056338255ea968d28ab462295b308cf475adcdde,chainer_/models/dpn.py,DPN,__init__,#DPN#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#,358

Before Change

                self.features.add(stage)
            self.features.add(DPNFinalBlock(channels=in_channels))

            self.output = nn.HybridSequential(prefix="")
            if for_training or not test_time_pool:
                self.output.add(nn.GlobalAvgPool2D())
                self.output.add(conv1x1(
                    in_channels=in_channels,
                    out_channels=classes,
                    use_bias=True))
                self.output.add(nn.Flatten())
            else:
                self.output.add(nn.AvgPool2D(
                    pool_size=7,
                    strides=1))
                self.output.add(conv1x1(
                    in_channels=in_channels,
                    out_channels=classes,
                    use_bias=True))
                self.output.add(GlobalAvgMaxPool2D())
                self.output.add(nn.Flatten())

    def hybrid_forward(self, F, x):
        x = self.features(x)
        x = self.output(x)

After Change

                    setattr(self.features, "stage{}".format(i + 1), stage)
                setattr(self.features, "final_block", DPNFinalBlock(channels=in_channels))

            self.output = SimpleSequential()
            with self.output.init_scope():
                if for_training or not test_time_pool:
                    setattr(self.output, "final_pool", GlobalAvgPool2D())
                    setattr(self.output, "final_conv", conv1x1(
                        in_channels=in_channels,
                        out_channels=classes,
                        use_bias=True))
                    setattr(self.output, "final_flatten", partial(
                        F.reshape,
                        shape=(-1, classes)))
                else:
                    setattr(self.output, "avg_pool", partial(
                        F.average_pooling_2d,
                        ksize=7,
                        stride=1))
                    setattr(self.output, "final_conv", conv1x1(
                        in_channels=in_channels,
                        out_channels=classes,
                        use_bias=True))
                    setattr(self.output, "avgmax_pool", GlobalAvgMaxPool2D())
                    setattr(self.output, "final_flatten", partial(
                        F.reshape,
                        shape=(-1, classes)))

    def __call__(self, x):
        x = self.features(x)
        x = self.output(x)
        return x

Instances