This project includes a model training tool to train models to anlayse emotions from face pictures. The dataset used for training is FER2013.
1.Install dependencies
python -m pip install -r requirements.txt2.Preprocess datasets
Preprocessing pipeline for AFEW and Emotic datasets. Refercences, download links and licences are given in the Datasets section. Datasets will be saved in /data/{split}_{dataset}.csv .
python -m src.data.preprocessOptions:
-h, --help show this help message and exit
--dataset {afew,emotic,all}
Dataset to preprocess
--data_dir DATA_DIR Root path to input dataset directory
--output_dir OUTPUT_DIR
Path to output processed CSVs
--image_size IMAGE_SIZE
Output image size (width & height)
--include_extra Include EMOTIC extra training data3.Train a single model
Pipeline for training models from /models on datasets in /data. When saved, the weights which gave the best performances can be found in /output/{model}/seed{seed}dataset-{dataset}{dataaug if data augmentation enabled}criterion-{criterion}{CCCweight{CCC weight} if criterion = "combined"}_V{V weight}_A{A weight}_D{D weight}_opt-{optimizer}_lr{learning}_backboneLRscale{backbone learning rate scale}_bs{batch size}_dropout{dropout rate}/best_model_state.pth .
python -m src.training.trainOptions:
-h, --help show this help message and exit
--seed SEED Random seed for reproducibility (default: 42)
--dataset {fer,caers,afew}
Dataset to use for training and evaluation (default: fer)
--input_size INPUT_SIZE
Image spatial resolution (default: 112)
--num_workers NUM_WORKERS
DataLoader subprocess workers (default: 4)
--early_stopping_patience EARLY_STOPPING_PATIENCE
Number of epochs with no improvement after which training will be stopped (default: 20)
--output_dir OUTPUT_DIR
Directory to save logs and model checkpoints (default: ./output)
--batch_size BATCH_SIZE
Batch size for training (default: 32)
--epochs EPOCHS Number of training epochs (default: 100)
--learning_rate LEARNING_RATE
Learning rate for the optimizer (default: 1e-4)
--weight_decay WEIGHT_DECAY
Weight decay for the optimizer (default: 5e-4)
--model {vgg11,vgg13,vgg16,vgg19,resnet18,resnet34,resnet50,efficientnet,mobilenet,mobilefacenet}
Model architecture to use (default: vgg16)
--pretrained Use pre-trained weights (default: False)
--freezed Freeze the convolutional layers of the model (default: False)
--dropout_rate DROPOUT_RATE
Dropout rate for the regression head (default: 0.5)
--optimizer {adam,sgd,adamw}
Optimizer to use for training (default: sgd)
--device {cuda,cpu} Device to use for training (default: cuda)
--grad_clip GRAD_CLIP
Gradient clipping value (default: 0.0, no clipping)
--data_augmentation Apply data augmentation during training (default: False)
--resume Resume training from a previous checkpoint (default: False)
--no_checkpoint Do not save checkpoints (default: False)
--no_model_save Do not save the model (default: False)
--criterion {mse,ccc,combined}
Loss function to use for training (default: mse)
--VAD_weights VAD_WEIGHTS
Weights for the VAD loss (default: [1.0, 1.0, 1.0])
--orth_loss_weight ORTH_LOSS_WEIGHT
Weight for the orthogonal loss (default: 0.5)
--ccc_weight CCC_WEIGHT
Weight for the CCC loss (default: 0.5)
--lr_factor LR_FACTOR
Factor by which the learning rate will be reduced (default: 0.1)
--lr_patience LR_PATIENCE
Number of epochs with no improvement after which the learning rate will be reduced (default: 10)
--lr_threshold LR_THRESHOLD
Minimum change in the monitored quantity to qualify as an improvement (default: 1e-4)
--lr_threshold_mode LR_THRESHOLD_MODE
Mode to compare the monitored quantity to the threshold (default: rel)
--lr_cooldown LR_COOLDOWN
Number of epochs to wait before resuming normal operation after a reduction in the learning rate (default: 0)
--lr_min LR_MIN Lower bound on the learning rate (default: 0.0)4.Evaluate a model
Script to evaluate models trained with the pipeline above on the test dataset.
python -m src.evaluation.testOptions:
-h, --help show this help message and exit
--input-size INPUT_SIZE
Image spatial resolution (default: 48).
--device {cuda,cpu} Device to use for evaluation (default: cuda if available, otherwise cpu).
--split {Test,Val,Train}
Data split to evaluate on (default: Test).
--state-dict-path STATE_DICT_PATH
Path to state dict with weights (.pth).5.Plot a loss curve
Plots the curves corresponding to all log.csv file located in the specified directory.
python -m src.evaluation.plot_loss_curveOptions:
-h, --help show this help message and exit
--dir DIR Directory containing log.csv files.data/: untracked, store the dataset csvsmodels/: model architectures for emotion analysisoutput/: training logs and model weightssrc/data/: data preparation and preprocessingsrc/evaluation/: evaluation scriptssrc/training/: training scriptssrc/transforms/: transforms methods for data loading and data augmentationsrc/utils/: shared dataset and utility modules
- Source: Huo, Y., & Ge, Y. (2025). VAD-Net: Multidimensional Facial Expression Recognition in Intelligent Education System.
- Available at: https://github.com/YeeHoran/VAD-Net/
- Licence: MIT License
- Sources: Kossaifi, J., Tzimiropoulos, G., Todorovic, S., & Pantic, M. (2017). AFEW-VA database for valence and arousal estimation in-the-wild, In Image and Vision Computing Dhall, A., Goecke, R., Lucey, S., & Gedeon, T. (2012). Collecting Large, Richly Annotated Facial-Expression Databases from Movies, In IEEE MultiMedia, (vol. 19, no. 3, pp. 34-41)
- Available at: https://ibug.doc.ic.ac.uk/resources/afew-va-database/
- License: Available for research purpose only