This repository is intended to provide a quick implementation of the LHC-Net and to replicate the results in this paper on FER2013 by downloading our trained models or, in case of hardware compatibility, by training the models from scratch. A fully custom training routine is also available.
Bit-exact replicability is strongly hardware dependent. Since the results we presented depend on the choice of a very good performing starting ResNet34v2 model, we strongly recommend to run the replicability script before attempting to execute our training protocol which is computational intensive and time consuming.
Execute the following commands in your terminal:
python Download_Data.py
python ETL.py
python check_rep.py
Ore equivalently:
python main_check_rep.py
If you get the output "Replicable Results!" you will 99% get our exact result, otherwise if you get "Not Replicable Results. Change your GPU!" you won't be able to get our results.
Please note that Download_Data.py will download the FER2013 dataset in .csv format while ETL.py will save all the 28709 images of the training set in .jpeg format in order to allow the use of TensorFlow image data generator and save some memory.
Recommended setup for full replicability:
Nvidia Geforce GTX-1080ti (other Pascal-based GPUs might work)
GPU Driver 457.51
Cuda Driver 11.1.1*
CuDNN v8.0.5 - 11.1
Python 3.8.5
requirements.txt
*After Cuda installation rename C:...\NVIDIA GPU Computing Toolkit\CUDA\v11.1\bin\cusolver64_11.dll in cusolver64_10.dll
Execute the following commands in your terminal:
python Download_Data.py
python Download_Models.py
python LHC_Downloaded_Eval.py
python Controller_Downloaded_Eval.py
Ore equivalently:
python main_downloaded.py
To train an LHC-Net using a generically imagenet pre-trained ResNet backbone edit the configuration files in the Settings folder and execute the following commands in your terminal:
python Download_Data.py
python ETL.py
python LHC_Net_Train.py
python LHC_Net_Eval.py
Ore equivalently:
python main_standalone.py
If the replicability check gave a positive result you could replicate our results by integrating and training the LHC modules on a ResNet backbone already trained on FER2013, according with our first experimental protocol. To do that execute the following commands in your terminal:
python Download_Data.py
python ETL.py
python ResNet34_Train.py
python LHC_Train.py
python Controller_Train.py
python LHC_Eval.py
python Controller_Eval.py
Ore equivalently:
python main_modular.py
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent