ENH: Surface more errors from the glmnet solver #29
Conversation
The glmnet solver uses integer codes to communicate errors and warnings. The error code returned by then solver is saved to the attribute `jerr` after fitting a model. Negative values denote warnings such as convergence issues, positive values denote fatal conditions such as memory allocation problems, and a value of zero is used when the solver runs successfully without error. Initially we translated the convergence warnings into more complete messages from the numeric code. For all other errors, we just returned something opaque like "glmnet error no. 123." In this PR, we add some additional messages and raise the relevant type of warning or exception.
glmnet/errors.py
Outdated
| _convergence_errors(jerr, n_lambda) | ||
|
|
||
|
|
||
| def _fatal_errors(jerr, n_lambda): |
There was a problem hiding this comment.
Is there a reason to pass n_lambda here? Looks like the function doesn't refer to it
There was a problem hiding this comment.
No, it's only there to make this function consistent with _convergence_errors. Although it looks like it may be necessary to decode the additional errors that could be returned by lognet.
There was a problem hiding this comment.
Hmm... looks like n_lambda isn't needed in either case. The convergence code is in [-k, -1].
glmnet/errors.py
Outdated
|
|
||
| def _check_glmnet_error_flag(jerr, n_lambda): | ||
| """Check the error flag. Issue warning on convergence errors (jerr < 0) | ||
| and exception on anything else.""" |
There was a problem hiding this comment.
Your intro paragraph that describes the interpretation of jerr is good -- maybe stash it here in the docstring for a quick reference?
It's only a few additional messages. I'll update the PR to include these. |
|
Hi Bill! (disappear) |
The glmnet solver uses integer codes to communicate errors and warnings. The error code returned by then solver is saved to the attribute
jerrafter fitting a model. Negative values denote warnings such as convergence issues, positive values denote fatal conditions such as memory allocation problems, and a value of zero is used when the solver runs successfully without error.Initially we translated the convergence warnings into more complete messages from the numeric code. For all other errors, we just returned something opaque like "glmnet error no. 123."
In this PR, we add some additional messages and raise the relevant type of warning or exception.