wangshankun / cifar10-100-fast-training

https://github.com/WeihaoZhuang/cifar10-100-fast-training

使用NVIDIA docker shm-size稍微大些

下载对应cuda版本号的NVIDIA-pytorch镜像，我这里cuda是11.4,对应版本是21.07

docker pull nvcr.io/nvidia/pytorch:21.07-py3

docker run --gpus all -it --shm-size="8g" -v /home/wsk/data:/root/data nvcr.io/nvidia/pytorch:21.07-py3 bash

进入后执行main.py 开始训练

CIFAR10-100-fast-training

This implementation is for cifar10/100 fast training base on (https://github.com/davidcpage/cifar10-fast)

Please check demo.ipynb.

You can run 70 epochs rather than 350 (https://github.com/kuangliu/pytorch-cifar) or 200 (https://github.com/weiaicunzai/pytorch-cifar100) epochs.

Requirements:

    pytorch
    apex (https://nvidia.github.io/apex/)
    numpy

Results:

Model	CIFAR-10 accuracy	CIFAR-100 accuracy
ResNet18	95.32	76.75
ResNet34	95.57	77.48
ResNet50	95.62	77.66
MobileNet	92.25	60.35
MobileNetV2	93.69	62.56
DenseNet-cifar	94.7	72.61
DenseNet121	95.1	76.97
DenseNet201	94.89	77.28
Wide-ResNet40	95.13	74.79
Wide-ResNet16	95.4	78.52
Wide-ResNet28	96.21	79.86
VGG11	92.41	70.07
VGG16	94.27	72.68
VGG19	94.22	71.11
GoogleNet	95.35	79.24
InceptionV3	95.55	79.29