The hyper-parameters for reproduce the deform conv model on VOC

[yuzcccc]

How many iterations does the model need to reproduce the results? I use the same hyper-parameters as for the baseline faster rcnn w/o deform conv., the final loss is much higher than the baseline, and the mAP is not satisfactory.

My setting is max_iter = 30k (with 4 gpus, iter_size 2) . base_lr = 0.001, gamma = 0.1, stepsize = 20k The baseline mAP is 77.9 while the deform conv. one is 73.x. The backbone model is ResNet-101