Mboost - Test Predict Function - stmboost #3
Assignees
Comments
avinashbarnwal
changed the title
|
@avinashbarnwal Can you also post the error log? |
|
@hcho3, please find the error - |
|
What is the content of |
|
These are internal calls. I am not sure about the internals. |
|
Try adding some printing statements in the lines in the error trace. |
|
It is throwing an error when we are calling - |
|
IMO the example is still not minimal enough. does the error message happen on all data sets or just this one? |
|
example(stmboost) works for me, > example(stmboost) |
|
this is more minimal |
|
here is another MRE with simulated data. library(tram)
library(tbm)
N <- 100
set.seed(1)
x <- runif(N)
y <- 10+30*x+rnorm(N)
y.lower <- y-runif(N)
y.upper <- y+runif(N)
plot(y ~ x)
segments(x, y.lower, x, y.upper)
df <- data.frame(x, y.lower, y.upper)
f <- survival::Surv(y.lower, y.upper, type="interval2") ~ x
m_mlt <- Survreg(f, data = df, dist = "lognormal")
bm <- stmboost(m_mlt, formula = f, data = df,control = boost_control(mstop=200,nu=0.01,trace=TRUE),method = quote(mboost::mboost))
q = seq(from = min(df$y.lower), to = max(df$y.upper), length.out = 100)
sessionInfo()
d = predict(bm, newdata = df, type = "density", q = q)output: |
|
Thanks @tdhock. |
|
Torsten's new code works for me @avinashbarnwal library(tram)
library(tbm)
N <- 100
set.seed(1)
x <- runif(N, min = -.5, max = .5)
### I made this a log-normal DGP
y <- exp(1+2*x+rnorm(N))
### smaller than 0 is not allowed
y.lower <- pmax(y-runif(N), 0.001)
y.upper <- y+runif(N)
plot(y ~ x)
segments(x, y.lower, x, y.upper)
df <- data.frame(x, y.lower, y.upper)
### saver: generate a new variable with a simple name
### because the deparsed call might be a trouble maker
df$y <- survival::Surv(y.lower, y.upper, type="interval2")
### the model is UNCONDITIONAL (no x here)
### (you can have conditional models, but this is more complex)
m_mlt <- Survreg(y ~ 1, data = df, dist = "lognormal")
### but boosting needs the x
f <- y ~ x
bm <- stmboost(m_mlt, formula = f, data = df,control = boost_control(mstop=200,nu=0.01,trace=TRUE),method = quote(mboost::mboost))
q = seq(from = min(df$y.lower), to = max(df$y.upper), length.out = 10)
sessionInfo()
### this is dim(grid) x dim(df), so obs are in columns !!!
dim(d <- predict(bm, newdata = df, type = "density", q = q))Above is the contents of toby.R script from Torsten. Below is the output of running that in R: |
|
Hi Prof. @tdhock, I am not sure why we have different y~1 and y~x. m_mlt <- Survreg(y ~ 1, data = df, dist = "lognormal") I generally keep the same formula for both the places. I am feeling we should make software, where we can reproduce these things more clearly. Things are very hay-wire. |
|
Hi Prof. @tdhock, I have updated the minimal reproducible example code and I have changed. to Then it gave the same error as before. Also for d is null. Check this notebook - https://github.com/avinashbarnwal/aftXgboostPaper/blob/master/src/R/testing/testing_mboost/mboost_MRE1.ipynb I am not sure whats the difference between y~x and y~1 in Survreg. |
|
Hi Prof. @tdhock, Please let me know if you have got the time to look into this issue. |
|
I think the two formulas are for different purposes. In Torsten's email he indicated that y ~ 1 should be used in one case and not the other. So please do that in your code. |
|
Thanks for the update Prof. @tdhock. I will use it accordingly. But i am confused what the difference? Do you think we should ask Prof. Torsten and also predicted values from y~1 are NAs? |
|
I don't know what is the difference, you should ask Torsten for clarification. but I don't think it is relevant to what you are doing. I think you should just copy/modify the code he sent us, and that should be fine for baseline predictions. |
|
Yes, but even using y~1 predictions are nulls. |
|
I don't understand / can't replicate the nulls you describe. When I do library(tram)
library(tbm)
N <- 100
set.seed(1)
x <- runif(N, min = -.5, max = .5)
### I made this a log-normal DGP
y <- exp(1+2*x+rnorm(N))
### smaller than 0 is not allowed
y.lower <- pmax(y-runif(N), 0.001)
y.upper <- y+runif(N)
plot(y ~ x)
segments(x, y.lower, x, y.upper)
df <- data.frame(x, y.lower, y.upper)
### saver: generate a new variable with a simple name
### because the deparsed call might be a trouble maker
df$y <- survival::Surv(y.lower, y.upper, type="interval2")
m_mlt <- Survreg(
y ~ 1, #DO NOT CHANGE THIS.
data = df,
dist = "lognormal")
bm <- stmboost(
m_mlt,
formula = y ~ x, #CHANGE x to all input variables.
data = df,
control = boost_control(mstop=200,nu=0.01,trace=TRUE),
method = quote(mboost::mboost))
pred.mat <- predict(
bm,
newdata = df,
type = "density",
q = 0.5)#help(predict.mlt) q: quantiles at which to evaluate the model
pred.mat[, 1:5]
sessionInfo()I get the following result: |
|
the important part with the real-valued predictions is |
|
Hi Prof. @tdhock, I am still finding null values. Below is the code - I am not sure how you are getting different results? |
|
Hi Prof. @tdhock , We are running on a different OS. I am using Mac and you are using Windows. Also, my version for the tram is latest - tram_0.3-1. You are using tram_0.3-0. Please let me know if you think these are the reasons for difference in results. |
|
Hi Prof. @tdhock, I have been able to reproduce the minimal reproducible example in Linux. This is AWS EC2 provided by @hcho3. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hi Prof. @tdhock and @hcho3,
Please find the minimal reproducible example to test the predict function of mboost -
The text was updated successfully, but these errors were encountered: