This is probably something I'm doing wrong, but I managed to get the TSAN algorithm working on my dataset (I think). However, the results are largely random noise. See image below of predicted vs. actual ages.

I used the two-stage model and the shell script bash_test_second_stage.sh. In addition, I used the pretrained model Second_Stage_ScaleDesne/Second_ScaleDense. I did have to alter the Python script prediciton_second_stage.py from

    if opt.model == 'ScaleDense':
        model = Second_stage_ScaleDense.second_stage_scaledense(8, 5, opt.use_gender)
    else:
        print('Wrong model choose')

to

    if opt.model == 'Second_ScaleDense':
        model = Second_stage_ScaleDense.second_stage_scaledense(8, 5, opt.use_gender)
    else:
        print('Wrong model choose')

Also, the output of the the code when it was ran is:

======= start prediction =========

TEST : [steps 2], Loss 278.7934, MAE: 13.3956

STD_err = 15.033906
CC: [[ 1. -0.20014753]
[-0.20014753 1. ]]
PAD spear man cc SpearmanrResult(correlation=-0.8653139925024903, pvalue=2.9300691966550956e-20)
spear man cc SpearmanrResult(correlation=-0.1856620308418212, pvalue=0.14187989318250752)
mean pad: -7.2646427

==========================