bupt-priv pytorch-classification
Classification on CIFAR-10/100 and ImageNet with PyTorch.
Disclaimer
The official pytorch-classification code is available here, thanks for bearpaw's job!
The bupt-priv pytorch-classification is modified version of pytorch-classification, and not for commercial use.
Features
- Unified interface for different network architectures
- Multi-GPU support
- Training progress bar with rich info
- Training log and training curve visualization code (see
./utils/logger.py
) - Add Aligned Inception Resnet/ResNeXt (by soeaver)
- Add ResNeXt26-32x4d (by soeaver)
Install
- Install PyTorch
- Clone recursively
git clone --recursive https://github.com/soeaver/pytorch-classification
Training
Please see the Training recipes for how to train the models.
Results
CIFAR
Top1 error rate on the CIFAR-10/100 benchmarks can be found in official pytorch-classification
ImageNet
Single-crop (224x224) validation error rate is reported.
Model | Params (M) | Top-1 Error (%) | Top-5 Error (%) |
---|---|---|---|
ResNet18 | 11.69 | 30.09 | 10.78 |
ResNeXt50-32x4d | 25.03 | 22.60 | 6.29 |
ResNet18 (bupt-priv) | 11.69 | 29.11 | 10.07 |
ResNeXt26-32x4d (bupt-priv) | 15.39 | 24.93 | 7.75 |
Se-ResNeXt50-32x4d (bupt-priv) | 27.56 | 21.91 | 6.02 |
AIR101 (bupt-priv) | 64.38 | 20.74 | 5.56 |
AIRX50-24x4d (bupt-priv) | 28.09 | 21.88 | 5.95 |
AIRX101-32x4d (bupt-priv) | 64.17 | 20.62 | 5.50 |
AIRXt152-32x4d (bupt-priv) | 89.34 | 20.29 | 5.24 |
- All the bupt-priv models are trained for 130/140 epochs, the lr schedule is 61 91 121.
- AIR/AIRX is short for Aligned-Inception-ResNe(X)t.
Supported Architectures
CIFAR-10 / CIFAR-100
Since the size of images in CIFAR dataset is 32x32
, popular network structures for ImageNet need some modifications to adapt this input size. The modified models is in the package models.cifar
:
- AlexNet
- VGG (Imported from pytorch-cifar)
- ResNet
- Pre-act-ResNet
- ResNeXt (Imported from ResNeXt.pytorch)
- Wide Residual Networks (Imported from WideResNet-pytorch)
- DenseNet
ImageNet
- All models in
torchvision.models
(alexnet, vgg, resnet, densenet, inception_v3, squeezenet) - ResNeXt
- AIR/AIRX
Contribute
Feel free to create a pull request if you find any bugs or you want to contribute (e.g., more datasets and more network structures).