Tian Zhi

@chenweifu2008 Our project is based on Detectron2, so you can refer to the documents of Detectron2 for this. https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html The easiest way should be to convert your data to the...

@APeiZou It seems your data format is not correct.

@Willy0919 Please manually specify `--dist-url` (with a different port) in the training command line.

Please place `--dist-url tcp://127.0.0.1:50001` before options, for example. ``` python tools/train_net.py \ --config-file configs/FCOS-Detection/R_50_1x.yaml \ --dist-url tcp://127.0.0.1:50001 \ --num-gpus 4 OUTPUT_DIR training_dir/fcos_R_50_1x_3d_ctr_real ```

@sankin97 Could you please post your results here?

@sankin97 You should use COCO pre-training.

@sankin97 The results do not seem normal. For both CondInst and BoxInst, we use the same settings as in https://github.com/facebookresearch/detectron2/blob/master/configs/Cityscapes/mask_rcnn_R_50_FPN.yaml.

@Alxx999 `gt_inds` means the ground-truth indices that the training samples correspond to, which is computed at https://github.com/aim-uofa/AdelaiDet/blob/73182b7b2216e6bd93fdd1ae576166bc1871f2f6/adet/modeling/fcos/fcos_outputs.py#L253

@huangzicheng Please try to use the latest PyTorch and onnx.

@BlueBlueFF No. You can try it by yourself, but I think data augmentation is important on PASCAL VOC because the number of images is small.