How to save glm result without data or only with coeffients for prediction?

Question

When I use the following R code,

model_glm=glm(V1~. , data=xx,family="binomial");
save(file="modelfile",model_glm);

The size of modelfile will be as much as the data, which will be 1gig in my case. How can I remove the data part in the result of model_glm, so I can only save a small file.

Set `model = FALSE` in your call to `glm` to prevent the model.frame from being returned. — BenBarnes, Nov 14 '12 at 21:29

score 11 · Accepted Answer · answered Nov 14 '12 at 21:36

11

Setting model = FALSE in your call to glm should prevent the model.frame from being returned. Also setting y = FALSE will prevent the response vector from being returned. x = FALSE is the default setting and prevents the model.matrix from being returned.

This combination should shrink the size of your glm object.

Of course, you can also extract the coefficients with coef(model_glm) or, with standard errors,

summary(model_glm)$coef

answered Nov 14 '12 at 21:36

BenBarnes

19,114
6
56
74

1

Have you looked at resulting model? It is still very heavy: residuals, fitted values and the full qr matrix. The `biglm` package returns much lighter model objects. – hadley Nov 15 '12 at 04:00
Yes, the `bigglm` function may be a viable alternative. If you post it an an answer, it'd get a vote from me. Nonetheless, it's still possible to select only the desired components from the `glm` object before saving. – BenBarnes Nov 15 '12 at 05:34

score 8 · Answer 2 · edited Jan 23 '15 at 22:53

8

I had this issue where I was running the GLM as part of an R in production and the size of the GLM greatly slowed me down. I found I needed to kill off more than just the $data. Here is my post on it, with an example below.

> object.size(sg)
96499472 bytes
> sg$residuals <- NULL
> sg$weights <- NULL
> sg$fitted.values <- NULL
> sg$prior.weights <- NULL
> sg$na.action<- NULL
> sg$linear.predictors <- NULL
> sg$fitted.values <- NULL
> sg$effects <-NULL
> sg$data <- NULL
> object.size(sg)
3483976 bytes
> sg$qr$qr <- NULL
> object.size(sg)
79736 bytes

edited Jan 23 '15 at 22:53

Nic

6,211
10
46
69

answered Jan 23 '15 at 21:30

Levi Bowles

81
1
2

Just a comment... In your blog you can barely read the things highlighted in yellow – Rich Scriven Jan 23 '15 at 22:58
1

You could also do something like `sg[c("residuals", "weights", "fitted.values")] <- NULL` where you make that vector include the names of everything you want to get rid of. – Dason Jan 23 '15 at 22:58
1

@RichardScriven Oh my - it burned my eyes. – Dason Jan 23 '15 at 22:59
1

@RichardScriven thanks, fixed, good thing I'm not in data vis. – Levi Bowles Jan 26 '15 at 14:59

score 1 · Answer 3 · answered May 09 '18 at 22:06

object.size() is misleading because it ignores the environment attributes. If you want to assess the true size, use:

length(serialize(model_glm, NULL))

Apart from the data stored, if you want to significantly reduce the size of your glm do:

rm(list=ls(envir = attr(model_glm$terms, ".Environment")),
     envir = attr(model_glm$terms,
              ".Environment"))

This comes from a well detailed article

score 0 · Answer 4 · edited Jul 28 '14 at 04:52

0

You can NULL the data in the model object before saving it. I did a quick test and still generated predictions.

model_glm$data <- NULL

edited Jul 28 '14 at 04:52

Yu Hao

119,891
44
235
294

answered Jul 28 '14 at 04:50

Anon

1

How to save glm result without data or only with coeffients for prediction?

4 Answers4

Linked