ResNeXt.pytorch

Reproduces ResNet-V3 (Aggregated Residual Transformations for Deep Neural Networks) with pytorch.

Tried on pytorch 1.6
Trains on Cifar10 and Cifar100
Upload Cifar Training Curves
Upload Cifar Trained Models
Pytorch 0.4.0
Train Imagenet

Download

git clone https://github.com/prlz77/resnext.pytorch
cd resnext.pytorch
# git checkout R4.0 or R3.0 for backwards compatibility (not recommended).

Usage

To train on Cifar-10 using 2 gpu:

python train.py ~/DATASETS/cifar.python cifar10 -s ./snapshots --log ./logs --ngpu 2 --learning_rate 0.05 -b 128

It should reach ~3.65% on Cifar-10, and ~17.77% on Cifar-100.

After train phase, you can check saved model.

Thanks to @AppleHolic we have now a test script:

To test on Cifar-10 using 2 gpu:

python test.py ~/DATASETS/cifar.python cifar10 --ngpu 2 --load ./snapshots/model.pytorch --test_bs 128

Configurations

From the original paper:

cardinality	base_width	parameters	Error cifar10	error cifar100	default
8	64	34.4M	3.65	17.77	x
16	64	68.1M	3.58	17.31

Update: widen_factor has been disentangled from base_width because it was confusing. Now widen factor is set to consant 4, and base_width is the same as in the original paper.

Trained models and curves

Link to trained models corresponding to the following curves:

Update: several commits have been pushed after training the models in Mega, so it is recommended to revert to e10c37d8cf7a958048bc0f58cd86c3e8ac4e707d

Other frameworks

torch (@facebookresearch). (Original) Cifar and Imagenet
caffe (@terrychenism). Imagenet
. Imagenet

Cite

@article{xie2016aggregated,
  title={Aggregated residual transformations for deep neural networks},
  author={Xie, Saining and Girshick, Ross and Doll{\'a}r, Piotr and Tu, Zhuowen and He, Kaiming},
  journal={arXiv preprint arXiv:1611.05431},
  year={2016}
}