Github user takuti commented on the issue:
https://github.com/apache/incubator-hivemall/pull/87
> It may exists bugs in the Optimizer implementation. Too large losses are
found when doing early termination.
Too large losses are probably observed as a consequence of inappropriate
hyper-parameters. In the last implementation, I tried to test all loss
functions under the same `-eta0` and `-lambda` as:
```java
for (String opt : optimizers) {
for (String reg : regularizations) {
if (reg == "RDA" && opt != "AdaGrad") {
continue;
}
for (String loss : lossFunctions) {
String options = "-opt " + opt + " -reg " + reg + " -loss " +
loss
+ " -lambda 1e-6 -cv_rate 0.005 -iter 512";
```
However, since each loss function draws different curve, optimal `-eta0`
and `-lambda` should be different.
Finding optimal hyper-parameters for each combination of `-opt` + `-reg` +
`-loss` is really hard, so I've updated `assert` condition; a new version of
`GeneralRegressionUDTFTest` simply checks if iterative training improves MAE to
some degree, by comparing with MAE w/ initial state of model.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
