Fix group normalization (#477)

elixir-nx · Feb 17, 2023 · 4fac82b · 4fac82b
1 parent 884107c
commit 4fac82b
Show file tree

Hide file tree

Showing 5 changed files with 145 additions and 6 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,11 @@
 # Changelog
 
+## v0.5.1 (2022-02-17)
+
+### Bug Fixes
+
+* Fixed incorrect results from group normalization
+
 ## v0.5.0 (2022-02-16)
 
 ### Enhancements

diff --git a/lib/axon/layers.ex b/lib/axon/layers.ex
@@ -1207,24 +1207,29 @@ defmodule Axon.Layers do
   defn group_norm(input, gamma, beta, opts \\ []) do
     opts = keyword!(opts, [:num_groups, epsilon: 1.0e-5, channel_index: -1, mode: :inference])
 
+    channel_axis = normalize_group_norm_channel_axis(input, opts[:channel_index])
+
     group_shape = Axon.Shape.group_norm_shape(input, opts[:num_groups], opts[:channel_index])
     num_channels = Nx.axis_size(input, opts[:channel_index])
 
     parameter_shape = norm_parameter_reshape(input, num_channels, opts[:channel_index])
-
     gamma = Nx.reshape(gamma, parameter_shape)
     beta = Nx.reshape(beta, parameter_shape)
 
     x = Nx.reshape(input, group_shape)
 
-    axes = Axon.Shape.group_norm_axes(x, opts[:channel_index])
+    axes = Axon.Shape.group_norm_axes(x, channel_axis)
 
     {mean, var} = mean_and_variance(x, axes: axes)
     x = (x - mean) * Nx.rsqrt(var + opts[:epsilon])
     x = Nx.reshape(x, input)
     x * gamma + beta
   end
 
+  deftransformp normalize_group_norm_channel_axis(input, channel_index) do
+    Nx.Shape.normalize_axis(Nx.shape(input), channel_index, Nx.shape(input))
+  end
+
   @doc ~S"""
   Functional implementation of instance normalization.
 

diff --git a/lib/axon/shape.ex b/lib/axon/shape.ex
@@ -942,16 +942,16 @@ defmodule Axon.Shape do
   @doc """
   Calculates the reduction axes for group normalization.
   """
-  deftransform group_norm_axes(input, channel_index) do
-    Enum.to_list(1..(Nx.rank(input) - 1)) -- [channel_index]
+  deftransform group_norm_axes(x, channel_index) do
+    Enum.to_list(1..(Nx.rank(x) - 1)) -- [channel_index]
   end
 
   @doc """
   Calculates the reshape for group normalization.
   """
   deftransform group_norm_shape(input, num_groups, channel_index) do
     shape = Nx.shape(input)
-    channel_index = Nx.Shape.normalize_axis(shape, channel_index, Nx.names(input))
+    channel_index = Nx.Shape.normalize_axis(Nx.shape(input), channel_index, Nx.names(input))
 
     channels = elem(shape, channel_index)
     group_size = div(channels, num_groups)

diff --git a/mix.exs b/mix.exs
@@ -2,7 +2,7 @@ defmodule Axon.MixProject do
   use Mix.Project
 
   @source_url "https://github.com/elixir-nx/axon"
-  @version "0.5.0"
+  @version "0.5.1"
 
   def project do
     [

diff --git a/test/axon/layers_test.exs b/test/axon/layers_test.exs
@@ -1226,4 +1226,132 @@ defmodule Axon.LayersTest do
       )
     end
   end
+
+  describe "group_norm" do
+    test "matches pytorch" do
+      a =
+        Nx.tensor([
+          [
+            0.8423,
+            1.9226,
+            -1.1295,
+            -1.3154,
+            1.2963,
+            -0.6821,
+            -0.0519,
+            0.6875,
+            -0.0313,
+            -0.3328,
+            -0.2821,
+            -2.3289,
+            -1.7641,
+            -1.3184,
+            -0.0890,
+            0.0625
+          ],
+          [
+            -1.0853,
+            0.8060,
+            -0.1397,
+            -0.2169,
+            0.9605,
+            0.3947,
+            0.4760,
+            0.8097,
+            0.0380,
+            -0.6314,
+            0.5761,
+            1.9309,
+            0.5038,
+            -0.1892,
+            1.8476,
+            0.0517
+          ]
+        ])
+
+      b =
+        Nx.tensor([
+          -0.3101,
+          -1.5896,
+          -1.4963,
+          0.1278,
+          -1.4580,
+          1.3832,
+          0.5709,
+          0.5531,
+          -0.0588,
+          1.0411,
+          1.3503,
+          -1.2166,
+          0.7133,
+          0.0694,
+          0.3150,
+          -0.1306
+        ])
+
+      c =
+        Nx.tensor([
+          1.6585,
+          2.3515,
+          -1.3456,
+          0.2376,
+          -0.1333,
+          0.5068,
+          0.2441,
+          1.0382,
+          0.6879,
+          -0.5402,
+          -1.8304,
+          -0.8906,
+          -0.5329,
+          -0.3390,
+          -0.1877,
+          0.1405
+        ])
+
+      expected =
+        Nx.tensor([
+          [
+            1.4768,
+            -0.1375,
+            0.4536,
+            0.0623,
+            -1.5881,
+            -0.5951,
+            0.1157,
+            1.2847,
+            0.6378,
+            -0.0194,
+            -1.0751,
+            1.3407,
+            -1.3700,
+            -0.3844,
+            0.0597,
+            0.0149
+          ],
+          [
+            2.2986,
+            0.9877,
+            -0.4434,
+            0.1453,
+            -1.7321,
+            0.8146,
+            0.4430,
+            1.5159,
+            0.7202,
+            -1.9153,
+            -1.7368,
+            -2.8723,
+            -0.5429,
+            -0.3954,
+            0.2952,
+            0.2103
+          ]
+        ])
+
+      actual = Axon.Layers.group_norm(a, b, c, num_groups: 2)
+
+      assert_all_close(expected, actual, atol: 1.0e-3)
+    end
+  end
 end