e15bc113b83c2bdf980d2557971888d10011ef29,neuralstyle/transformernet.py,InstanceNormalization,forward,#InstanceNormalization#Any#,117

Before Change

        t = x.resize(x.size()[0], x.size()[1], 1, n)
        mean = torch.mean(t, 3).repeat(1, 1, x.size()[2], x.size()[3])
        // Calculate the biased var. torch.var returns unbiased var
        var = torch.var(t, 3).repeat(1, 1, x.size()[2], x.size()[3]) * ((n - 1) / float(n))
        res = (x - mean) / torch.sqrt(var + 1e-5)
        // TODO: Check if you need to add scaling and shifting here
        return res

After Change

        mean = torch.mean(t, 3).expand_as(x)
        // Calculate the biased var. torch.var returns unbiased var
        var = torch.var(t, 3).expand_as(x)
        scale_broadcast = self.scale.unsqueeze(1).unsqueeze(1).unsqueeze(0)
        scale_broadcast = scale_broadcast.expand_as(x)
        shift_broadcast = self.shift.unsqueeze(1).unsqueeze(1).unsqueeze(0)
        shift_broadcast = shift_broadcast.expand_as(x)
        out = (x - mean) / torch.sqrt(var + 1e-5)
        out = (out * scale_broadcast) + shift_broadcast
        return out

Instances