query2labels
query2labels copied to clipboard
SlongLiu
Official implementation of paper "Query2Label: A Simple Transformer Way to Multi-Label Classification".
backbone使用的是 swin_B_224_22k，在coco2014上做的训练，下面是其他训练参数： --batch-size 256 \ --gamma_pos 0 --gamma_neg 2 --dtgfl \ --epochs 80 \ --lr 1e-5 \ --optim AdamW \ --pretrained \ --weight-decay 1e-2 \ --cutout --n_holes 1 --cut_fact 0.5...
Is there a config file for training on the VOC dataset? If other backbones are used, such as TResNet-M, do the corresponding model parameters need to be adjusted?
Why when I train "!python query2labels/main_mlc.py \ --backbone resnet101 --dataname coco14 --batch-size 1 --print-freq 100 \ --output "path/to/output" \ --world-size 1 --rank 0 --dist-url tcp://127.0.0.1:3717 \ --gamma_pos 0 --gamma_neg 2...
there is a little bug in q2l_infer.py which leads to args.resume will be covered by the "resume " in config.json, and make some people can not repreduct paper's results。 for...
how to set other hyperparameters to achieve paper results in batch size 32? Has anyone tried it
看源代码，在训练时，main_mlc.save_checkpoint里面，结果保存在model_best.pth.tar，这样的pth文件。 在推理的时候，作者给的脚本 ``` python q2l_infer.py -a modelname --config /path/to/json/file --resume /path/to/pkl/file [other args] e.g. python q2l_infer.py -a 'Q2L-R101-448' --config "pretrained/Q2L-R101-448/config_new.json" -b 16 --resume 'pretrained/Q2L-R101-448/checkpoint.pkl' ``` 为啥是这样啊？应该怎么处理？
Liu, thank you for your work. I am very interested in the attention visualization in the paper. I tried to implement this in my project, but encountered some problems. Can...
Why can I obtain correct results during inference, but during training, I can only achieve a MAP (Mean Average Precision) of 3.7
Why can I obtain correct results during inference, but during training, I can only achieve a MAP (Mean Average Precision) of 3.7
Thanks for your work. Have you ever tried not to use the ASL loss function? Will a simple ce loss function cause an obvious performance decrease?
