Scaling the unit covariance gaussian to the PCA space

[chahuja]

As eigen values of the covariance matrix (1/n * X^T @ X) are the square of the eigen values of the random variable X, the scaling factor should be e**0.5 instead of e.

FaceEditor/encoder.py

Line 144 in 6951960

x_vecs = x_mean + np.dot(v, (rand_vecs * e).T).T

[martianmartina]

Yes, one singular value of X should be the square root of an eigenvalue of the covariance matrix of X. I think the problem is fixed in face_edit.py where he took the square root after loading evals.npy, but I agree that this could cause a problem in encoder.py. Good catch!