Source codes for our ACII'21 paper: Contrastive Learning for Domain Transfer in Cross-Corpus Emotion Recognition.
- This code is tested with PyTorch 1.2.0 and Python 3.7.6
Aff-Wild2 and SEWA. All the frames should be cropped and aligned.
IDs for face warping with Aff-Wild2 and SEWA are provided in Aff-Wild2-train-id.txt, Aff-Wild2-val-id.txt, and SEWA-id.txt.
Put these datasets into the folder "data". The directory structure should be modified to match:
├── code
| ├── ..
├── contrastive
| ├── ..
├── mini_datasets
| ├── Aff-Wild2
| | ├── train.csv
| | ├── val.csv
| | ├── test.csv
| ├── SEWA
| | ├── train.csv
| | ├── val.csv
| | ├── test.csv
├── data
| ├── Aff-Wild2
| | ├── cropped_aligned
| | | ├── ..
| ├── SEWA
| | ├── porep_SEWA
| | | ├── ..
| ├── Real
| | ├── Aff-Wild2_v3
| | | ├── ..
| ├── Fake
| | ├── Aff-Wild2_v3
| | | ├── ..
Train base model with Aff-Wild2/SEWA:
python train_bl.py --source Aff-Wild2/SEWA --label arousal/valence
Train domain adaptation models with Aff-Wild2 and SEWA:
python train_dann.py/train_dan.py/train_adda.py --source Aff-Wild2 --target SEWA --label arousal/valence
Pre-train FATE with Aff-Wild2 and SEWA (in /contrastive folder):
python pretrain.py
Fine-tune FATE with SEWA (in /code folder):
python fine-tune.py --label arousal/valence
Test on SEWA (base model and domain adaptation baselines):
python test_models.py --source Aff-Wild2 --target SEWA --model bl/dann/dan/adda
Test on SEWA (FATE):
python test_FATE.py --source Aff-Wild2 --target SEWA
Please put /checkpoints folder in /code folder and /checkpoints_v3 folder in /contrastive folder when loading weights. All the weights are available on Google Drive.
In /checkpoints folder, /model folder saves the weights for FATE.