---
license: apache-2.0
datasets:
- clip-benchmark/wds_fer2013
base_model:
- microsoft/resnet-50
pipeline_tag: image-classification
library_name: keras
---
## Model Description

This model is a **ResNet-50** deep convolutional neural network fine-tuned for the **FER-2013 (Facial Expression Recognition 2013)** dataset. The dataset consists of low-resolution (48x48) grayscale images of faces categorized into seven core emotional states.

This project focused on maximizing the performance of the pre-trained ResNet-50 architecture on this particularly challenging, noisy, and imbalanced dataset.

## Training Details

### Architecture

  * **Base Model:** ResNet-50 (pre-trained on ImageNet).
  * **Head:** Custom dense layers (224 units) with a high 0.5 dropout rate.
  * **Transfer Learning Strategy:** **Deep Freezing**. The model base was frozen up to the `conv5` block, meaning only the final convolutional block (`conv5`) and the custom head were fine-tuned. This prevents early layers, which are optimized for high-resolution images, from being corrupted by the 48x48 input.

### Optimization & Regularization

| Technique | Rationale |
| :--- | :--- |
| **Class Weighting** | Applied inverse frequency weights to mitigate the severe class imbalance (e.g., Disgust is rare, Happy is abundant). |
| **Data Augmentation** | Used random flips, translations, rotations, and zooms to artificially expand the small dataset and combat overfitting. |
| **High Dropout** | Increased dropout to 0.5 to aggressively regularize the model and prevent the divergence seen in earlier training runs. |
| **Optimizer** | Adam with a very low fine-tuning learning rate of 5e-6. |

## Evaluation Results

The final model achieved its **highest stability and best performance** after 50 epochs of fine-tuning, demonstrating strong generalization given the difficulty of the data.

### Overall Performance

| Metric | Result |
| :--- | :--- |
| **Test Accuracy** | **45.70\%** |
| **Test Loss** | 1.4929 |
| **Training Accuracy (End)** | 63.25\% |

### Per-Class F1-Scores

The F1-Score highlights the model's difficulty with ambiguous negative emotions.

| Emotion | F1-Score | Support (Test Count) | Notes |
| :--- | :--- | :--- | :--- |
| **Neutral** | **0.6386** | 831 | Highest precision, well-distinguished class. |
| **Happy** | 0.6037 | 1774 | Strongest recall, the most abundant class. |
| **Disgust** | 0.4659 | 111 | Significantly improved performance on this rare class. |
| **Sad** | $0.3995$ | 1233 | Ambiguous. |
| **Surprise** | 0.3531 | 1247 | Ambiguous. |
| **Fear** | 0.3374 | 1024 | Ambiguous. |
| **Angry** | **0.3312** | 958 | Lowest F1-score, indicating high confusion. |

## 💡 Usage and Limitations

### Inputs

  * **Image Format:** Grayscale (48x48 pixels).
  * **Normalization:** Pixel values must be scaled to [0, 1] (by dividing by 255.0).

### Recommended Libraries

  * `tensorflow` (for loading the model)
  * `numpy` (for array manipulation)

### Limitations

1.  **Low Accuracy:** The 45.70\% accuracy is limited by the **low resolution** (48x48) and **noisy labels** of the FER-2013 dataset. It is not comparable to modern human performance (65\%-68\% on FER-2013) or models trained on high-quality, high-resolution "in-the-wild" datasets like AffectNet.
2.  **Overfitting:** Despite aggressive regularization, the model remains highly overfit (Training vs. Test gap), which is characteristic of this dataset.

### ❓ Troubleshooting the Error

If you encounter `ValueError` upon loading, ensure you are loading the model with the `.keras` extension:

```python
import tensorflow as tf
loaded_model = tf.keras.models.load_model("./best_fer_resnet_local/best_model.keras")
```