Error in example code - ordinary regression

[maswiebe]

When I run:

library(tidyverse)
library(gridExtra)
library(zaminfluence)
library(AER)

compare <- function(x, y) { return(max(abs(x - y))) }
check_equivalent  <- function(x, y) { stopifnot(compare(x, y) < 1e-8) }

num_obs <- 10000

set.seed(42)

SummarizeReruns <- function(reruns, preds) {
    reruns_df <- GetSignalsAndRerunsDataframe(signals, reruns, model_grads)
    preds_df <- GetSignalsAndRerunsDataframe(signals, preds, model_grads)

    summary_df <-
        rbind(reruns_df %>% mutate(method="rerun"),
              preds_df %>% mutate(method="prediction")) %>%
        pivot_wider(names_from=method, values_from=value)
    return(summary_df)
}

#############################
# Oridinary regression.

# Generate data.
set.seed(42)
x_dim <- 3
param_true <- 0.1 * runif(x_dim)
df <- GenerateRegressionData(num_obs, param_true, num_groups=NULL)

# Fit a regression model.
x_names <- sprintf("x%d", 1:x_dim)
reg_form <- formula(sprintf("y ~ %s - 1", paste(x_names, collapse=" + ")))
fit_object <- lm(data = df, formula=reg_form, x=TRUE, y=TRUE)

# Get influence and reruns.
# Derivatives are only computed for keep_pars, which can be in any order.
model_grads <-
    ComputeModelInfluence(fit_object, keep_pars=c("x2", "x1")) %>%
    AppendTargetRegressorInfluence("x1") %>%
    AppendTargetRegressorInfluence("x2")

I get the error: Error in ComputeModelInfluence(fit_object, keep_pars = c("x2", "x1")) : unused argument (keep_pars = c("x2", "x1"))

The code seems to work fine if I drop keep_pars and just run ComputeModelInfluence(fit_object), which is what the readme example does.

[rgiordan]

I just added the keep_pars argument last week. I see it in the master branch here: https://github.com/rgiordan/zaminfluence/blob/14fc32b43638b64b3603849450da3af73760170d/zaminfluence/R/ols_iv_grads_torch_lib.R#L447

So I think this could be caused by trying to run the new example script with an old version of zaminfluence. Are you sure you've got the latest version of zaminfluence installed?

[maswiebe]

Ok, just restarted R.

packageVersion("zaminfluence")

[1] ‘0.0.0.9000’

Now I can get through ComputeModelInfluence(fit_object, keep_pars=c("x2", "x1")), but it breaks on AppendTargetRegressorInfluence("x1"):

Error in cpp_torch_float64() : Lantern is not loaded. Please use install_torch() to install additional dependencies.

But this doesn't work: install_torch()

trying URL 'https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-1.10.2%2Bcpu.zip' Error in utils::download.file(library_url, temp_file) : cannot open URL 'https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-1.10.2%2Bcpu.zip' In addition: Warning message: In utils::download.file(library_url, temp_file) : URL 'https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-1.10.2%2Bcpu.zip': status was 'SSL peer certificate or SSH remote key was not OK'

Fyi sessionInfo()

R version 4.1.3 (2022-03-10) Platform: x86_64-pc-linux-gnu (64-bit) Running under: Ubuntu 20.04.4 LTS

[rgiordan]

This sounds like a torch problem, though of course it may be somehow related to how I have set up dependencies.

Are you able to succesfully follow the installation instructions for the torch package?

https://torch.mlverse.org/docs/

[maswiebe]

Ok, install.packages("torch") did the trick. Works fine now!