hello，I want to ask a question about yolox distillation about fgd HOT 7 CLOSED

For mmdet < 3.0, refer this for optimizer

from fgd.

I revised the optimizer，but it still has error.
File "/home/yxy/anaconda3/envs/yolox/lib/python3.7/site-packages/mmdet/models/backbones/csp_darknet.py
patch_top_left = x[..., ::2,::2]
TypeError: string indices must be integers
But I use the code to distill retinanet,there is no error.

from fgd.

Do you try yolox without distillation ? The error may in mmdet codes instead of distillation.

from fgd.

I try yolox without distillation ,there is no error. I used the version mmdet==2.19.0 and mmcv-full==1.5.0 , is it maybe related to mmdet and mmcv version mismatch? Thank you for your reply.

from fgd.

Fine, I do not know how to fix it either. You can run yolox, that means the mmdet and mmcv-full is matched.

from fgd.

        scale_y = 1.0
        scale_x = 1.0
        gt = kwargs['gt_bboxes']
        if self.yolox:
            scale_y = self.student._input_size[0] / self.student._default_input_size[0]
            scale_x = self.student._input_size[1] / self.student._default_input_size[1]
            student_loss, img, gt = self.student.forward_train(img, img_metas, **kwargs)
        else:
            student_loss = self.student.forward_train(img, img_metas, **kwargs)
        
        with torch.no_grad():
            fea_t = self.teacher.extract_feat(img)

您好，在in the mmdet/distillation/distillers/detection_distiller.py 这个文件里
当self.yolox is true, fea_t = self.teacher.extract_feat(img) 里面的img传入的实际并不是图像，而是loss_bbox.
然后，您能看下，是不是这个问题导致输入的类型错误。但是当我改成student_loss = self.student.forward_train(img, img_metas, **kwargs)的时候，又会报下面的错误。但是我看代码里面有特征图对齐的代码，不知道这是怎么搞的了
assert preds_S.shapel-2:] == preds_T.shapel-2:],'the output dim of teacher and student differ AssertionError: the output dim of teacher and student differ

from fgd.

返回的应该就是img，你看看你有没有对yolox这个detetcor做修改，这里执行forward_train后会返回loss, img, bbox

from fgd.