Github user mktal commented on the issue:
https://github.com/apache/incubator-madlib/pull/75
- You definitely need to be careful with the `step size`, or learning rate,
to make sure that it is either small enough or decayed properly, like other
variants of sgd. For example, we can decay step size for every call of
`structure_svm_step` but within `structure_svm_step` we keep it constant. `
- `epoch` should not be chosen too large otherwise we will overfit one
buffer too much (< 10)
- `batch_size` does not affect the convergence too much and often 32, 64 or
128 are some of the common choices
---
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 [email protected] or file a JIRA ticket
with INFRA.
---
