Skip to content

Latest commit

 

History

History

ImageNet-ResNet-PyTorch

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 

To Train:

This command trains a ResNet50 with GN on ImageNet:

python main.py /path/to/imagenet --multiprocessing-distributed -j 8

Trained on 8 GPUs with a total batch size of 256, it should achieve 24.1~24.3 Top-1 accuracy. This is about 0.1 worse than the paper, probably because this implementation uses fewer augmentations than the paper.

Training takes ~26 hours on 8 V100s. A training log is included.

The code is modified slightly from pytorch official examples and uses the same data augmentations available there.

To evaluate with the pretrained model:

python main.py /path/to/imagenet/ -e --resume model.pth --batch 64