thank you for sharing your code!
If a subject has many observations, e.g., some subjects in MPII Gaze have ~50.000 observations, you are going to invert a matrix of that size:
|
V_hat_inv_i = V_hat_i \ eye(n_i); |
Did you sub-sample your training/validation data (and were very patient at test time)? The Woodbury matrix identity can theoretically be used here but it is unfortunately numerically very unstable.
thank you for sharing your code!
If a subject has many observations, e.g., some subjects in MPII Gaze have ~50.000 observations, you are going to invert a matrix of that size:
MeNets/scripts/em_lmecnn.m
Line 45 in 3537e99
Did you sub-sample your training/validation data (and were very patient at test time)? The Woodbury matrix identity can theoretically be used here but it is unfortunately numerically very unstable.