hdjang / feature-selective-anchor-free-module-for-single-shot-object-detection Goto Github PK

I train the FSAF and RetinaNet model respectively with default config on single PC, but get a strange result which the mAP of RetinaNet is 28.7% and that of FSAF is 27.0%, but according to your paper, the mAP of FSAF might be 0.2% higher than RetinaNet model.

I train model like that:
python tools/train.py configs/retinanet_r50_400_050x.py
python tools/train.py configs/fsaf_r50_400_050x.py

run on unbuntu 18.04
python 3.7
Pillow version is 6.1.0
pytorch version is 1.1.0

Hi, I try to train and test on Windows 10, but the mmdetection don't support using on windows. Can you provide a pure pytorch version.
Very appreciate!

I run the command: python tools/test.py
but appear an error :ImportError: cannot import name 'init_dist' from 'mmdet.apis' (/home/yu/Downloads/mmdetection/mmdet/apis/init.py)
how to solve the problem Its very important for me.
TorchVision: 0.4.0a0
OpenCV: 4.2.0
MMCV: 0.4.4
MMDetection: 1.1.0+unknown
MMDetection Compiler: GCC 7.5
MMDetection CUDA Compiler: 10.0

I noticed that the calculation of anchor-free cls & reg preds used following codes:

      def forward_single(self, x):
        cls_feat = x
        reg_feat = x
        for cls_conv in self.cls_convs:
            cls_feat = cls_conv(cls_feat)
        for reg_conv in self.reg_convs:
            reg_feat = reg_conv(reg_feat)
        
        cls_score = self.fsaf_cls(cls_feat)
        bbox_pred = self.relu(self.fsaf_reg(reg_feat))
        
        return cls_score, bbox_pred

    def forward(self, feats):
        return multi_apply(self.forward_single, feats)

where multi_apply() used the map() function to apply forward_single() to each feature map in feats. However the self.cls_conv, self.reg conv etc. should share their weights across different scale feature map, is this true ?

thanks

Hi, @hdjang. I find the loss of bounding box is IoU loss according to this paper. But the config file shows the loss of bounding box is SmoothL1 loss.

有没有不是基于mmdetection的版本的代码呀

hi, ty for ur sharing. It's very helpful.
I have some question with lr.
I saw ur GPU_number is 4, num_images_per_GPU is 8. And lr rate is 0.01.
But according to the original mmdetection setting, shouldn't it to be 0.01*2?
(mmdetection GPU=8,num_images_per_GPU=2 lr is 0.01)

hi @hdjang ， have you reproduced the performance reported in the paper? thks