Project GDL - AEGNN

@authors: Filippo Casari, Alessandro De Grandi, Anthony Bugatto

Setup

We suggest to run our code with a linux OS.
Create a conda environment by running:

conda env create --file=environment.yml

Preprocessing

To run the preprocessing script:

AEGNN_DATA_DIR=$(pwd)/data python3 scripts/preprocessing.py --dataset "ncaltech101" --num-workers 4 --gpu 0

In case you want to work on Ncars dataset, substitute Ncaltech with 'ncars' AEGNN_DATA_DIR is the dir in which the dataset should be. You can change this env variable if your dataset is elsewhere. Change the number of workers according with your hardware. Alternatively, use experiments jupyter notebook that includes also dataset splitting, and some useful plots for accuracy per event.

Training

AEGNN_DATA_DIR=$(pwd)/data AEGNN_LOG_DIR=$(pwd)/data/log python scripts/train.py graph_res --task recognition --dataset "ncaltech101" --gpu 0 --batch-size 8 --max-epochs 20 --num-workers 8

As before, you can change the name of the dataset you are working on as well as the batch size and number of epochs. The authors used for Ncaltech a batch size of 16 and 64 for Ncars. If your GPU memory runs out, you may want to reduce the batch size or use:

trainer_kwargs["accumulate_grad_batches"] = 2

that is set true by default within the code.

Flops Evaluation

PYTHONPATH=$(pwd) AEGNN_DATA_DIR=$(pwd)/data/storage python evaluation/flops.py

Accuracy per event

To run the accuracy per event, you have to pass as args the path of your trained model, the device you're using, the dataset, and the batch size. We added the last argument to the script to avoid running out of memory.

python evaluation/accuracy_per_events.py /home/ale/Downloads/GDL/data/log/checkpoints/ncaltech101/recognition/20230505200347/epoch\=19-step\=8199.pt --device cuda --dataset ncaltech101 --bsize 8

Suggestion: for Ncars use as batch size 64.

Plot flops

We created a script for plotting number of flops per layer. You can run:

python example_plot_flops.py