Hi，does s2anet support multi anchor_ratios? about s2anet

[csuhan]

Hi, @PressEtoRace . Following the trend of anchor-free models, s2anet uses one anchor per location in the feature maps, which saves computations and runs faster. So we haven't implemented multiple anchors learning. But it is not impossible to train s2anet with multiple anchors. Some methods in s2anet_head should be modified:

1. _init_layers, forward_single. You need add multiple anchors like RetinaNet. Also, modify anchor generation method and bbox_decode method for a batch of anchors. Besides, modify anchor2offset in AlignConv so that multiple anchors in per location can get multiple offset (see here).
2. loss and get_refine_anchors. Modify get_refine_anchors, loss_fam_single and loss_odm_single.
3. get_bboxes.

Take Retinanet as a reference, it will be not hard if you fully understand our paper and mmdetection.

[vision-zhu]

Have you implemented multiple scales? @PressEtoRace , Can you share me the model after change? Thank you very much! My e-mail is [email protected]

[LoveDLWujing]

i just modified the AlignConv to support multi scale and multi ratio anchor setting, it seems to work

    def forward(self, x, anchors, stride):
        out = list()
        for i in range(self.anchor_num):
            offset = anchor2offset(anchors[..., i * 5: (i+1) * 5], self.kernel_size, stride)
            out.append(self.relu(self.deform_conv(x, offset)))
        # TODO: max-pooling over multi anchor to get feature
        out = torch.stack(out, dim=1)
        out, _ = torch.max(out, dim=1, keepdim=False)
        return out

and modified the _init_layers

self.fam_reg = nn.Conv2d(self.feat_channels, self.num_anchors * 5, 1)
self.fam_cls = nn.Conv2d(self.feat_channels, self.num_anchors * self.cls_out_channels, 1)

self.odm_cls = nn.Conv2d(self.feat_channels, self.num_anchors * self.cls_out_channels, 3, padding=1)
self.odm_reg = nn.Conv2d(self.feat_channels, self.num_anchors * 5, 3, padding=1)

[taoboyang]

i just modified the AlignConv to support multi scale and multi ratio anchor setting, it seems to work

    def forward(self, x, anchors, stride):
        out = list()
        for i in range(self.anchor_num):
            offset = anchor2offset(anchors[..., i * 5: (i+1) * 5], self.kernel_size, stride)
            out.append(self.relu(self.deform_conv(x, offset)))
        # TODO: max-pooling over multi anchor to get feature
        out = torch.stack(out, dim=1)
        out, _ = torch.max(out, dim=1, keepdim=False)
        return out

and modified the _init_layers

self.fam_reg = nn.Conv2d(self.feat_channels, self.num_anchors * 5, 1)
self.fam_cls = nn.Conv2d(self.feat_channels, self.num_anchors * self.cls_out_channels, 1)

self.odm_cls = nn.Conv2d(self.feat_channels, self.num_anchors * self.cls_out_channels, 3, padding=1)
self.odm_reg = nn.Conv2d(self.feat_channels, self.num_anchors * 5, 3, padding=1)

Hi I want to know How to set self.anchor_num

[seacloud-0420]

@LoveDLWujing
Hi，I followed the code you posted with no success. Could you please post all the changed code?

[seacloud-0420]

@PressEtoRace @vision-zhu @taoboyang
Use multiple anchors to train, have you implemented it successfully? Can you share me the model after change? Thank you very much!