>>> a = np.random.default_rng(0).random((2,4))
>>> a
array([[0.63696169, 0.26978671, 0.04097352, 0.01652764],
[0.81327024, 0.91275558, 0.60663578, 0.72949656]])
>>> b = np.random.default_rng(1).random((2,4))
>>> b
array([[0.51182162, 0.9504637 , 0.14415961, 0.94864945],
[0.31183145, 0.42332645, 0.82770259, 0.40919914]])
>>> math.sqrt(sklearn.metrics.mean_squared_error(a, b))
0.5001152789168666
>>> np.mean(np.abs(a - b))
0.42166963796730605
besides, the whole util module is missing.
Evaluation of ID is totally missing! Please release a detailed code or guide of computing ID (especially FR net weights and preprocessing). I have followed description from your paper and utilize resnet-50 from arcface_torch and got 0.4386 ID score (far from 0.84 reported in paper) with publised next3d_ffhq_512.pkl. The render images seems OK.
evaluation.py
calculate AED/APD withmath.sqrt(sklearn.metrics.mean_squared_error(np.array(exps), np.array(exps)[:, :50]))
cal_3dmm_distance.py
calculate AED/APD withAED = np.mean(np.abs(s_exp - t_exp))
these two are obviously different evidenced as below:besides, the whole
util
module is missing.next3d_ffhq_512.pkl
. The render images seems OK.