search

XiaozhiChen / resnet-generator

Generate Caffe Prototxt for Deep Residual Learning Network
71 stars 51 forks source link

training result with faster-rcnn #4

Open banxiaduhuo opened 8 years ago

banxiaduhuo commented 8 years ago

@XiaozhiChen Hi, Xiaozhi. I used your program to create the resnet50's train.prototxt and test.protxt, then trained with faster-rcnn. And I got a very low result on voc2007, about 0.47, even lower than ZF model's 0.62. What's the result you got on voc2007?

Is my solver correct? base_lr: 0.001 lr_policy: "multistep" gamma: 0.1 stepvalue: 300000 stepvalue: 500000 display: 20 momentum: 0.9 weight_decay: 0.0001 snapshot: 0

Thank you very much!

XiaozhiChen commented 8 years ago

@banxiaduhuo A minor bug was fixed in the script. I got 71.8 mAP on VOC2007 test set using the the prototxt and solver with model trained on VOC07 trainval.

Higher AP could be obtained with some other changes.

banxiaduhuo commented 8 years ago

@XiaozhiChen Thank you~ I train the ResNet50 and fix all BN+Scale like your fine-tune.prototxt. After 70k iterations, I got a 0.65 mAP on VOC2007. The results of previous 0.47 does not fix the BN+Scale, It seems BN must be fixed when fine-tuning. Did you also use the fine-tune.prototxt to train? And how many iterations?

XiaozhiChen commented 8 years ago

@banxiaduhuo I was using the finetune.prototxt and 70K iterations.

banxiaduhuo commented 8 years ago

@XiaozhiChen Thank you very much!