LNSU

CS231-FER Weights

Pretrained checkpoints for the CS231 – Facial Expression Recognition Benchmark project.

Project Structure

CS231/
├── dataset/
│   ├── RAF/
│   └── list_patition_label.txt
│
├── methods/
│   ├── Advanced-FER-CNN/
│   ├── EfficientNetV2-S/
│   ├── hybrid-vit/
│   ├── LNSU/
│   └── ResNet18-Discriminative-loss/
└── README.md

---

Available Checkpoints

1. Advanced-FER-CNN

Files

best_cnn_model.h5
best_CNNModel.keras

Destination

Place either checkpoint inside:

methods/Advanced-FER-CNN/
├── best_cnn_model.h5
└── best_CNNModel.keras

---

2. EfficientNetV2-S

Files

best_checkpoint.pth
best_EfficientNetV2_RAFDB_rafdb.pth

Destination

methods/EfficientNetV2-S/
├── best_checkpoint.pth
└── best_EfficientNetV2_RAFDB_rafdb.pth

---

3. Hybrid Vision Transformer (Hybrid-ViT)

File

best_model.pth

Destination

methods/hybrid-vit/
└── best_model.pth

---

4. LNSU

Files

start_0.pt
checkpoint_step_59999_gpu_0.pt

Destination

methods/LNSU/
├── start_0.pt
└── results/
    └── checkpoint_step_59999_gpu_0.pt

Notes

start_0.pt is the initialization checkpoint required by the training pipeline.

Training script:

cd methods/LNSU
./train_exp.sh

---

DAN (ResNet18 + Discriminative Loss)

Required Backbone

resnet18_msceleb.pth

Place into:

methods/ResNet18-Discriminative-loss/models/

Experimental Notebook

MLP_Affinity_RAF_DB_Training.ipynb

The notebook reproduces all DAN experiments on RAF-DB.

Available Checkpoints

rafdb_resnet18_mlp_pretrained_ce_affinity.pth
rafdb_resnet18_mlp_pretrained_ce.pth
rafdb_resnet18_mlp_random_ce_affinity.pth
rafdb_resnet18_mlp_random_ce.pth

Placement

methods/ResNet18-Discriminative-loss/
├── rafdb_resnet18_mlp_pretrained_ce_affinity.pth
├── rafdb_resnet18_mlp_pretrained_ce.pth
├── rafdb_resnet18_mlp_random_ce_affinity.pth
└── rafdb_resnet18_mlp_random_ce.pth

Dataset

This repository does not redistribute RAF-DB.

Please obtain RAF-DB from the official source and place it in:

dataset/
├── RAF/
│   ├── train_00001.jpg
│   ├── train_00002.jpg
│   └── ...
└── list_patition_label.txt

---

Benchmark Methods

The benchmark includes:

1. Advanced-FER-CNN
2. EfficientNetV2-S
3. Hybrid Vision Transformer (Hybrid-ViT)
4. LNSU
5. DAN (ResNet18 + Discriminative Loss)

All methods are evaluated on RAF-DB using a unified project structure.