(inp.size()[1]+1) * out.size()[1], if bias is True.
In
compute_madds.py, forcompute_Linear_madd, the number of additions computed is computed this way :add = num_in_features - 1.I guess it comes from the additions of the dot product between the input vector and the weights vector. But if the
biasparameter of thetorch.nn.Linearlayer isTrue, an additional addition will be performed with that bias.So the correct formula should beadd = num_in_features, in case there is a bias.There is the same problem with
compute_Linear_flopsincompute_flops,inp.size()[1] * out.size()[1]should be replaced by(inp.size()[1]+1) * out.size()[1], ifbiasisTrue.