Baselines (currently just ResNets and MobileNets) on Stanford Cars Using a minimally modified version of the canonical PyTorch ImageNet example.
Model | Accuracy |
---|---|
ResNet-18 | 86.0 |
ResNet-50 | 90.0 |
ResNet-101 | 90.1 |
ResNet-152 | 90.1 |
MobileNetv2 | 87.1 |
ResNet-18* | 61.5 |
ResNet-50* | 7.9 |
ResNet-101* | 5.8 |
ResNet-152* | 4.4 |
MobileNetv2* | 66.4 |
* denotes no pretraining. Each model just uses the default learning rate schedule (decay by 10 every 30 epochs), and 90 epochs of training. Models without pretraining seem to have stability issues as sometimes they converge to less than 10% accuracy.
Dependencies:
- PyTorch
- SciPy (for parsing original
.mat
metadata files)
To reproduce a result:
run download_and_prepare.sh
to download the original images and organize them into an ImageFolder
dataset.
Run main.py
(exact same options as the official PyTorch ImageNet example).
e.g., python3 main.py --pretrained --arch resnet50
Please submit a PR if you have an improvement to a baseline that is established in the literature (e.g., new data augmentation strategy, regularization, additional image resolution, etc.) and requires minimal hyperparameter tuning.