Fixing Overconfidence in Dynamic Neural Networks

This repository is the official implementation of the methods in the publication:

• Lassi Meronen, Martin Trapp, Andrea Pilzer, Le Yang, and Arno Solin (2024). Fixing overconfidence in dynamic neural networks. In IEEE Winter Conference on Applications of Computer Vision (WACV). [arXiv preprint]

Installing required Python packages in a virtual environment

• Start by installing Python version 3.7.4
• Create and start a virtual environment named MSDNet:

python -m venv MSDNet
source MSDNet/bin/activate

• Install the required packages into the newly created virtual environment:

python -m pip install -r requirements.txt

Obtaining the CIFAR-100, ImageNet, and Caltech-256 data sets

CIFAR-100

• The CIFAR-100 and Caltech-256 data sets will be automatically downloaded by the training script if you don't already have them (also CIFAR-10 if you wish to experiment on that). Note that on Caltech-256 the dowloaded image folders may contain some additional non-image files that need to be manually removed from the folders for the training scripts to run.

ImageNet

• The ImageNet data set can be downloaded at image-net.org. You should download the Training images (Task 1 & 2)(138GB) and Validation images (all tasks)(6.3GB) from the ILSVRC2012 version.
• After this you should have ILSVRC2012_img_train.tar and ILSVRC2012_img_val.tar. Extract ILSVRC2012_img_train.tar once to obtain a folder ILSVRC2012_img_train containing 1000 .tar subfolders.
• Make sure ILSVRC2012_img_train and ILSVRC2012_img_val.tar are in the same folder, we refer to this as /path_to_imagenet. Go to this folder to perform the following data set extraction.
• The complete training set can be extracted from the .tar files for example by doing the following:

mkdir train && cd train
find /path_to_imagenet/ILSVRC2012_img_train -name "*.tar" | while read NAME ; do SUBSTRING=$(echo $NAME| cut -d'/' -f 7) ; mkdir -p "${SUBSTRING%.tar}"; tar -xf "${NAME}" -C "${SUBSTRING%.tar}"; done
cd ..

• The validation set can be extracted as follows:

mkdir val && cd val && tar -xf /path_to_imagenet/ILSVRC2012_img_val.tar -C /path_to_imagenet/val
wget -qO- https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh | bash
cd ..

• After these steps you should have folders /path_to_imagenet/train and /path_to_imagenet/val which both contain 1000 subfolders, each containing the sample images for one class.

Training MSDNet backbone for CIFAR100

• To train the small model for CIFAR-100, run the following command:

python main.py --data-root /path_to_CIFAR100/ --data cifar100 \
	--save /savepath/MSDNet/cifar100_4  \
    --arch msdnet --batch-size 64 --epochs 300 --nBlocks 4 \
    --stepmode lin_grow --step 1 --base 1 --nChannels 16 --use-valid \
    -j 1 --var0 2.0 --laplace_temperature 1.0

• To run the medium and large models, you need to change the --nBlocks argument to 6 and 8 respectively. Note that you also need to change the path in --save to not overwrite previously trained models.

Training MSDNet backbone for ImageNet

• To train the small model for ImageNet, run the following command:

python main.py --data-root /path_to_imagenet --data ImageNet \
	--save /savepath/MSDNet/imagenet_base4 \
    --arch msdnet --batch-size 256 --epochs 90 --nBlocks 5 \
    --stepmode even --step 4 --base 4 --nChannels 32 --use-valid \
    --growthRate 16 --grFactor 1-2-4-4 --bnFactor 1-2-4-4 \
    -j 4 --gpu 0 --var0 2.0 --laplace_temperature 1.0

• To run the medium and large models, you need to change both the --step and --base arguments to 6 for the medium model, and to 7 for the large model. Note that you also need to change the path in --save to not overwrite previously trained models.

Training MSDNet backbone for Caltech-256

• To train the small model for Caltech-256, run the following command:

python main.py --data-root /path_to_caltech --data caltech256 \
	--save /savepath/MSDNet/caltech_base4 \
    --arch msdnet --batch-size 128 --epochs 180 --nBlocks 5 \
    --stepmode even --step 4 --base 4 --nChannels 32 --use-valid \
    --growthRate 16 --grFactor 1-2-4-4 --bnFactor 1-2-4-4 \
    -j 4 --gpu 0 --var0 2.0 --laplace_temperature 1.0

• To run the medium and large models, you need to change both the --step and --base arguments to 6 for the medium model, and to 7 for the large model. Note that you also need to change the path in --save to not overwrite previously trained models.

Compute Laplace approximation separately after training

• The Laplace approximation is precomputed automatically at the end of training. However, if you wish to separately recompute the Laplace approximation, you can do so as follows:

python main.py --data-root /path_to_CIFAR100/ --data cifar100 \
	--save /savepath/MSDNet/cifar100_4 \
    --arch msdnet --batch-size 64 --epochs 300 --nBlocks 4 \
    --stepmode lin_grow --step 1 --base 1 --nChannels 16 --use-valid \
    -j 1 --compute_only_laplace --resume /savepath/MSDNet/cifar100_4/save_models/model_best_acc.pth.tar \
    --var0 2.0

• Note that you need to change the arguments --save and --resume to the correct path that contains the trained model that you want to calculate the Laplace approximation for.
• The example command is for CIFAR-100 but the same can be done for ImageNet or Caltech-256, by adding the --compute_only_laplace and --resume /path_to_saved_model/save_models/model_best_acc.pth.tar arguments to the ImageNet or Caltech-256 training command.

Test vanilla MSDNet on CIFAR-100

• To test the small vanilla MSDNet model on CIFAR-100, run the following:

python main.py --data-root /path_to_CIFAR100/ --data cifar100 --save /savepath/MSDNet/cifar100_4 \
    --arch msdnet --batch-size 64 --epochs 300 --nBlocks 4 --stepmode lin_grow --step 1 --base 1 \
    --nChannels 16 --use-valid -j 1 --evalmode dynamic \
    --evaluate-from /savepath/MSDNet/cifar100_4/save_models/model_best_acc.pth.tar

• Note that the --save and --evaluate-from arguments have to be the correct paths to the saved model directory.
• For medium and large models you need to again change the --nBlocks argument accordingly, as well as the paths in --save and --evaluate-from.

Test vanilla MSDNet on ImageNet

• To test the small vanilla MSDNet model on ImageNet, run the following:

python main.py --data-root /path_to_imagenet --data ImageNet --save /savepath/MSDNet/imagenet_base4 \
    --arch msdnet --batch-size 256 --epochs 90 --nBlocks 5 --stepmode even --step 4 --base 4 \
    --nChannels 32 --use-valid -j 4 --gpu 0 --evalmode dynamic \
    --growthRate 16 --grFactor 1-2-4-4 --bnFactor 1-2-4-4 \
    --evaluate-from /savepath/MSDNet/imagenet_base4/save_models/model_best_acc.pth.tar

• Here again the medium and large models require changing the --step and --base arguments as described in the model training, and the paths in --save and --evaluate-from need to be set correctly to utilize the correct saved model that you want to evaluate.

Test vanilla MSDNet on Caltech-256

• To test the small vanilla MSDNet model on Caltech-256, run the following:

python main.py --data-root /path_to_caltech --data caltech256 --save /savepath/MSDNet/caltech_base4 \
    --arch msdnet --batch-size 128 --epochs 180 --nBlocks 5 --stepmode even --step 4 --base 4 \
    --nChannels 32 --use-valid -j 4 --gpu 0 --evalmode dynamic \
    --growthRate 16 --grFactor 1-2-4-4 --bnFactor 1-2-4-4 \
    --evaluate-from /savepath/MSDNet/caltech_base4/save_models/model_best_acc.pth.tar

• Here again the medium and large models require changing the --step and --base arguments as described in the model training, and the paths in --save and --evaluate-from need to be set correctly to utilize the correct saved model that you want to evaluate.

Test models that use Laplace and/or model-internal ensembling (MIE)

• To use Laplace approximation in the evaluation of a model, add the following arguments to the model testing commands:

--laplace --laplace_temperature 1.0 --var0 2.0 --n_mc_samples 50 --optimize_temperature --optimize_var0

• To use MIE in the evaluation of a model, add the argument --MIE to the model testing commands.
• To test with both Laplace and MIE (our model) add both of the above mentioned arguments into the testing command.

License

This software is provided under the MIT License.