update-README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,60 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
base_model:
|
| 6 |
+
- WinKawaks/vit-tiny-patch16-224
|
| 7 |
+
datasets:
|
| 8 |
+
- deanngkl/ferplus-7cls
|
| 9 |
+
- deanngkl/raf-db-7emotions
|
| 10 |
+
- deanngkl/affectnet_no_contempt
|
| 11 |
+
metrics:
|
| 12 |
+
- accuracy
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Vision Transformer for Facial Expression Classifier
|
| 16 |
+
|
| 17 |
+
A deep learning project that fine-tunes a Vision Transformer (ViT-Tiny) model for 7-class facial emotion classification using cleaned versions of FER+, AffectNet, and RAF-DB datasets.
|
| 18 |
+
|
| 19 |
+
## 📌 Project Highlights
|
| 20 |
+
|
| 21 |
+
- 🔍 7-class emotion classification: `['anger', 'disgust', 'fear', 'happiness', 'neutral', 'sadness', 'surprise']`
|
| 22 |
+
- 🧠 Model: ViT-Tiny (`timm` implementation)
|
| 23 |
+
- 🎯 Achieved 82% validation accuracy on a blended hold-out set (8 377 images)
|
| 24 |
+
- 📚 Cleaned & uploaded datasets to Hugging Face Datasets
|
| 25 |
+
- 🧪 Integrated CutMix, cosine decay scheduler, and AMP for training
|
| 26 |
+
|
| 27 |
+
---
|
| 28 |
+
|
| 29 |
+
## 📦 Datasets
|
| 30 |
+
|
| 31 |
+
| Dataset | Link | Notes |
|
| 32 |
+
|-------------|----------------------------------------------------------------------|--------------------------------|
|
| 33 |
+
| FER+ | [Hugging Face](https://huggingface.co/datasets/deanngkl/ferplus-7cls) | Filtered to 7 basic emotions |
|
| 34 |
+
| AffectNet | [Hugging Face](https://huggingface.co/datasets/deanngkl/affectnet_no_contempt) | Removed 'contempt' class |
|
| 35 |
+
| RAF-DB | [Hugging Face](https://huggingface.co/datasets/deanngkl/raf-db-7emotions) | Added proper emotion labels |
|
| 36 |
+
|
| 37 |
+
The total amount of datasets
|
| 38 |
+
|
| 39 |
+
```html
|
| 40 |
+
Loaded 75398 training samples from 3 sources
|
| 41 |
+
Loaded 8377 validation samples from 3 sources
|
| 42 |
+
Training-set distribution:
|
| 43 |
+
0: 0 : 9738
|
| 44 |
+
1: 1 : 3385
|
| 45 |
+
2: 2 : 4313
|
| 46 |
+
3: 3 : 18315
|
| 47 |
+
4: 4 : 20987
|
| 48 |
+
5: 5 : 9289
|
| 49 |
+
6: 6 : 9371
|
| 50 |
+
Emotion batch torch.Size([64, 3, 224, 224])
|
| 51 |
+
```
|
| 52 |
+
|
| 53 |
+
---
|
| 54 |
+
|
| 55 |
+
🙋♂️ Author
|
| 56 |
+
### Dean Ng Kwan Lung
|
| 57 |
+
Blog : [Portfolio](https://kwanlung.github.io/)
|
| 58 |
+
LinkedIn : [LinkedIn](https://www.linkedin.com/in/deanng00/)
|
| 59 |
+
GitHub : [GitHub](https://github.com/kwanlung)
|
| 60 |
+
Email : kwanlung123@gmail.com
|