Book driven updates

NAMESPACE

@@ -1,16 +1,19 @@
 # Generated by roxygen2: do not edit by hand
 
 S3method("!=",keras_shape)
+S3method("$",python.builtin.super)
 S3method("$",python_builtin_super_getter)
 S3method("$<-",keras.src.callbacks.callback.Callback)
 S3method("+",keras.src.backend.common.keras_tensor.KerasTensor)
 S3method("==",keras.src.backend.common.keras_tensor.KerasTensor)
 S3method("==",keras_shape)
 S3method("[",keras_shape)
+S3method("[[",python.builtin.super)
 S3method("[[",python_builtin_super_getter)
 S3method(Arg,keras.src.backend.Tensor)
 S3method(Arg,keras.src.backend.common.keras_tensor.KerasTensor)
 S3method(Summary,keras_shape)
+S3method(all,equal.numpy.ndarray)
 S3method(as.array,jax.Array)
 S3method(as.array,jaxlib._jax.ArrayImpl)
 S3method(as.array,jaxlib.xla_extension.ArrayImpl)
@@ -494,6 +497,7 @@ export(metric_sum)
 export(metric_top_k_categorical_accuracy)
 export(metric_true_negatives)
 export(metric_true_positives)
+export(named_list)
 export(new_callback_class)
 export(new_layer_class)
 export(new_learning_rate_schedule_class)
@@ -806,6 +810,7 @@ export(optimizer_sgd)
 export(pad_sequences)
 export(pop_layer)
 export(predict_on_batch)
+export(py_help)
 export(py_require)
 export(py_to_r)
 export(quantize_weights)
@@ -887,6 +892,7 @@ importFrom(reticulate,py_func)
 importFrom(reticulate,py_get_attr)
 importFrom(reticulate,py_get_item)
 importFrom(reticulate,py_has_attr)
+importFrom(reticulate,py_help)
 importFrom(reticulate,py_install)
 importFrom(reticulate,py_is_null_xptr)
 importFrom(reticulate,py_iterator)

NEWS.md

@@ -1,60 +1,88 @@
 # keras3 (development version)
 
-- Added S3 methods for JAX array: `str`, `as.array`, `as.double`, `as.integer`, `as.numeric`.
-
-- Added `str` S3 method for Keras Variables.
-
-- `layer_reshape()` can now accept `-1` as a sentinel for an automatically calculated axis size.
+- Expanded numeric operations with `op_layer_normalization()`, `op_cbrt()`,
+  `op_corrcoef()`, `op_deg2rad()`, `op_heaviside()`, the new `op_sparse_sigmoid()`
+  plus matching `activation_sparse_sigmoid()`, and an `attn_logits_soft_cap`
+  argument for `op_dot_product_attention()`.
 
-- Updated dependencies declared by `use_backend("jax", gpu=TRUE)`
-  for compatability with `keras-hub`.
+- Added signal window operations: `op_bartlett()`, `op_blackman()`,
+  `op_hamming()`, `op_hanning()`, and `op_kaiser()`.
 
-- Added training loop configuration helpers:
-  `config_max_epochs()`, `config_set_max_epochs()`, `config_max_steps_per_epoch()`,
-  and `config_set_max_steps_per_epoch()`. The caps can also be set via the
-  `KERAS_MAX_EPOCHS` and `KERAS_MAX_STEPS_PER_EPOCH` environment variables.
-  Added `config_is_nnx_enabled()` to check whether JAX NNX features are enabled.
+- Added `loss_categorical_generalized_cross_entropy()` for training with noisy
+  labels.
 
 - LoRA-enabled layers (`layer_dense()`, `layer_embedding()`, `layer_einsum_dense()`)
   gain a `lora_alpha` argument to scale the adaptation delta independently of the
   chosen rank.
 
-- `keras_variable()` now accepts a `synchronization` argument for distributed
-  strategies.
+- Added complex-valued helpers: S3 `Arg()` methods for tensors, `op_angle()`,
+  and conversions `op_view_as_real()` / `op_view_as_complex()`.
 
-- `Layer$add_weight()` gains an `overwrite_with_gradient` option and
-  layers now provide a `symbolic_call()` method.
+- Added the Muon optimizer via `optimizer_muon()`.
+
+- Added elastic deformation utilities for images: `layer_random_elastic_transform()`
+  and the lower-level `op_image_elastic_transform()`.
 
 - Transposed convolution utilities now follow the latest Keras API:
   `op_conv_transpose()` defaults `strides = 1` and the `layer_conv_*_transpose()`
   layers expose `output_padding` for precise shape control.
 
-- `layer_torch_module_wrapper()` gains an `output_shape` argument to help Keras
-  infer shapes when wrapping PyTorch modules.
+- `register_keras_serializable()` now returns a registered Python callable,
+  making it easier to use with bare R functions.
 
 - `save_model_weights()` adds a `max_shard_size` argument to split large weight
   files into manageable shards.
 
-- Added elastic deformation utilities for images: `layer_random_elastic_transform()`
-  and the lower-level `op_image_elastic_transform()`.
+- `keras_variable()` now accepts a `synchronization` argument for distributed
+  strategies.
 
-- Added `loss_categorical_generalized_cross_entropy()` for training with noisy
-  labels.
+- `layer_layer_normalization()` removes the `rms_scaling` argument.
 
-- Added the Muon optimizer via `optimizer_muon()`.
+- `layer_reshape()` can now accept `-1` as a sentinel for an automatically calculated axis size.
 
-- Added complex-valued helpers: S3 `Arg()` methods for tensors, `op_angle()`,
-  and conversions `op_view_as_real()` / `op_view_as_complex()`.
+- `layer_torch_module_wrapper()` gains an `output_shape` argument to help Keras
+  infer shapes when wrapping PyTorch modules.
 
-- Added signal window operations: `op_bartlett()`, `op_blackman()`,
-  `op_hamming()`, `op_hanning()`, and `op_kaiser()`.
+- `Layer$add_weight()` gains an `overwrite_with_gradient` option and
+  layers now provide a `symbolic_call()` method.
 
-- Expanded numeric operations with `op_layer_normalization()`, `op_cbrt()`,
-  `op_corrcoef()`, `op_deg2rad()`, `op_heaviside()`, the new `op_sparse_sigmoid()`
-  plus matching `activation_sparse_sigmoid()`, and an `attn_logits_soft_cap`
-  argument for `op_dot_product_attention()`.
+- Added `str()` S3 method for Keras Variables.
 
-- `layer_layer_normalization()` removes the `rms_scaling` argument.
+- Added S3 methods for JAX array:
+  `str()`, `as.array()`, `as.double()`, `as.integer()`, `as.numeric()`.
+
+- Added base-array compatibility methods for backend tensors: `t()`,
+  `aperm()`, and `all.equal()`.
+
+- Added `pillar::type_sum()` for JAX variables and `JaxVariable`;
+  extended `str()` coverage to the new JAX variable class.
+
+- `config_max_epochs()`, `config_set_max_epochs()`, `config_max_steps_per_epoch()`,
+  and `config_set_max_steps_per_epoch()`. The caps can also be set via the
+  `KERAS_MAX_EPOCHS` and `KERAS_MAX_STEPS_PER_EPOCH` environment variables.
+  Added `config_is_nnx_enabled()` to check whether JAX NNX features are enabled.
+
+- Built-in dataset loaders now accept `convert = FALSE` to return NumPy arrays
+  instead of R arrays.
+
+- Updated `plot(history, theme_bw = TRUE)` for `ggplot2` 3.4.0
+  compatibility.
+
+- `plot(model)` DPI is now globally configurable via
+  `options(keras.plot.model.dpi = )`, (defaults to  `200`).
+
+- Reexported reticulate functions: `py_help()`, `py_to_r()`, `r_to_py()`,
+  `py_require()`, and `import()`.
+
+- Support `super()$initialize()` in subclassed Keras classes; improved
+  `super()` behavior in subclasses.
+
+- Updated dependencies declared by `use_backend("jax", gpu=TRUE)`
+  for compatability with `keras-hub`.
+
+- Exported `named_list()` utility.
+
+- Fixed an issue when switching backends twice in a row.
 
 # keras3 1.4.0
 

R/install.R

@@ -388,7 +388,7 @@ uv_unset_override_never_tensorflow <- function() {
   if (is.na(override)) return()
   cpu_override <- pkg_file("never-tensorflow-override.txt")
   if (override == cpu_override) {
-    Sys.unsetenv(override)
+    Sys.unsetenv("UV_OVERRIDE")
   } else {
     new <- gsub(cpu_override, "", override, fixed = TRUE)
     new <- gsub(" +", " ", new)

R/model-persistence.R

@@ -585,8 +585,8 @@ function (object, filepath, call_endpoint = "serve", call_training_endpoint = NU
 #' @param object
 #' A keras object.
 #'
-#' @returns `object` is returned invisibly, for convenient piping. This is
-#'   primarily called for side effects.
+#' @returns The registered `object` (and converted) is returned. This returned object is what you
+#' should must use when building and serializing the model.
 #' @export
 #' @family saving and loading functions
 #' @family serialization utilities
@@ -605,7 +605,7 @@ function (object, name = NULL, package = NULL)
                 c("", "base", "R_GlobalEnv"), "Custom")
 
   keras$saving$register_keras_serializable(package, name)(py_object)
-  invisible(object)
+  py_object
 }
 
 

R/package.R

@@ -186,7 +186,12 @@ keras <- NULL
     keras <- import("keras")
     convert_to_tensor <- import("keras.ops", convert = FALSE)$convert_to_tensor
     with(keras$device("cpu:0"), {
-      backend_tensor_class <- class(convert_to_tensor(array(1L)))[1L]
+      all_backend_tensor_s3_classes <- class(convert_to_tensor(array(1L)))
+      backend_tensor_class <- all_backend_tensor_s3_classes[1L]
+      if ("jax.Array" %in% all_backend_tensor_s3_classes)
+        backend_tensor_class <- "jax.Array"
+      # message("setting methods on backend_tensor_class: ", backend_tensor_class,
+      #         "\nother options: ", paste0(all_backend_tensor_s3_classes, collapse = " "))
     })
     symbolic_tensor_class <- nameOfClass__python.builtin.type(keras$KerasTensor)
 
@@ -207,6 +212,9 @@ keras <- NULL
     registerS3method("as.array", backend_tensor_class, op_convert_to_array, baseenv())
     registerS3method("^", backend_tensor_class, `^__keras.backend.tensor`, baseenv())
     registerS3method("%*%", backend_tensor_class, op_matmul, baseenv())
+    registerS3method("t", backend_tensor_class, op_transpose, baseenv())
+    registerS3method("aperm", backend_tensor_class, op_transpose, baseenv())
+    registerS3method("all.equal", backend_tensor_class, all.equal.numpy.ndarray, baseenv())
 
     if(keras$config$backend() == "jax") {
       for(py_type in import("jax")$Array$`__subclasses__`()) {
@@ -271,6 +279,13 @@ keras <- NULL
 
 }
 
+## should this live in reticulate?? probably...
+#' @export
+all.equal.numpy.ndarray <- function(target, current, ...) {
+  # or use numpy.allequal?
+  all.equal(as.array(target), as.array(current), ...)
+}
+
 
 at.keras_backend_tensor <-  function(object, name) {
   out <- rlang::env_clone(object)

R/py-classes.R

@@ -119,12 +119,16 @@ function(classname,
       type = `__class__`,
       object_or_type = base::get("self", envir = base::parent.frame()))
       {
-        convert <- base::get("convert", envir = base::as.environment(object_or_type))
-        py_builtins <- reticulate::import_builtins(convert)
-        reticulate::py_call(py_builtins$super, type, object_or_type)
-      }
+      convert <- base::get("convert", object_or_type)
+      py_super <- reticulate::py_eval(
+        "__import__('builtins').super",
+        convert = convert
+      )
+      py_super(type, object_or_type)
+    }
     class(super) <- "python_builtin_super_getter"
-  }))
+  })
+)
 
 
   py_class
@@ -137,15 +141,23 @@ function(classname,
 #' @export
 `$.python_builtin_super_getter` <- function(x, name) {
   super <- do.call(x, list(), envir = parent.frame()) # call super()
+  `[[.python.builtin.super`(super, name)
+}
+
+#' @export
+`[[.python_builtin_super_getter` <- `$.python_builtin_super_getter`
+
+#' @export
+`$.python.builtin.super` <- function(x, name) {
   name <- switch(name, initialize = "__init__", finalize = "__del__", name)
-  out <- py_get_attr(super, name)
+  out <- py_get_attr(x, name)
   convert <- get0("convert", as.environment(out), inherits = FALSE,
                   ifnotfound = TRUE)
   if (convert) py_to_r(out) else out
 }
 
 #' @export
-`[[.python_builtin_super_getter` <- `$.python_builtin_super_getter`
+`[[.python.builtin.super` <- `$.python.builtin.super`
 
 # No .DollarNames.python_builtin_super_getter because the python.builtin.super
 # object doesn't have populated attributes itself, only a dynamic `__getattr__`

R/r-utils.R

@@ -40,15 +40,34 @@ drop_nulls <- function(x, i = NULL) {
   x[!drop]
 }
 
+#' Create a named list from arguments
+#'
+#' Constructs a list from the provided arguments where all elements are named.
+#' This wraps [rlang::dots_list()] but changes two defaults:
+#' - `.named` is set to `TRUE`
+#' - `.homonyms` is set to `"error"`
+#'
+#' Other parameters retain their defaults from [rlang::dots_list()]:
+#' - `.ignore_empty = "trailing"`
+#' - `.preserve_empty = FALSE`
+#' - `.check_assign = FALSE`
+#'
+#' @inheritParams rlang::dots_list
+#'
+#' @inheritParams dots_list
+#'
+#' @return A named list.
+#'
+#' @seealso [rlang::dots_list()]
+#'
+#' @export
 #' @importFrom rlang dots_list
-# identical to rlang::list2(), except .named = TRUE
 named_list <- function(...)
   dots_list(...,
-            .named = TRUE,
-            # not the default
+            .named = TRUE, # not default
             .ignore_empty = "trailing",
             .preserve_empty = FALSE,
-            .homonyms = "error",
+            .homonyms = "error", # not default
             .check_assign = FALSE)
 
 `append1<-` <- function(x, value) {

R/reexports.R

@@ -77,6 +77,9 @@ reticulate::py_to_r
 #' @export
 reticulate::r_to_py
 
+#' @export
+reticulate::py_help
+
 #' @importFrom tensorflow tensorboard
 #' @export
 tensorflow::tensorboard

R/utils.R

@@ -369,6 +369,8 @@ to_categorical <-
 function (x, num_classes = NULL)
 {
   if (inherits(x, "factor")) {
+    # if (length(DIM(x)) == 1)
+    #    return(diag(nrow = num_classes %||% length(levels(x)))[as.integer(x), ])
     x <- array(as.integer(x) - 1L, dim = dim(x) %||% length(x))
     if (is.null(num_classes))
       num_classes <- length(levels(x))
@@ -631,7 +633,7 @@ function(x,
          ...,
          rankdir = "TB",
          expand_nested = FALSE,
-         dpi = 200,
+         dpi = getOption("keras.plot.model.dpi", 200L),
          layer_range = NULL,
          show_layer_activations = FALSE,
          show_trainable = NA,