lifeiteng/Optimizers

Tensorflow Optimizers

RAdam & Lookahead Tensorflow Optimizers

Papers

• RAdam: On The Variance Of The Adaptive Learning Rate And Beyond
  • official pytorch code
• Lookahead: Lookahead Optimizer: k steps forward, 1 step back

Test on Tensorflow==1.13.1

pip install tensorflow-gpu==1.13.1
# If you are in China
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple tensorflow-gpu==1.13.1

Tests

python lookahead_optimizer_test.py
python radam_optimizer_test.py

TODOs

• Lookahead: Find a simle way to init local variables at the first step.

cifar10 example

• copied and modified from tensorflow/models

pip install tensorflow_datasets
# If you are in China
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple tensorflow_datasets

dir=exp/cifar10
mkdir -p $dir

for lr in "0.01" "0.001" "0.1";do
	python cifar10_train.py --learning_rate $lr --optimizer   sgd --train_dir $dir/train_SGD_lr${lr} || exit 1
	python cifar10_train.py --learning_rate $lr --optimizer  adam --train_dir $dir/train_Adam_lr${lr}
	python cifar10_train.py --learning_rate $lr --optimizer radam --train_dir $dir/train_RAdam_lr${lr}
	python cifar10_train.py --learning_rate $lr --optimizer lookaheadradam --train_dir $dir/train_RAdam_lr${lr}_Lookahead
done

• 20000 step results
  • SGD likes large learning rate. LR = 0.1 0.01 0.001
  • Adam is excellent when LR=0.001. LR = 0.1 0.01 0.001
  • RAdam is not sensitive to learning rate. LR = 0.1 0.01 0.001, but failed when LR=0.1
  • Best results(SGD LR=0.1, Adam LR=0.001, RAdam LR=0.001)
  • Lookahead is helpful sometimes.