e27bd53c38aefc3e544944ec5b3a4ec340471c01,chainercv/links/model/faster_rcnn/faster_rcnn_vgg.py,FasterRCNNVGG16,__init__,#FasterRCNNVGG16#,79

Before Change

        else:
            vgg_kwargs = {}
        bbox_kwargs = {"initialW": chainer.initializers.Normal(0.001)}
        score_kwargs = {"initialW": chainer.initializers.Normal(0.01)}

        feature = VGG16FeatureExtractor(vgg_kwargs)
        rpn = RegionProposalNetwork(
            512, 512,
            ratios=ratios,
            anchor_scales=anchor_scales,
            feat_stride=self.feat_stride,
            proposal_creator_params=proposal_creator_params,
        )
        head = VGG16RoIPoolingHead(
            n_class,
            roi_size=7, spatial_scale=1. / self.feat_stride,
            vgg_kwargs=vgg_kwargs,

After Change

        if loc_initialW is None:
            loc_initialW = chainer.initializers.Normal(0.001)
        if score_initialW is None:
            score_initialW = chainer.initializers.Normal(0.01)
        if rpn_initialW is None:
            rpn_initialW = chainer.initializers.Normal(0.01)

        if vgg_initialW is None and pretrained_model:
            vgg_initialW = chainer.initializers.constant.Zero()

        feature = VGG16FeatureExtractor(initialW=vgg_initialW)
        rpn = RegionProposalNetwork(
            512, 512,
            ratios=ratios,
            anchor_scales=anchor_scales,
            feat_stride=self.feat_stride,
            initialW=rpn_initialW,
            proposal_creator_params=proposal_creator_params,
        )
        head = VGG16RoIPoolingHead(
            n_class,
            roi_size=7, spatial_scale=1. / self.feat_stride,
            vgg_initialW=vgg_initialW,

Instances