why are the metrics divided by 1000 here?

[Serge-weihao]

MGMatting/code-base/utils/evaluate.py

Line 102 in 671678d

def comput_sad_loss(pred, target, trimap):

and can the code evaluates real-world portrait dataset?
https://github.com/yucornetto/MGMatting/blob/main/code-base/evaluation.py

[yucornetto]

Hi, the implementation of this function is adopted from GCA-Matting, which I suppose is based on the official Matlab evaluation code.

You can evaluate the real-world portrait dataset using the "Whole Image" score, or you can modify to support the detail map in the real-world portrait dataset (e.g., detail_map[detail_map>0] = 128). Sorry for being busy these days, I will also update the code to give better support later.

[Serge-weihao]

does evaluation on real-world portrait also use metrics divided by 1000?

[yucornetto]

Yes, same metrics are used in real-world portrait dataset

[Serge-weihao]

Are your Trimap computed on "unknown" and TrimapFG computed on "Whole Image"?

[yucornetto]

Hi, in this table. TrimapFG and Trimap refers to different types of inputs to our model. Since DIM dataset only provides trimap as guidance inputs, we use it to simulate a rough segmentation mask in two ways: TrimapFG: we only use the definite foreground region of the trimap; Trimap: we use a soft version of trrimap (i.e. trimap / 255.0). You can refer to our paper or code for more details :)

[Serge-weihao]

thanks for updating, i will solve them with your code. But one more question, does self._noupdate_u_v() updating your network weight during evaluation？

class SpectralNorm(nn.Module):
    """
    Based on https://github.com/heykeetae/Self-Attention-GAN/blob/master/spectral.py
    and add _noupdate_u_v() for evaluation
    """
    def __init__(self, module, name='weight', power_iterations=1):
        super(SpectralNorm, self).__init__()
        self.module = module
        self.name = name
        self.power_iterations = power_iterations
        if not self._made_params():
            self._make_params()

    def _update_u_v(self):
        u = getattr(self.module, self.name + "_u")
        v = getattr(self.module, self.name + "_v")
        w = getattr(self.module, self.name + "_bar")

        height = w.data.shape[0]
        for _ in range(self.power_iterations):
            v.data = l2normalize(torch.mv(torch.t(w.view(height,-1).data), u.data))
            u.data = l2normalize(torch.mv(w.view(height,-1).data, v.data))

        sigma = u.dot(w.view(height, -1).mv(v))
        setattr(self.module, self.name, w / sigma.expand_as(w))

    def _noupdate_u_v(self):
        u = getattr(self.module, self.name + "_u")
        v = getattr(self.module, self.name + "_v")
        w = getattr(self.module, self.name + "_bar")

        height = w.data.shape[0]
        sigma = u.dot(w.view(height, -1).mv(v))
        setattr(self.module, self.name, w / sigma.expand_as(w))

    def _made_params(self):
        try:
            u = getattr(self.module, self.name + "_u")
            v = getattr(self.module, self.name + "_v")
            w = getattr(self.module, self.name + "_bar")
            return True
        except AttributeError:
            return False

    def _make_params(self):
        w = getattr(self.module, self.name)

        height = w.data.shape[0]
        width = w.view(height, -1).data.shape[1]

        u = Parameter(w.data.new(height).normal_(0, 1), requires_grad=False)
        v = Parameter(w.data.new(width).normal_(0, 1), requires_grad=False)
        u.data = l2normalize(u.data)
        v.data = l2normalize(v.data)
        w_bar = Parameter(w.data)

        del self.module._parameters[self.name]

        self.module.register_parameter(self.name + "_u", u)
        self.module.register_parameter(self.name + "_v", v)
        self.module.register_parameter(self.name + "_bar", w_bar)

    def forward(self, *args):
        # if torch.is_grad_enabled() and self.module.training:
        if self.module.training:
            self._update_u_v()
        else:
            self._noupdate_u_v()
        return self.module.forward(*args)

[Serge-weihao]

thanks for updating, i will solve them with your code. But one more question, does self._noupdate_u_v() updating your network weight during evaluation？

class SpectralNorm(nn.Module):
    """
    Based on https://github.com/heykeetae/Self-Attention-GAN/blob/master/spectral.py
    and add _noupdate_u_v() for evaluation
    """
    def __init__(self, module, name='weight', power_iterations=1):
        super(SpectralNorm, self).__init__()
        self.module = module
        self.name = name
        self.power_iterations = power_iterations
        if not self._made_params():
            self._make_params()

    def _update_u_v(self):
        u = getattr(self.module, self.name + "_u")
        v = getattr(self.module, self.name + "_v")
        w = getattr(self.module, self.name + "_bar")

        height = w.data.shape[0]
        for _ in range(self.power_iterations):
            v.data = l2normalize(torch.mv(torch.t(w.view(height,-1).data), u.data))
            u.data = l2normalize(torch.mv(w.view(height,-1).data, v.data))

        sigma = u.dot(w.view(height, -1).mv(v))
        setattr(self.module, self.name, w / sigma.expand_as(w))

    def _noupdate_u_v(self):
        u = getattr(self.module, self.name + "_u")
        v = getattr(self.module, self.name + "_v")
        w = getattr(self.module, self.name + "_bar")

        height = w.data.shape[0]
        sigma = u.dot(w.view(height, -1).mv(v))
        setattr(self.module, self.name, w / sigma.expand_as(w))

    def _made_params(self):
        try:
            u = getattr(self.module, self.name + "_u")
            v = getattr(self.module, self.name + "_v")
            w = getattr(self.module, self.name + "_bar")
            return True
        except AttributeError:
            return False

    def _make_params(self):
        w = getattr(self.module, self.name)

        height = w.data.shape[0]
        width = w.view(height, -1).data.shape[1]

        u = Parameter(w.data.new(height).normal_(0, 1), requires_grad=False)
        v = Parameter(w.data.new(width).normal_(0, 1), requires_grad=False)
        u.data = l2normalize(u.data)
        v.data = l2normalize(v.data)
        w_bar = Parameter(w.data)

        del self.module._parameters[self.name]

        self.module.register_parameter(self.name + "_u", u)
        self.module.register_parameter(self.name + "_v", v)
        self.module.register_parameter(self.name + "_bar", w_bar)

    def forward(self, *args):
        # if torch.is_grad_enabled() and self.module.training:
        if self.module.training:
            self._update_u_v()
        else:
            self._noupdate_u_v()
        return self.module.forward(*args)

ok, it does not change the weight after eval forward