This repository contains the PyTorch implementation of the paper Reducing inference energy consumption using dual complementary CNNs, published in the Future Generations Computer System journal. It contains three Python scripts that run the experiments as described in the paper, and are explained bellow.
- Python 3
- Pytorch >= 2.1.0
- Torchvision >= 0.16.0
To start with the experiments you begin by cloning the repository:
git clone https://github.com/michaelkinnas/Reducing-Inference-Energy-Consumption-Using-Two-Complementary-CNNs
cd Reducing-Inference-Energy-Consumption-Using-Two-Complementary-CNNsTo install the depedencies run:
pip install -r requirements.txtIt is recommended to use a python virtual environment:
- with python venv:
python -m venv <path-to-environment>/<environment-name>
source <path-to-environment>/<environment-name>/bin/activate
pip install -r requirements.txt- with conda:
conda create --name <environment-name> --file requirements.txtThe pipeline as described in the paper has three main steps:
- Calculate complementarity matrix from given pretrained CNN models and select two CNN models.
- Calculate λ hyperparamter value for selected CNN pair.
- Run the inference methodology on test dataset with the selected CNN pair and calculated λ parameter.
The script file complementarity.py will calculate a complementarity matrix for a given pretrained CNN model pool.
To run it use the command python3 complementarity.py plus some additional parameters as described bellow:
-h, --help show this help message and exit
-D, --dataset {cifar10,imagenet,intel,fashionmnist}
Define which dataset models to use.
-f, --dataset-root DATASET_ROOT
The root file path of the validation or test dataset. (e.g. For CIFAR-10 the directory
containing the 'cifar-10-batches-py' folder, etc.)
-t, --train Define whether to use the training or test split, for datasets that require that
parameter.
-w, --weights WEIGHTS
Optional. The path directory of custom weights for all the models used in the process.
The files should be in '.pth' extension and named after the original CIFAR-10 model name
(e.g. 'resnet20.pth'). If not set the default pretrained CIFAR-10 model weights will be
used.
The results are saved in a complementarity.csv file in the same directory that you run the script.
python3 complementarity.py -D cifar10 -f "<path to dataset root>"NOTICE: If using the Intel or FashionMNIST datasets you must provide the directory that contain all the .pth model weights for all corresponding CIFAR-10 models as shown bellow. (You can download the .tar.gz files containing the pretrained weights for Intel here and for FashionMNIST here, and then unzip them in that directory), otherwise the default CIFAR-10 weights from PyTorch will be used.
The script file threshold.py will calculate the optimal threshold hyperparameter for a given CNN pair.
To run it use the command python3 threshold.py plus some additional parameters as described bellow:
-h, --help show this help message and exit
-D, --dataset {cifar10,imagenet,intel,fashionmnist}
Define which dataset models to use.
-f, --dataset-root DATASET_ROOT
The root file path of the validation or test dataset. (e.g. For CIFAR-10 the directory
containing the 'cifar-10-batches-py' folder, etc.)
-m1, --model1 MODEL1 The first model, required.
-m2, --model2 MODEL2 The second model, required.
-t, --train Only valid for the CIFAR-10 dataset. Define wether to use the training or test dataset.
-n, --n_threshold_values N_THRESHOLD_VALUES
Define the number of threshold values to check between 0 and 1. Higher numbers will be
slower. Default is 2000
-w1, --weights1 WEIGHTS1
Optional. Directory of the '.pth' weights file for the first model.
-w2, --weights2 WEIGHTS2
Optional. Directory of the '.pth' weights file for the second model.
To calculate the best threshold hyperparameter for the selected CNN pair, you can type:
python3 threshold.py --model1 resnet20 --model2 mobilenetv2_x0_5 --filepath "<path to dataset root>" -tNOTICE: If using Intel of FashionMNIST datasets you must provide the weights .pth file for each model that is trained on this dataset. By default the pretrained model weights are for the CIFAR-10 dataset. You can download the .tar.gz files containing all the pretrained weights for Intel here and for FashionMNIST here. You can unzip the .tar.gz files and point to the corresponding CIFAR-10 model you are using.
-h, --help show this help message and exit
-y, --yml-file YML_FILE
Use .yml configuration file instead of cli arguments. In this case you must provide the
location of the .yml file and the rest of the arguments are ignored.
-D, --dataset {cifar10,imagenet,intel,fashionmnist}
The dataset to use.
-m1, --model-1 MODEL_1
The first model name, required. It must be included in the provided lists of available
models.
-m2, --model-2 MODEL_2
The second model. It must be included in the provided lists of available models.
-w1, --weights-1 WEIGHTS_1
Optional. A file path to the first model's weights file.
-w2, --weights-2 WEIGHTS_2
Optional. A file path to the second model's weights file.
-f, --dataset-root DATASET_ROOT
The root file path of the validation or test dataset. (e.g. For CIFAR-10 the directory
containing the 'cifar-10-batches-py' folder, etc.)
-s, --scorefn {maxp,difference,entropy,truth}
Score function to use.
-t, --threshold THRESHOLD
The threshold value to use for the threshold check. (λ parameter)
-p, --postcheck Enable post check. Default is false.
-m, --memory {dhash,invariants}
Enable memory component. Default is None.
-d, --duplicates DUPLICATES
Set the percentage of the original training set for duplication. Default is 0 (No
duplicates). Range [0-1]
-r, --rotations If set the duplicated samples will be randomly rotated or mirrored.
-rp, --root-password ROOT_PASSWORD
Optional. If provided the password will be used to command the computer to shutdown after finishing.
python3 main.py --model1 resnet20 --model2 mobilenetv2_x0_5 --filepath "<path to dataset root>" --scorefn difference --threshold 0.8724 --postcheckInstead of cli parameters you can use a configuration file. If you use the -y parameter you must provide the *.yml file to read from. An example yml configuration file is provided in this repo.
NOTICE: If using Intel of FashionMNIST datasets you must provide the weights .pth file for each model that is trained on this dataset. By default the pretrained model weights are for the CIFAR-10 dataset. You can download the .tar.gz files containing all the pretrained weights for Intel here and for FashionMNIST here. You can unzip the .tar.gz files and point to the corresponding CIFAR-10 model you are using.
- mobilenetv2_x0_5
- mobilenetv2_x0_75
- mobilenetv2_x1_0
- mobilenetv2_x1_4
- repvgg_a0
- repvgg_a1
- repvgg_a2
- resnet20
- resnet32
- resnet44
- resnet56
- shufflenetv2_x0_5
- shufflenetv2_x1_0
- shufflenetv2_x1_5
- shufflenetv2_x2_0
- vgg11_bn
- vgg13_bn
- vgg16_bn
- vgg19_bn
- convnext_tiny
- densenet121
- densenet161
- densenet169
- densenet201
- googlenet
- inception_v3
- mnasnet1_0
- mnasnet1_3
- mobilenet_v3_large
- regnet_x_16gf
- regnet_x_1_6gf
- regnet_x_3_2gf
- regnet_x_400mf
- regnet_x_800mf
- regnet_x_8gf
- regnet_y_1_6gf
- regnet_y_3_2gf
- regnet_y_400mf
- regnet_y_800mf
- regnet_y_8gf
- resnext50_32x4d
- resnet101
- resnet34
- resnet50
- shufflenet_v2_x2_0
- swin_s
- swin_t
- swin_v2_s
- swin_v2_t
This work was funded by the European Union’s Horizon Europe research and innovation program under grant agreement No. 101120237 (ELIAS)
This project is licensed under the Apache License 2.0. You can red the LICENSE file for more details.