Use this link to download weights (you have to download and extract manually).
To set the path to the downloaded weights folder specify path_to_state_dict in vgg.py.
Install dependecies by running:
pip install -r requirements.txt-net - choice of architecture (default: resnet18)
-dataset - choice of dataset (default: cifar10)
-tr-batch - training batch size (default: 512)
-val-batch - validation batch size (default: 512)
-lr - learning rate (default: 0.1)
-wd - weight decay (default: 5e-4)
-epochs - number of epochs to train (default: 300)
-cpu - cpu flag
-reinit - diversity loss usage flag
-mode - choice between transfer and default training
python train.py -lr 0.1 -gpu -dataset cifar100 -reinit -alpha-drop -net resnet18 -b 512
python train.py -lr 0.1 -gpu -dataset cifar100 -net resnet18 -b 512
| No. | Model | default | cosine | dataset | alpha | epochs |
|---|---|---|---|---|---|---|
| 1 | resnet18 | 75.21% | 75.82% | cifar100 | from 1 to 0 | 200 |
| 2 | resnet34 | 75.54% | 76.91% | cifar100 | from 1 to 0 | 200 |
| 3 | resnet50 | 75.52% | 78.14% | cifar100 | from 1 to 0 | 200 |
| 4 | seresnet18 | 74.56% | 75.58% | cifar100 | from 1 to 0 | 200 |
| 5 | mobilenet | 67.03% | 67.25% | cifar100 | from 1 to 0 | 200 |
| 6 | mobilenetv2 | 69.53% | 69.72% | cifar100 | from 1 to 0 | 200 |
| 7 | squeezenet | 68.91% | 69.44% | cifar100 | from 1 to 0 | 200 |
We use pre-commit hooks with black and flake8 to align code format.
pip install pre-commit black flake8Initialize it from the folder with the repo:
pre-commit install