**Warning : You need to have access to a GPU to launch the experiments in a reasonable time**
- Of course, start by cloning this repo
- Make sure you have at least **Python 3.8**
- Then we strongly advised to the readers to create a new environment for this project for instance with the command `conda create -n the_name_you_want`
- You will need after to install pip `conda install pip` and launch `pip install -r requirements.txt`
- Then we strongly advise to the readers to create a new environment for this project for instance with the command `conda create -n the_name_you_want`
- You will need after to install pip with `conda install pip`
- If you have a gpu do `conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch` otherwise if you prefer to use your CPU do `conda install pytorch torchvision torchaudio cpuonly -c pytorch`
- At last launch `pip install -r requirements.txt`
### Construction of the domains and running of our experiments
After preparing the required environment, you need to construct the bases which will be used for the training and evaluation phases.
...
...
@@ -21,10 +21,10 @@ After preparing the required environment, you need to construct the bases which
*To be sure that you construct the same domains as ours, we saved a .txt file containing the order of the image paths we extracted from the Splicing category. By default, the code available here take into account that list and not the one you could obtain from your side that can lead to a different order.*
- When this is finished, you can reproduce our experiments lauching the script `simulations.py` in the folder `Experiments`.
Each experiment of our paper is associated to a code. Giving the code to the function `reproduce(code)` from `simulations.py` enables to reproduce our experiments. However, this does not ensure that you will obtain the same results since it can change according to your GPU/CPU.
**All the experiments have been launched with the GPU NVIDIA GeForce GTX 1060 6GB**. You can be at least sure to obtain the same datasets and to start the training with the same weights.
- When this is finished, you can reproduce our experiments lauching the script `simulations.py` moving to the folder `Experiments`. By default it will reproduce all the experiments of the paper given that you created previously all the necessary domains.
Each experiment of our paper is associated to a code. Giving the code to the function `reproduce` from `simulations.py` enables to reproduce our experiments. However, this does not ensure that you will obtain the same results since it can change according to your GPU/CPU.
You can be at least sure to start the trainings with the same weights.
Please find below a table linking a code to an experiment presented in the paper :
