FF-Net implementation is inconsistent to the description in TABLE 1.

[songkq]

Hi, @MinZHANG-WHU
As for the conv_f layer of FF-Net, it is composed of Conv+ReLU in TABLE 1 in the paper.
However, FF-Net implementation didn't include the ReLU operation. Does it matter?

        n.concat_1 = self.concat(n.data_t12, n.fd_1, n.up_2, n.up_3)
        n.conv_t = self.conv(n.concat_1, 3, self.ff_channel, stride=1, pad=1,
                             name_w="conv_t_w", name_b="conv_t_b",
                             lr_mult_w=1, lr_mult_b=1,
                             decay_mult=1, bias_term=True)
        n.conv_prob = self.conv(n.conv_t, 1, 1, stride=1, pad=0,
                                name_w="conv_prob_w", name_b="conv_prob_b",
                                lr_mult_w=1, lr_mult_b=1,
                                decay_mult=1, bias_term=True)

        n.sig = L.Sigmoid(n.conv_prob, in_place=False)

[MinZHANG-WHU]

Sorry for the confusing. This is a careless mistake. Actually FF-Net implementation didn't include the ReLU layer.

[songkq]

OK.