train_2step_wma.py

[youtang1993]

Thank you for your excellent work and great effort to improve performance! I was having a little trouble when I tried to retrain the 2step model using the updated "train_2step_wma.py" file. I set the parameters exactly as described in the paper. In fact, I had not modified any parameters in the program at all and everything remained the default. But I can not get the same results as provided in the description. That makes me very troubled and it must be something important to be ignored. Would you please help me?

The MR I got is 13.25, which corresponds to model “resnet_e80_l0.0647282644082.hdf5”. Reduce learning rate to 1e-5 and 1e-6 can not significantly improve performance. Obviously, there is still a big gap between this and the best results (below 12) provided in “results_2step_wma.txt”.

The other thing I want to confirm is that the download link for the backbone Resnet-50 points to "resnet50_weights_tf_dim_ordering_tf_kernels.h5". In fact, we need "resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5". Does this mean that I should change the file name?