Environmental impact of computer vision

Analysis of data

Notebooks

1. Training and FLOP analysis - get the data for model accuracy, power draw from power monitor readings and FLOPs. Linear regression is performed to get the relation between FLOPs and energy use.
   Code for training in https://github.com/MScDisseration/model_hub/tree/master/Training
   Code for FLOPs in https://github.com/MScDisseration/model_hub/blob/master/models.py

2. Vision models inference analysis - similar to previous notebook. Data is collected for power draw during inference (10000 inference run for each model).
   Code for running inference in https://github.com/MScDisseration/model_hub/tree/master/Inference

Training and inference on computer vision models

pip install -r requirements.txt

https://pytorch.org/docs/master/torchvision/models.html#classification

Get Floating Point Operations, FLOPs

python models.py

Using https://github.com/Lyken17/pytorch-OpCounter

Get data from training

Download folder of images to train on
curl https://download.pytorch.org/tutorial/hymenoptera_data.zip --output /media/data/hymenoptera_data.zip

Train models
cd Training
python finetune_multiple.py

Get power data for Inference

cd Inference
python multipleRuns.py

Extras

Run inference on one model

cd Inference
python vision.py --path imagepath --model modelname

E.g. python vision.py --path "../data/butterfly.jpg" --model "alexnet"