search

tianzhi0549 / FCOS

FCOS: Fully Convolutional One-Stage Object Detection (ICCV'19)
https://arxiv.org/abs/1904.01355
Other
3.27k stars 629 forks source link

If I use other backbone to train rather than fcos_imprv_R_50_FPN_1x, I need to change something for inference? #323

Closed pp00704831 closed 3 years ago

pp00704831 commented 3 years ago

Hello, I use fcos_imprv_R_50_FPN_1x as a backbone network for my own datasets, it can work successfully for training and testing like below: image

But when I use other backbones such as fcos_imprv_R_101_FPN_1x or fcos_imprv_dcnv2_R_50_FPN_1x, it can train successfully because its loss value will decline gradually. image

But when I use inference code " python tools/test_net.py --config-file configs/fcos/fcos_imprv_dcnv2_R_50_FPN_1x.yaml MODEL.WEIGHT ./training_dir/fcos_imprv_dcnv2_R_50_FPN_1x/model_0027500.pth " for other backbones rather than fcos_imprv_R_50_FPN_1x, it will get zero value for all AP image

Because it can be successfully trained, as shown above. Do I need to change something for testing with other backbones rather than fcos_imprv_R_50_FPN_1? I have already change the --config-file and MODEL.WEIGHT.

tianzhi0549 commented 3 years ago

@pp00704831 It seems your model is different from the one during training. Please check all the weights are successfully loaded during testing.

pp00704831 commented 3 years ago

Hello, But I just used your training codes, and your training and testing command, the training weights will be saved at ./training_dir/fcos_imprv_dcnv2_R_50_FPN_1x. The most strange thing is that only fcos_imprv_R_50_FPN_1x can be successfully tested @@, others will show 0 AP value. I am keeping trying training on other backbones to check if this situation will keep occurring.