From 4e01b4bc89356c86f5e75323cfdb81a17eed7e13 Mon Sep 17 00:00:00 2001 From: Paul Balanca Date: Thu, 11 Jan 2024 15:47:00 +0000 Subject: [PATCH] Use `float32` scale in AutoScale MNIST training example. (#81) In some configurations, `float16` does not have enough dynamic range to represent scaling, meaning the scale factor can quickly overflow. Using `float32` (or equivalently `bfloat16`) should allow a bit more margin.