Merge branch 'main' of https://github.com/rstudio/keras into main

Author: t-kalinowski

NEWS.md

@@ -1,5 +1,8 @@
 # keras (development version)
 
+- Fixed issue where `input_shape` supplied to custom layers defined with `new_layer_class()`  
+  would result in an error (#1338)
+
 - New `callback_backup_and_restore()`, for resuming an interrupted `fit()` call.
 
 - The merging family of layers (`layer_add`, `layer_concatenate`, etc.) gain the ability

R/install.R

@@ -33,10 +33,7 @@ install_keras <- function(method = c("auto", "virtualenv", "conda"),
       method = method,
       conda = conda,
       version = version,
-      extra_packages = c("pandas", "Pillow",
-                         "pydot",
-                         "tensorflow-hub",
-                         "tensorflow-datasets",
+      extra_packages = c(default_extra_packages(),
                          extra_packages),
       ...))
   }
@@ -66,6 +63,7 @@ default_version <- numeric_version("2.9")
 default_extra_packages <- function(tensorflow_version = "default") {
   pkgs <- c(
     "tensorflow-hub",
+    "tensorflow-datasets",
     "scipy",
     "requests",
     "pyyaml",

R/layer-custom.R

@@ -196,7 +196,16 @@ py_formals <- function(py_obj) {
 create_layer_wrapper <- function(Layer, modifiers = NULL, convert = TRUE) {
 
   force(Layer)
-  force(modifiers)
+  modifiers <- utils::modifyList(
+    list(
+      # include helpers for standard layer args by default,
+      # but leave an escape hatch allowing users to override/opt-out.
+      input_shape = as_tf_shape,
+      batch_input_shape = as_tf_shape,
+      batch_size = as.integer
+    ),
+    as.list(modifiers)
+  )
 
   wrapper <- function(object) {
     args <- capture_args(match.call(), modifiers, ignore = "object")
@@ -241,3 +250,11 @@ r_to_py.keras_layer_wrapper <- function(fn, convert = FALSE) {
     layer <- r_to_py(layer, convert)
   layer
 }
+
+
+as_tf_shape <- function (x) {
+  if (inherits(x, "tensorflow.python.framework.tensor_shape.TensorShape"))
+    x
+  else
+    shape(dims = x)
+}

R/layer-methods.R

@@ -12,6 +12,8 @@
 #'
 #' @param object Layer or model object
 #' @param config Object with layer or model configuration
+#' @param custom_objects list of custom objects needed to instantiate the layer,
+#'   e.g., custom layers defined by `new_layer_class()` or similar.
 #'
 #' @return `get_config()` returns an object with the configuration,
 #'   `from_config()` returns a re-instantiation of the object.

man/get_config.Rd

@@ -1,4 +1,8 @@
-# Sys.setenv(TF_CPP_MIN_LOG_LEVEL = 1)
+Sys.setenv(TF_CPP_MIN_LOG_LEVEL = 1)
+# 0 = all messages are logged (default behavior)
+# 1 = INFO messages are not printed
+# 2 = INFO and WARNING messages are not printed
+# 3 = INFO, WARNING, and ERROR messages are not printed
 
 
 # Sys.setenv(RETICULATE_PYTHON = "~/.local/share/r-miniconda/envs/tf-2.7-cpu/bin/python")

man/install_keras.Rd

@@ -197,14 +197,11 @@ test_succeeds("create_layer_wrapper", {
 })
 
 
-
-
-
 test_succeeds("create_layer_wrapper", {
 
   layer_sampler <- new_layer_class(
     classname = "Sampler",
-    call = function(self, z_mean, z_log_var) {
+    call = function(z_mean, z_log_var) {
       epsilon <-  k_random_normal(shape = k_shape(z_mean))
       z_mean + exp(0.5 * z_log_var) * epsilon
     }
@@ -217,3 +214,43 @@ test_succeeds("create_layer_wrapper", {
   expect_equal(dim(res), c(128, 2))
 
 })
+
+
+test_succeeds("custom layers can accept standard layer args like input_shape", {
+  # https://github.com/rstudio/keras/issues/1338
+  layer_simple_dense <- new_layer_class(
+    classname = "SimpleDense",
+
+    initialize = function(units, activation = NULL, ...) {
+      super$initialize(...)
+      self$units <- as.integer(units)
+      self$activation <- activation
+    },
+
+    build = function(input_shape) {
+      input_dim <- input_shape[length(input_shape)]
+      self$W <- self$add_weight(shape = c(input_dim, self$units),
+                                initializer = "random_normal")
+      self$b <- self$add_weight(shape = c(self$units),
+                                initializer = "zeros")
+    },
+
+    call = function(inputs) {
+      y <- tf$matmul(inputs, self$W) + self$b
+      if (!is.null(self$activation))
+        y <- self$activation(y)
+      y
+    }
+  )
+
+  model <- keras_model_sequential() %>%
+    layer_simple_dense(20, input_shape = 30) %>%
+    layer_dense(10)
+
+  expect_identical(dim(model$input), c(NA_integer_, 30L))
+  expect_true(model$built)
+
+  res <- model(random_array(1, 30))
+  expect_tensor(res, shape = c(1L, 10L))
+
+})