Open chuangchuangtan opened 3 years ago
Thanks for your attention, we did not use CRF and its variants.
Can you release the details of training strategy for CAM, Has, ACol, CutMix, SPG and ADL? When implementing those methods, I get different results comparing with Table 3 of SPA paper. Thank you.
Did you use 23151 or 50000 images to calculate results of Peak-T, Peak-IoU and GT-Known Loc in Table 3?
around 23k images
I implement CutMix twice according to evaluation script, train script and evaluation hyperparameter, and con't get results of Peak-T, Peak-IoU in Table 3. I achieve Peak-IoU 52.96 and 52.80, instead of Peak-IoU 54.54.
CUDA_VISIBLE_DEVICES=0 python main.py --dataset_name ILSVRC --architecture inception_v3 --wsol_method cutmix --experiment_name ILSVRC_inception_v3_cutmix_20210507 --pretrained TRUE --num_val_sample_per_class 5 --large_feature_map TRUE --batch_size 32 --epochs 10 --lr 0.00024974608 --lr_decay_frequency 3 --weight_decay 5.00E-04 --cutmix_beta 0.08 --cutmix_prob 0.49 --override_cache FALSE --workers 8 --box_v2_metric FALSE --iou_threshold_list 50 --eval_checkpoint_type last --cam_curve_interval 0.001
@chuangchuangtan I checked our code and models. The main differences compared with your configuration are: we used 8V100 GPUs and set bacth_size=1024, lr=0.005, lr_decay=10,15, total_epoch=20, cutmix_beta=1.0, cutmix_prob=0.5
It is confusing in section 4.1 that "For fair comparisons, we adopt the same training strategy with SEM."
Sorry for the confusing description. It means that we add SEM and SPA on the same model re-implemented by us, respectively. Since both methods work in the testing phase, it is a fair comparison on the basis of the same model in our opinion.
Thank you for your attention.
Hi , Thanks for sharing the code. When calculating Peak-IoU, did you apply CRF or ConvCRF on localization map as SEM?