normalize_shape() accepts tensors

Author: t-kalinowski

NEWS.md

@@ -37,6 +37,9 @@
   Standard layer arguments include: `input_shape`, `batch_input_shape`, `batch_size`, `dtype`, `name`, `trainable`, `weights`.
   Layers updated: `layer_global_{max,average}_pooling_{1,2,3}d()`, `time_distributed()`, `bidirectional()`.
 
+- All the backend function with a shape argument `k_*(shape =)` that now accept a 
+  a mix of integer tensors and R numerics in the supplied list.
+
 - `k_random_uniform()` now automatically coerces `minval` and `maxval` to the output dtype.
 
 # keras 2.6.1

R/backend.R

@@ -3103,10 +3103,5 @@ backend_normalize_shape <- function(shape) {
   if (inherits(shape, "python.builtin.object"))
     return(shape)
 
-  if (is.list(shape)) {
-    if (any(sapply(unlist(shape), function(x) inherits(x, "python.builtin.object"))))
-      return(shape)
-  }
-
   normalize_shape(shape)
 }

R/layers-core.R

@@ -433,6 +433,13 @@ normalize_shape <- function(shape) {
   # if it's a list or a numeric vector then convert to integer
   if (is.list(shape) || is.numeric(shape)) {
     shape <- lapply(shape, function(value) {
+
+      # Pass through python objects unmodified, only coerce R objects
+      # supplied shapes, e.g., to tf$random$normal, can be a list that's a mix
+      # of scalar integer tensors and regular integers
+      if(inherits(value, "python.builtin.object"))
+        return(value)
+
       if (!is.null(value))
         as.integer(value)
       else

tests/testthat/test-backend.R

@@ -132,3 +132,12 @@ test_backend("k_set_epsilon", {
   k_set_epsilon(1e-07)
   expect_equal(k_epsilon(), 1e-07)
 })
+
+
+test_backend("k_random_normal", {
+  chk <- function(x) expect_tensor(x, shape = c(1L, 2L, 3L))
+  chk(k_random_normal(c(1,2,3)))
+  chk(k_random_normal(as_tensor(c(1,2,3), "int32")))
+  chk(k_random_normal(lapply(c(1,2,3), as_tensor, "int32")))
+  chk(k_random_normal(list(1, as_tensor(2, "int32"), 3)))
+})

vignettes/new-guides/customizing_what_happens_in_fit.Rmd

@@ -356,8 +356,6 @@ GAN(keras$Model) %py_class% {
     # Sample random points in the latent space
     batch_size <- tf$shape(real_images)[1]
 
-    # TODO: shape() should be able to handle a scalar tensor as input
-    # also, backend functions like k_random_normal
     random_latent_vectors <- tf$random$normal(list(batch_size, self$latent_dim))
 
     # Decode them to fake images