Bug report: `remove_keras_spec` deletes multiple models if their names share a common prefix

[davidrsch]

Description

The function remove_keras_spec has an issue where it removes more models than intended. When creating the to_kill vector, it deletes all objects whose names start with the name of the model to be deleted. This can cause problems if a user has defined two model specs, for example, my_mlp and my_mlp_2. When remove_keras_spec is called to delete my_mlp, it will also delete my_mlp_2 because its name also starts with "my_mlp".

A more specific approach is needed to create the to_kill vector. Instead of using a broad pattern, it should use a specific list of suffixes to identify the objects to be removed.

Reprex

# Load necessary libraries
library(kerasnip)
#> Warning: package 'kerasnip' was built under R version 4.5.1
library(parsnip)
#> 
#> Attaching package: 'parsnip'
#> The following object is masked from 'package:kerasnip':
#> 
#>     get_model_env
library(keras3)
#> Warning: package 'keras3' was built under R version 4.5.1

# Create first model spec
create_keras_sequential_spec(
  "my_mlp",
  list(
    input = \(model, input_shape) {
      keras_model_sequential(input_shape = input_shape)
    },
    dense = \(model, units = 16) model |> layer_dense(units = units)
  ),
  "regression"
)

# Create second model spec
create_keras_sequential_spec(
  "my_mlp_2",
  list(
    input = \(model, input_shape) {
      keras_model_sequential(input_shape = input_shape)
    },
    dense = \(model, units = 32) model |> layer_dense(units = units)
  ),
  "regression"
)

# Both models should exist
print(paste("my_mlp exists:", exists("my_mlp")))
#> [1] "my_mlp exists: TRUE"
print(paste("my_mlp_2 exists:", exists("my_mlp_2")))
#> [1] "my_mlp_2 exists: TRUE"

# Remove the first model
remove_keras_spec("my_mlp")
#> Removed from parsnip registry objects: my_mlp, my_mlp_2, my_mlp_2_args, my_mlp_2_encoding, my_mlp_2_fit, my_mlp_2_modes, my_mlp_2_pkgs, my_mlp_2_predict, my_mlp_args, my_mlp_encoding, my_mlp_fit, my_mlp_modes, my_mlp_pkgs, my_mlp_predict
#> Removed 'my_mlp' from parsnip:::get_model_env()$models

# Check if the second model still exists
print(paste("my_mlp exists after removal:", exists("my_mlp")))
#> [1] "my_mlp exists after removal: FALSE"
print(paste("my_mlp_2 exists after removal:", exists("my_mlp_2")))
#> [1] "my_mlp_2 exists after removal: TRUE"

# my_mlp_2 exist in the models list of the parsnip environment but the
# associated objects were deleted as informed by `remove_keras_spec`

Session Info

sessionInfo()
#> R version 4.5.0 (2025-04-11 ucrt)
#> Platform: x86_64-w64-mingw32/x64
#> Running under: Windows 10 x64 (build 19045)
#> 
#> Matrix products: default
#>   LAPACK version 3.12.1
#> 
#> locale:
#> [1] LC_COLLATE=English_United Kingdom.utf8 
#> [2] LC_CTYPE=English_United Kingdom.utf8   
#> [3] LC_MONETARY=English_United Kingdom.utf8
#> [4] LC_NUMERIC=C                           
#> [5] LC_TIME=English_United Kingdom.utf8    
#> 
#> time zone: Europe/Madrid
#> tzcode source: internal
#> 
#> attached base packages:
#> [1] stats     graphics  grDevices utils     datasets  methods   base     
#> 
#> other attached packages:
#> [1] keras3_1.4.0       parsnip_1.3.1      kerasnip_0.0.2.900
#> 
#> loaded via a namespace (and not attached):
#>  [1] gtable_0.3.6        xfun_0.52           ggplot2_3.5.2      
#>  [4] recipes_1.3.0       lattice_0.22-6      vctrs_0.6.5        
#>  [7] tools_4.5.0         tfruns_1.5.3        generics_0.1.3     
#> [10] parallel_4.5.0      tibble_3.2.1        pkgconfig_2.0.3    
#> [13] Matrix_1.7-3        data.table_1.17.0   lifecycle_1.0.4    
#> [16] compiler_4.5.0      munsell_0.5.1       codetools_0.2-20   
#> [19] htmltools_0.5.8.1   class_7.3-23        yaml_2.3.10        
#> [22] prodlim_2024.06.25  pillar_1.10.2       whisker_0.4.1      
#> [25] tidyr_1.3.1         MASS_7.3-65         gower_1.0.2        
#> [28] rpart_4.1.24        parallelly_1.43.0   lava_1.8.1         
#> [31] tidyselect_1.2.1    digest_0.6.37       future_1.40.0      
#> [34] dplyr_1.1.4         purrr_1.0.4         listenv_0.9.1      
#> [37] splines_4.5.0       fastmap_1.2.0       grid_4.5.0         
#> [40] colorspace_2.1-1    cli_3.6.4           magrittr_2.0.3     
#> [43] base64enc_0.1-3     survival_3.8-3      dotty_0.1.0        
#> [46] future.apply_1.11.3 withr_3.0.2         scales_1.3.0       
#> [49] lubridate_1.9.4     timechange_0.3.0    rmarkdown_2.29     
#> [52] globals_0.17.0      nnet_7.3-20         timeDate_4041.110  
#> [55] reticulate_1.42.0   png_0.1-8           evaluate_1.0.3     
#> [58] knitr_1.50          hardhat_1.4.1       rlang_1.1.6        
#> [61] Rcpp_1.0.14         zeallot_0.2.0       glue_1.8.0         
#> [64] ipred_0.9-15        reprex_2.1.1        jsonlite_2.0.0     
#> [67] R6_2.6.1            fs_1.6.6            tensorflow_2.16.0

Expected Behavior

When remove_keras_spec("my_mlp") is called, only the my_mlp model and its associated objects should be removed. The my_mlp_2 model and associated objects should remain untouched.

Actual Behavior

When remove_keras_spec("my_mlp") is called, objects from both my_mlp and my_mlp_2 models are removed from the parsnip registry. This is because the function uses a broad grep pattern that matches any model name starting with "my_mlp".

Additional Context

The issue is located in the remove_keras_spec.R file, specifically in the line where the to_kill vector is created: to_kill <- grep(paste0("^", model_name), all_regs, value = TRUE). This should be replaced with a more specific pattern to avoid unintended deletions.