Closed Radonirinaunimi closed 1 year ago
This seems to be a bug - else how come the chi2_val is so huge?
It may also be useful to plot theory as ratio to data, to better identify this kind of discrepancies. But I don't see how one can get such a poor chi2 since the data in the validation subset agrees rather well with the theory
This is admittedly an odd issue. For a different replica, the results are given below. Notice how the $\chi^2{\rm exp}$ has improved significantly while the $\chi^2{\rm vl}$ is still (slightly) worse.
Dataset | Epoch | REP ID | $\rm{N}_{\rm tr}$ | $\chi^2_{\rm tr}$ | $\rm{N}_{\rm vl}$ | $\chi^2_{\rm vl}$ | $\rm{N}_{\rm tot}$ | $\chi^2_{\rm exp}$ |
---|---|---|---|---|---|---|---|---|
NUTEV_F2 | 5559 | 1 | 58 | 0.659 | 20 | 11.721 | 78 | 1.893 |
In the same way as before, the 20 data-points forming the validation set are shown in the table below: | $x$ | $Q^2_M~([0, 1])$ |
---|---|---|
0.015 | 0 | |
0.015 | 0.115 | |
0.015 | 0.323 | |
0.080 | 0. | |
0.125 | 0.115 | |
0.125 | 0.466 | |
0.125 | 0.609 | |
0.125 | 0.739 | |
0.175 | 0.206 | |
0.175 | 0.856 | |
0.225 | 0.466 | |
0.275 | 0.609 | |
0.275 | 0.856 | |
0.35 | 0.466 | |
0.45 | 0.323 | |
0.45 | 0.974 | |
0.55 | 0.609 | |
0.55 | 0.856 | |
0.55 | 0.934 | |
0.65 | 0.974 | |
So the problem was that some datapoints (in the above example, only a single point) were artificially large because the shifts were so large due to a bug that is fixed in #45. Now the results are reasonable (and converge faster): | Dataset | Epoch | REP ID | $\rm{N}_{\rm tr}$ | $\chi^2_{\rm tr}$ | $\rm{N}_{\rm vl}$ | $\chi^2_{\rm vl}$ | $\rm{N}_{\rm tot}$ | $\chi^2_{\rm exp}$ |
---|---|---|---|---|---|---|---|---|---|
NUTEV_F2 | 567 | 35 | 58 | 1.171 | 20 | 2.7866 | 78 | 2.501 |
As shown above, there shouldn't be a reason for the validation $\chi^2$ to be large this large given that no prediction is (significantly) far way from the true validation points.