OSUPCVLab/EmotionFromFaceImages: Rahul' project · GitHub

MotivNet is a model built on Meta-Sapiens and ML-Decoder for the FER task on AffectNet. See Paper

We provide code to run inference on the model and code to train the model on your own data with custom specifications.

For inference, the output labels are shown below

0: Neutral, 1: Happiness, 2: Sadness, 3: Surprise, 4: Fear, 5: Disgust, 6: Anger.

Getting started is very simple.

First, create a new virtual environment with and run pip install -r requirements.txt

Then, download the MotivNet checkpoint from OneDrive. https://buckeyemailosu-my.sharepoint.com/:u:/g/personal/medicharla_2_buckeyemail_osu_edu/EfnsSxS42JNDipAEU45o-bUBsGfXviOOgaWka5LBLBkvBA?e=1SF6v8

Place this checkpoint in the /checkpoints/ folder to start finetuning or running inference on the model (defined in model.py)

To run predictions on a set of images, put all of the images in one folder and use the inference.py file to run your predictions

To train MotivNet for your own custom use case, you need 4 values to pass into the train.py file

1. --train_data This value is the folder path for all the images that comprise your training dataset
2. --train_labelsThis value holdes the file path for the ground truths of your training dataset. It is a JSON file where each key is the emotion_label (0,1,...,7), as defined above, and it holds an array of the image file names belonging to that emotion class.
3. --test_data This value is the folder path for all the images that comprise your test dataset
4. --test_labelsThis value holdes the file path for the ground truths of your tests dataset. It's stored in the same format as the train_labels.