RichardYip

> 日志： > > RuntimeError：预计在开始新的迭代之前已经完成了先前迭代的缩减。此错误表明您的模块具有未用于生成其输出的参数（forward 的返回值）。您可以通过将关键字参数_**find_unused_pa​​rameters=True**_传递给 torch.nn.parallel.DistributedDataParallel 来启用未使用参数检测。如果您已经设置了此参数，则分布式数据并行模块无法在模块的 forward 函数的返回值中定位输出张量。报告此问题时请包括您的模块前向返回值的结构（例如列表、字典、可迭代） > > 使固定： > > https://github.com/open-mmlab/mmaction/blob/c7e3b7c11fb94131be9b48a8e3d510589addc3ce/mmaction/apis/train.py#L73 > > 从：find_unused_pa​​rameters = cfg.get('find_unused_pa​​rameters', False) 到：find_unused_pa​​rameters = cfg.get('find_unused_pa​​rameters', True) good thanks

> box = [xmin,ymin,xmax,ymax] new_box = [xmin/img_w,ymin/img_h,x_center/img_w,y_center/img_h]??? 关键点 kp0 = [x0,y0] new_kp0 = [x0/img_w,y0/img_h]?? I think it should be as follows: new_box = [(xmin+xmax)/2/img_w,(ymin+ymax)/2/img_h,(xmax-xmin)/img_w,(ymax+ymin)/img_h]

this is the problem caused by different version of face_alignment， you can change the function from `face_alignment.LandmarksType._2D` to `face_alignment.LandmarksType.TWO_D`

> > I built adaptive gridsampler as follows. > > * torch=1.8.1 > * cuda 11.1 > * RTX3080 > * modified setup.py as follows > > * changed cxx_args...

> I saw a model download has a model named "GFPGANv1_net_d_mouth.pth". Can this model only repair the mouth area and have a faster speed？ this is the weight of Discriminator,not...

same question ， don't know why