SE-AlexNet

SE-AlexNet: Enhancing Face Perception via Dimension Reduction

A collection of 34 fine-tuned convolutional neural networks for studying how Squeeze-and-Excitation (SE) modules affect facial emotion recognition. This model zoo supports the psychophysical analysis pipeline described in the companion paper.

📦 Project Page

📦 GitHub (Code & Analysis): SE-AlexNet | PsychometricFittingCurve | GradCAM-ROI-SaliencyMapDecoder

---

1. Model Details

This repository contains weights for 5 model architectures, systematically varied across 3 experimental axes:

Architecture Descriptions

⚠️ Important note on SE-Location-2: All 10 L2 variants share identical architecture (reduction=16). The squeeze-{r} label refers to a training/data configuration, not the architectural reduction ratio. This is preserved for reproducibility.

---

2. Intended Use

These models are designed for visual feature interpretability analysis in facial emotion recognition research. Specific use cases:

• Psychophysical model comparison: Compare model perceptual biases (Points of Subjective Equality, PSE) against human observers
• Grad-CAM ROI analysis: Visualize which facial regions (eyes, nose, mouth) different architectures attend to
• Dimension reduction study: Quantify how SE-based channel recalibration affects representational efficiency
• Pre-training effect study: Compare face-based (VGGFace2) vs. object-based (ImageNet) feature transfer

Not intended for: Production emotion recognition systems, clinical diagnosis, or surveillance applications. These are research models trained on controlled lab datasets.

---

3. Training Data

All models were fine-tuned on AffectNet, the largest facial expression dataset:

• Training set: 28,000 facial images (modified subset)
• Validation set: 1,000 facial images
• Classes: 8 basic emotions (Neutral, Happy, Sad, Surprise, Fear, Disgust, Anger, Contempt) — some variants use 11-class or binary labels
• Preprocessing: Standard ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

Pre-training sources:

• FaceBased: VGGFace2 (faces only — captures facial identity features)
• ObjectBased: ImageNet-1K (general objects — captures generic visual features)

Training hyperparameters:

• Batch size: 128
• Learning rate: 1e-4 (Adam optimizer)
• Epochs: 40
• Frozen conv layers (transfer learning), only SE blocks + classifier trained

---

4. Evaluation & Performance

Key Findings from Ablation Study

Psychophysical Benchmarking

Models were evaluated against human observers ($N=40$) in a 2AFC Happy/Sad discrimination task:

• Metric: RMSE between model PSE and human PSE distribution
• Statistical tests: One-sample t-tests, Cohen's d effect sizes
• Full results: See PsychometricFittingCurve repository

---

5. How to Get Started

Installation

pip install -r requirements.txt

Quick Start — Forward Pass in 10 Lines

from inference import SEModelPipeline

# Load the best model (SE-Location3, FaceBased, r=32)
pipe = SEModelPipeline('se-location3/facebased/squeeze-32')

# Run inference
probs = pipe.predict('path/to/face_image.jpg')
print(f'Prediction shape: {probs.shape}')  # (1, 11)

List Available Models

import os, json

for root, dirs, files in os.walk('.'):
    if 'config.json' in files:
        with open(os.path.join(root, 'config.json')) as f:
            cfg = json.load(f)
        print(f"{root}: {cfg['model_type']} | {cfg['pretraining']} | {cfg.get('reduction', 'N/A')}")

Load a Specific Model Programmatically

import json
from modeling import load_model_from_config

# Load config
with open('se-location3/facebased/squeeze-32/config.json') as f:
    config = json.load(f)

# Build model with weights
model = load_model_from_config(
    config,
    weights_path='se-location3/facebased/squeeze-32/model.safetensors',
    device='cpu'
)

# Forward pass
import torch
x = torch.randn(1, 3, 224, 224)  # dummy input
output = model(x)

Use for Grad-CAM Visualization

from modeling import load_model_from_config
import json

# Load any model
with open('se-location3/facebased/squeeze-32/config.json') as f:
    config = json.load(f)
model = load_model_from_config(
    config,
    'se-location3/facebased/squeeze-32/model.safetensors'
)

# Target the last conv layer for Grad-CAM
target_layer = model.features[-3]  # Last Conv2d in features
# ... apply standard Grad-CAM pipeline

---

6. Repository Structure

SE-AlexNet/
├── README.md                    # This Model Card
├── requirements.txt             # Python dependencies
├── modeling.py                  # Exact model architecture definitions
├── inference.py                 # Universal inference pipeline
├── convert_weights.py           # .pth → .safetensors conversion script
├── generate_configs.py          # Config file generator
│
├── alexnet/                     # Standard AlexNet baseline
│   ├── facebased/
│   │   ├── config.json
│   │   └── model.safetensors
│   └── objectbased/
│       ├── config.json
│       └── model.safetensors
│
├── vgg16/                       # VGG16 benchmark
│   ├── facebased/
│   └── objectbased/
│
├── se-location1/                # SE after all convolutions
│   ├── facebased/
│   │   ├── squeeze-2/ → squeeze-32/  (5 reduction ratios)
│   └── objectbased/
│       └── squeeze-2/ → squeeze-32/
│
├── se-location2/                # SE between fc6 → fc7
│   ├── facebased/
│   └── objectbased/
│
└── se-location3/                # SE in classifier (BEST)
    ├── facebased/
    └── objectbased/

---

7. Citation & Related Repositories

Companion Code Repositories

Technical Architecture

SE-AlexNet extends the classic AlexNet by inserting Squeeze-and-Excitation blocks at strategic locations:

• Squeeze: Global Average Pooling compresses spatial information into channel descriptors
• Excitation: A bottleneck MLP learns channel-wise scaling factors
• Recalibration: Feature maps are re-weighted by learned importance scores

The three location variants test the hypothesis that SE blocks are most effective when placed closer to the decision boundary (classifier), where channel-wise feature importance directly impacts classification.

---

8. Limitations

• Dataset bias: Trained only on AffectNet (posed + spontaneous expressions). Performance may degrade on in-the-wild or cross-cultural expressions.
• Image resolution: Input fixed at 224×224. Smaller or lower-quality faces may reduce accuracy.
• Binary vs. Multi-class: VGG16 models use binary classification (Happy/Sad); AlexNet/SE variants use 8-11 classes. Not directly comparable without task alignment.
• Transfer learning only: Convolutional layers are frozen. Models are not trained end-to-end from scratch.
• Architecture age: AlexNet and VGG16 are now classic architectures. These models are for scientific comparison, not state-of-the-art emotion recognition.

---

9. License

MIT License. See companion GitHub repositories for full license details.

---

Trained weights converted from original PyTorch .pth checkpoints. For questions about the training methodology or experimental design, please refer to the companion paper and GitHub repositories.