fix(layers): Update Conv2D docstring to clarify numerical precision a…

…cross backends (#20867) * fix(layers): Update Conv2D docstring to clarify numerical precision across backends Clarify that Conv2D operations may exceed the documented 1e-7 precision difference across backends Document that large convolutions can show notable variations due to accumulated floating-point operations * Update conv2d.py --------- Co-authored-by: François Chollet <[email protected]>
keras-team · Feb 6, 2025 · 3906e32 · 3906e32
1 parent 43f40d8
commit 3906e32
Showing 1 changed file with 9 additions and 0 deletions.
diff --git a/keras/src/layers/convolutional/conv2d.py b/keras/src/layers/convolutional/conv2d.py
@@ -12,6 +12,15 @@ class Conv2D(BaseConv):
     and added to the outputs. Finally, if `activation` is not `None`, it is
     applied to the outputs as well.
 
+    Note on numerical precision: While in general Keras operation execution
+    results are identical across backends up to 1e-7 precision in float32,
+    `Conv2D` operations may show larger variations. Due to the large
+    number of element-wise multiplications and additions in convolution
+    operations, especially with large inputs or kernel sizes, accumulated
+    floating-point differences can exceed this 1e-7 threshold. These variations
+    are particularly noticeable when using different backends (e.g., TensorFlow
+    vs JAX) or different hardware.
+
     Args:
         filters: int, the dimension of the output space (the number of filters
             in the convolution).