Wu0409
Base result reproduction problem
Hello, I'm very interested in your work. I'm currently trying to run a baseline using a single-branch CNN student network (resnet50 and resnet34). After disabling ptc_loss and sim_loss, I've tried adjusting hyperparameters like --lr and --cam_iter, but the training results are disappointing (CAM is only around 38%, SEG is only 35%).
Use the following command: python -m torch.distributed.run --nproc_per_node=2 train_final_voc.py --data_folder ../../VOCdevkit/VOC2012/ --cam_iters 4000 --high_thre 0.65 --w_seg 0.3 --lr 3e-5
Have you ever tried using a CNN backbone as the student network?
This problem has been bothering me for several days. No matter how I adjust the parameters, I can't further improve the results. I look forward to your reply.
DuPL has not been experimented on with CNN architectures. Single-stage methods generally use the ViT architecture.
The pipeline itself does not conflict with CNNs, but the experimental setup and other aspects might require significant adjustments. I would recommend using ViT as the baseline for experiments.
