fixes confusing source of value max for channel input normalization

Author: grzuy

examples/generative/fashionmnist_autoencoder.exs

@@ -11,6 +11,7 @@ defmodule FashionMNIST do
   @batch_size 32
   @image_channels 1
   @image_side_pixels 28
+  @channel_value_max 255
 
   defmodule Autoencoder do
     @image_channels 1
@@ -39,7 +40,7 @@ defmodule FashionMNIST do
     bin
     |> Nx.from_binary(type)
     |> Nx.reshape({elem(shape, 0), @image_channels, @image_side_pixels, @image_side_pixels})
-    |> Nx.divide(Nx.Constants.max(type))
+    |> Nx.divide(@channel_value_max)
     |> Nx.to_batched(@batch_size)
   end
 

examples/generative/mnist_gan.exs

@@ -16,12 +16,13 @@ defmodule MNISTGAN do
   @batch_size 32
   @image_channels 1
   @image_side_pixels 28
+  @channel_value_max 255
 
   defp transform_images({bin, type, shape}) do
     bin
     |> Nx.from_binary(type)
     |> Nx.reshape({elem(shape, 0), @image_channels, @image_side_pixels, @image_side_pixels})
-    |> Nx.divide(Nx.Constants.max(type))
+    |> Nx.divide(@channel_value_max)
     |> Nx.to_batched(@batch_size)
   end
 

examples/vision/cifar10.exs

@@ -11,13 +11,14 @@ defmodule Cifar do
   @batch_size 32
   @image_channels 3
   @image_side_pixels 32
+  @channel_value_max 255
   @label_values Enum.to_list(0..9)
 
   defp transform_images({bin, type, shape}) do
     bin
     |> Nx.from_binary(type)
     |> Nx.reshape({elem(shape, 0), @image_side_pixels, @image_side_pixels, @image_channels})
-    |> Nx.divide(Nx.Constants.max(type))
+    |> Nx.divide(@channel_value_max)
     |> Nx.to_batched(@batch_size)
     |> Enum.split(1500)
   end

examples/vision/cnn_image_denoising.exs

@@ -14,6 +14,7 @@ defmodule MnistDenoising do
   @epochs 25
   @image_channels 1
   @image_side_pixels 28
+  @channel_value_max 255
 
   def run do
     {images, _} = Scidata.MNIST.download()
@@ -49,7 +50,7 @@ defmodule MnistDenoising do
     bin
     |> Nx.from_binary(type)
     |> Nx.reshape({elem(shape, 0), @image_side_pixels, @image_side_pixels, @image_channels})
-    |> Nx.divide(Nx.Constants.max(type))
+    |> Nx.divide(@channel_value_max)
     |> Nx.to_batched_list(@batch_size)
     # Test split
     |> Enum.split(1750)

examples/vision/horses_or_humans.exs

@@ -18,7 +18,7 @@ defmodule HorsesOrHumans do
   @batch_size 32
   @image_channels 4
   @image_side_pixels 300
-  @input_type {:u, 8}
+  @channel_value_max 255
 
   def data() do
     Path.wildcard(@directories)
@@ -31,9 +31,9 @@ defmodule HorsesOrHumans do
     |> Stream.cycle()
   end
 
-  defnp augment(inp) do
+  defnp augment(%Nx.Tensor{type: type} = inp) do
     # Normalize
-    inp = inp / Nx.Constants.max(@input_type)
+    inp = inp / @channel_value_max
 
     # For now just a random flip
     if Nx.random_uniform({}) > 0.5 do
@@ -46,8 +46,8 @@ defmodule HorsesOrHumans do
   defp parse_png(filename) do
     class =
       if String.contains?(filename, "horses"),
-        do: Nx.tensor([1, 0], type: @input_type),
-        else: Nx.tensor([0, 1], type: @input_type)
+        do: Nx.tensor([1, 0], type: {:u, 8}),
+        else: Nx.tensor([0, 1], type: {:u, 8})
 
     {:ok, img} = StbImage.read_file(filename)
 

examples/vision/mnist.exs

@@ -11,13 +11,14 @@ defmodule Mnist do
 
   @batch_size 32
   @image_side_pixels 28
+  @channel_value_max 255
   @label_values Enum.to_list(0..9)
 
   defp transform_images({bin, type, shape}) do
     bin
     |> Nx.from_binary(type)
     |> Nx.reshape({elem(shape, 0), @image_side_pixels**2})
-    |> Nx.divide(Nx.Constants.max(type))
+    |> Nx.divide(@channel_value_max)
     |> Nx.to_batched(@batch_size)
     # Test split
     |> Enum.split(1750)