search

wagner-niklas / CAGE_expression_inference

Project to infere emotional expressions and benchmark datasets by Niklas Wagner, Felix Mätzler, Samed R. Vossberg, Helen Schneider and Svetlana Pavlitska.
MIT License
11 stars 2 forks source link

Trained weights #1

Closed JSOlier closed 4 months ago

JSOlier commented 4 months ago

Hi, I am trying to use these models for valence estimation. However, I am unable to find the trained weights. Many scripts are loading the file best_model_affectnet_improved7VA, which is not part of the repository. Where can I find that file? Thank.

wagner-niklas commented 4 months ago

Hi @JSOlier thanks for the interest in our project. We will update our files this afternoon. Best regards, Niklas

FelixMaetzler commented 4 months ago

Hi @JSOlier, i committed all the scripts we used for our final paper. You can find them in the models directory. In every subdirectory there is a train.py and a generate_csv.py file. The first one generates the model.pt (it's too large to commit to github) and the second uses this model.pt and creates the inference.csv. The evaluation.py then uses (or creates) the inference.csv and prints the results.

As described in the paper, the VA models where trained on top of the combined models.

I hope this helps you. Please feel free to contact us again if you have further questions.

Best regards, Felix

ZayneHuang commented 4 months ago

Hi @FelixMaetzler,

Thanks for your contribution. Could you share the checkpoint files on other platforms like Google Drive?

Or, if I want to train the model on AffectNet, I wonder which model achieved the best accuracy? According to the paper, MaxViT valence−arousal won the best performance in relation to valence and arousal prediction on AffectNet 8, which is this model in the respository. Am I right?

wagner-niklas commented 4 months ago

Hi @ZayneHuang,

your right, AffectNet8_Maxvit_VA is our overall best model regarding valence and arousal estimation. Please keep in mind that we initialized the model weights of AffectNet8_Maxvit_Combined in this Script. Therefore, you have to train this model too.

Best regards, Niklas

ZayneHuang commented 4 months ago

Hi @ZayneHuang,

your right, AffectNet8_Maxvit_VA is our overall best model regarding valence and arousal estimation. Please keep in mind that we initialized the model weights of AffectNet8_Maxvit_Combined in this Script. Therefore, you have to train this model too.

Best regards, Niklas

Thanks for your reply and advice : )

Varun-GP commented 4 months ago

Thanks for your contribution. Could you please share the checkpoint files maybe on a Google Drive Link? It would greatly help in understanding the model prediction from live videos.