From 798083264a5dd63e68f7c6a0e5704ed78b7e5436 Mon Sep 17 00:00:00 2001 From: Lauren Yu <6631887+laurenyu@users.noreply.github.com> Date: Wed, 27 Jun 2018 15:22:13 -0700 Subject: [PATCH] Lower learning rate for TF cifar integ test We've been seeing a lot of errors that say failure due to NaN loss. One suggestion from https://stackoverflow.com/questions/40050397/deep-learning-nan-loss-reasons is to use a lower number for learning rate. --- tests/data/cifar_10/source/resnet_cifar_10.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tests/data/cifar_10/source/resnet_cifar_10.py b/tests/data/cifar_10/source/resnet_cifar_10.py index 47d416ebcf..39101f1fb6 100644 --- a/tests/data/cifar_10/source/resnet_cifar_10.py +++ b/tests/data/cifar_10/source/resnet_cifar_10.py @@ -33,8 +33,8 @@ BATCH_SIZE = 1 # Scale the learning rate linearly with the batch size. When the batch size is -# 128, the learning rate should be 0.1. -_INITIAL_LEARNING_RATE = 0.1 * BATCH_SIZE / 128 +# 128, the learning rate should be 0.05. +_INITIAL_LEARNING_RATE = 0.05 * BATCH_SIZE / 128 _MOMENTUM = 0.9 # We use a weight decay of 0.0002, which performs better than the 0.0001 that