For training Resnet

[LifeBeyondExpectations]

Thanks for sharing the code. I would like to know solver.prototxt for Resnet-101 or Resnet-101-semmantic. How did you train Resnet-101 that scored like mAP(75.2), F1-O(74.4), F1-C(69.5) for coco dataset ???

If I trained Resnet-101 with provided prototxts and just only (step_1) / (step_1, step_4(finetune)), the scores are not close to the reported scores in the paper.

[LifeBeyondExpectations]

• I would add some more comments

If I trained Resnet-101 with " step_1_resnet101_solver.prototxt ", the results are mAP(73.8) F1-C(67.4) F1-O(72.9)

If I trained Resnet-101 with " step_1_resnet101_solver.prototxt " and "step_4_resnet101_srn_finetune_solver(with modification of network from SRN to Resnet-101(only))", the results are mAP(73.9) F1-C(67.5) F1-O(73.0)

However, reported scores for Resnet-101 in the paper are mAP(75.2) F1-C(69.5) F1-0(74.3)

So .. I would like to ask you the training prototxt of Resnet-101 only, not Resnet-101-SRN. For SRN, I checked results from the code are almost the same as scores in the paper.(Thanks for your sharing again)

[zhufengx]

I would add some more comments If I trained Resnet-101 with " step_1_resnet101_solver.prototxt ", the results are mAP(73.8) F1-C(67.4) F1-O(72.9)

If I trained Resnet-101 with " step_1_resnet101_solver.prototxt " and "step_4_resnet101_srn_finetune_solver(with modification of network from SRN to Resnet-101(only))", the results are mAP(73.9) F1-C(67.5) F1-O(73.0)

However, reported scores for Resnet-101 in the paper are mAP(75.2) F1-C(69.5) F1-0(74.3)

So .. I would like to ask you the training prototxt of Resnet-101 only, not Resnet-101-SRN. For SRN, I checked results from the code are almost the same as scores in the paper.(Thanks for your sharing again)

Hi, @LifeBeyondExpectations , thanks for your comments. "Resnet-101" is exactly trained by "step_1_resnet101_solver.prototxt". So, I am wondering if you have modified some settings in the solver? Did you use the provided pretrained model?