We used a new learning rate scheduler called HTD.
You can see the toy demo code here or our papar in arXiv.
This repository is about some experiments of learning rate with CIFAR-10 & CIFAR-100.
The original paper start with a learning rate of 0.1, divide it by 10 at 32k(81 epoch) and 48k(122 epoch) iterations, and terminate training at 64k iterations(200 epochs total).
I ran other experiments based on the same architecture. The only difference is learning rate schedule.
All of the tensorboard logs & pretrained models are available at BIGBALLON/pretrained_models
You can run the script run.sh
to start all the experiments.
Or just run the command like:
python3 ResNet_keras.py --epochs 200 --stack_n 3 --lr_scheduler 1 --dataset cifar100
Please feel free to contact me if you have any questions! ๐ธ