Error in mm %*% object$betas: non-conformable arguments

[Daviser95]

Hi,

I am using ENMEval with my case study, where basically I am running several models for different species.
I am testing different subset of environmental predictors, within a loop where also model's tuning is performed.
So far, I used this combination of tuned arguments:

tune.args <- list(fc = c("L","LQ","LQH"), rm = c(0.5, seq(1, 5, by = 1), 7.5, 10))

Now, I was testing also different features, specifically the hinge one:

tune.args <- list(fc = c("L","LQ","H", "LQH"), rm = c(0.5, seq(1, 5, by = 1), 7.5, 10))

When using these tuned arguments,within my loop, I got the error:

> for (i in seq_along(subset_final_bgbias_unique)) {
+   cat("/n➡️ Running subset", i, "of", n_subsets, ".../n")
+   
+   vars_subset <- subset_final_bgbias_unique[[i]]
+   gc()
+   env_subset <- terra::subset(envs_ca_set_eros, vars_subset)
+   gc()
+   eval_result <- tryCatch({
+     ENMevaluate(
+     occs = occs_eros[,1:2],
+     envs = env_subset,
+     bg = bg_bias[,1:2],
+     algorithm = "maxnet",
+     tune.args = tune.args,
+     partitions = "block",       # check "randomkfold", "checkerboard2", ecc,
+     parallel = TRUE,
+     numCores = 14
+   )
+ }, error = function(e) {
+   cat("❌ Error in subset", i, ":", conditionMessage(e), "/n")
+   return(NULL)
+ })
+   if (!is.null(eval_result)) {
+     results_list_bias[[names(subset_final_bgbias_unique)[i]]] <- eval_result #check subset list
+   gc()
+   
+   # Estrai la miglior combinazione per questo subset
+   res <- eval_result@results
+   best_index <- which.min(res$AICc)
+   best_row <- res[best_index, ]
+   gc()
+   model_summary_bias <- rbind(model_summary_bias, data.frame(
+     Subset = names(subset_final_bgbias_unique)[i], #check subset list
+     AICc = best_row$AICc,
+     AUC_val = best_row$auc.val.avg,
+     AUC_train = best_row$auc.train,
+     CBI_val = best_row$cbi.val.avg,
+     CBI_train = best_row$cbi.train,
+     fc = best_row$fc,
+     rm = best_row$rm,
+     Variables = paste(vars_subset, collapse = ", ")
+   ))
+   }
+ }
/n➡️ Running subset 1 of 18 .../n*** Running initial checks... ***

* Removed 1 occurrence points with NA predictor variable values.
* Removed 4 background points with NA predictor variable values.
* Clamping predictor variable rasters...
* Model evaluations with spatial block (4-fold) cross validation and lat_lon orientation...

*** Running ENMeval v2.0.5 with maxnet from maxnet package v0.1.4 ***


Of 24 total cores using 14...
Running in parallel ...
❌ Error in subset 1 : 4 nodes produced errors; first error: non-conformable arguments

However, for the 2nd and 3rd subsets of predictors the loop succedeed:

➡️ Running subset 2 of 18 .../n*** Running initial checks... ***

* Removed 1 occurrence points with NA predictor variable values.
* Removed 4 background points with NA predictor variable values.
* Clamping predictor variable rasters...
* Model evaluations with spatial block (4-fold) cross validation and lat_lon orientation...

*** Running ENMeval v2.0.5 with maxnet from maxnet package v0.1.4 ***


Of 24 total cores using 16...
Running in parallel ...
Making model prediction rasters...
ENMevaluate completed in 13 minutes 34.6 seconds.
/n➡️ Running subset 3 of 18 .../n*** Running initial checks... ***

* Removed 1 occurrence points with NA predictor variable values.
* Removed 4 background points with NA predictor variable values.
* Clamping predictor variable rasters...
* Model evaluations with spatial block (4-fold) cross validation and lat_lon orientation...

*** Running ENMeval v2.0.5 with maxnet from maxnet package v0.1.4 ***


Of 24 total cores using 16...
Running in parallel ...
Making model prediction rasters...
ENMevaluate completed in 11 minutes 20.1 seconds.

Basically, of the 18 subsets, 9 failed and 9 succeeded.
Thus, I tried with executing the first subset (that produced the error) outside the loop, and removing the parallelization, and I got:

vars_subset <- subset_final_bgbias_unique[[1]]
gc()
env_subset <- terra::subset(envs_ca_set_eros, vars_subset)
eval_result_P2 <- ENMevaluate(
  occs = occs_eros[,1:2],
  envs = env_subset,
  bg = bg_bias[,1:2],
  algorithm = "maxent.jar",
  tune.args = tune.args,
  partitions = "block",       # check "randomkfold", "checkerboard2", ecc,
  parallel = F # switch ogg
)
*** Running initial checks... ***

* Removed 1 occurrence points with NA predictor variable values.
* Removed 4 background points with NA predictor variable values.
* Clamping predictor variable rasters...
* Model evaluations with spatial block (4-fold) cross validation and lat_lon orientation...

*** Running ENMeval v2.0.5 with maxnet from maxnet package v0.1.4 ***

  |===================================                                                                    |  34%
Error in mm %*% object$betas: non-conformable arguments



The results were NULL for model with settings fc H, rm 2 for partition 3. These settings will have NA results.
  |================================================                                                       |  47%Errore in mm %*% object$betas : non-conformable arguments

I checked that this could be related to the hinge feature (#57 and #52), and so far, I decided to delete the hinge feature from the tuning arguments (and the loop is proceeding without errors), but I was wondering if this issue was fixed by the time. I specify that the ENMEval version is 2.0.5.

Thanks in advance,
Best,
Davide

[jamiemkass]

Yes, this is a known issue with maxnet that has not been fixed. It seems to occur when there many potential model features, and so removing hinge features is one way to address it. If you use maxent.jar instead, you should not have this issue.