The hardware and bandwidth for this mirror is donated by dogado GmbH, the Webhosting and Full Service-Cloud Provider. Check out our Wordpress Tutorial.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]dogado.de.

kindling (development version)

Ongoing features

• More visualization supports

  • NN architecture plot

kindling 0.3.2

Bug fixes

• "linear" used as an activations/output_activation value no longer crashes at training time. It was previously resolved to torch::nnf_linear() (an affine transform requiring its own weight/bias), instead of behaving as an identity/no-op activation. "linear" now consistently maps to identity() (#21).

Documentation

• Added missing @examples for act_funs(), args(), and early_stop().

• Fixed the {tidymodels} example: loading Ionosphere via box::use(mlbench[Ionosphere]) failed because {mlbench} does not export its datasets through NAMESPACE/LazyData. The example now uses data(Ionosphere, package = "mlbench") instead.

• Usage examples from README gets transferred to vignettes/kindling.Rmd.

kindling 0.3.1

New features

• autoplot() and plot() methods for nn_fit objects now visualize the training loss history, with optional validation loss and an early-stopping marker when early stopping fires.

• autoplot_diagnostics() and plot_diagnostics() produce prediction diagnostic plots for nn_fit objects: residuals vs fitted and actual vs fitted panels for regression, one panel per output for multi-output regression, and a confusion matrix heatmap for classification.

• {vip} is removed from {kindling}’s package dependencies, and moved to “Suggests” instead to avoid dependency redundancy.

kindling 0.3.0

New Experimental functions

• Generalized nn_module() expression generator to generate torch::nn_module() expression for the same sequential NN architectures

  • This is how you use to generate nn_module() for 1D-CNN (Convolutional Neural Networks) with 3 hidden layers:

  nn_module_generator(
      nn_name = "CNN1DClassifier",
      nn_layer = "nn_conv1d",
      layer_arg_fn = ~ if (.is_output) {
          list(.in, .out)
      } else {
          list(
              in_channels = .in,
              out_channels = .out,
              kernel_size = 3L,
              stride = 1L,
              padding = 1L 
          )
      },
      after_output_transform = ~ .$mean(dim = 2),
      last_layer_args = list(kernel_size = 1, stride = 2),
      hd_neurons = c(16, 32, 64),
      no_x = 1,
      no_y = 10,
      activations = "relu"
  )

• train_nn() to execute nn_module_generator()

  • nn_arch() must be supplied to inherit extra arguments from nn_module_generator() function.
  • Allows early stopping if early_stopping is supplied with early_stop().
  • Supported with several data types: matrix, data.frame, dataset ({torch} dataset), and a formula interface.
  • train_nnsnip() is now provided to bridge train_nn() with {tidymodels}

• You can supply customized activation function under act_funs() with new_act_fn().

  • Activation functions that especially don’t exist on torch::nnf_*().
  • Supply the argument with a function
  • The function supplied into new_act_fn() must return a torch tensor object.
  • Example: act_funs(new_act_fn(torch::torch_tanh)) or act_funs(new_act_fn(\(x) torch::torch_tanh(x)))
  • Use .name as a displayed name of the custom activation function.

Superset

• act_funs() as a DSL function now supports index-style parameter specification for parametric activation functions

  • Activation functions can now be modified using [ syntax (e.g. softplus[beta = 0.2])
  • The current args() (e.g. softplus = args(beta = 0.2)) is now superseded by that.

Bug Fixes

• No suffix generated for 13 by ordinal_gen(). Now fixed.

• hd_neurons for both ffnn_generator() and rnn_generator() accepts empty arguments, which implies there’s no hidden layers applied.

kindling 0.2.0

New features

• Added regularization support for neural network models

  • L1 regularization (Lasso) for feature selection via mixture = 1
  • L2 regularization (Ridge) for weight decay via mixture = 0
  • Elastic Net combining L1 and L2 penalties via 0 < mixture < 1
  • Controlled via penalty (regularization strength) and mixture (L1/L2 balance) parameters
  • Follows tidymodels conventions for consistency with glmnet and other packages

• n_hlayers() now fully supports tuning the number of hidden layers

• hidden_neurons() gains support for discrete values via the disc_values argument

  • e.g. disc_values = c(32L, 64L, 128L, 256L)) is now allowed
  • This allows tuning over specific common hidden unit sizes instead of (or in addition to) a continuous range

Implementation fixes

• Tuning methods and grid_depth() is now fixed

  • Parameter space for the number of hidden layers is now fixed and active
  • Corrected parameter space handling for n_hlayers (no more invalid sampling when x > 1)
  • Uses tidyr::expand_grid(), not purrr::cross*()
  • Fix randomization of parameter space which will produce NAs outside from {kindling}‘s own ’dials’
  • No more list columns when n_hlayers = 1

• The supported models now use hardhat::mold(), instead of model.frame() and model.matrix().

Documentation

• Add a vignette to showcase the comparison with other similar packages

• The package description has been clarified

• Vignette to showcase the comparison with other similar packages

• hidden_neurons parameter now supports discrete values specification

  • Users can specify exact neuron counts via values parameter (e.g., hidden_neurons(values = c(32, 64, 128)))
  • Maintains backward compatibility with range-based parameters (e.g., hidden_neurons(range = c(8L, 512L)) / hidden_neurons(c(8L, 512L)))

• Added \value documentation to kindling-nn-wrappers for CRAN compliance

• Documented argument handling and list-column unwrapping in tidymodels wrapper functions

• Clarified the relationship between grid_depth() and wrapper functions

kindling 0.1.0

• Initial CRAN release
• Higher-level interface for torch package to define, train, and tune neural networks
• Support for feedforward (multi-layer perceptron) and recurrent networks (RNN, LSTM, GRU)
• Integration with tidymodels ecosystem (parsnip, workflows, recipes, tuning)
• Variable importance plots and network visualization tools

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
Health stats visible at Monitor.