Overview

This repo focuses on the development, training, evaluation, and deployment evaluation of lightweight deep learning models optimized for resource-constrained embedded devices.

Key features:

• Model Training: Modified MobileNetV2 architecture for small input sizes
• Testing & Evaluation: CPU-based evaluation using accuracy, confusion matrices, classification predictions, and inference latency.
• Post-Training Quantization: Reduces model size and improves inference speed on embedded devices with minimal accuracy loss.
• Dataset: CIFAR-10

---

Code Structure

Scripts/
│
├── main.py                 # Entry point for training, testing, quantization[or combined via argument at run time]
├── train.py                # Model training module
├── test.py                 # Evaluation and testing 
├── quantize_model.py       # Post-training quantization pipeline
├── data_loader.py          # Dataset loading 
├── model_network.py        # Model architecture
├── params_init.py          # Model weights initializer
├── plot_log.py             # Model architecture
├── helpers/                # utils functions
│   ├── global_import.py    # packages
│   └── quantize_helper.py  

Installation

Follow the steps below to set up the required Python environment and install the dependencies.

1. Clone this project repo:

git clone git@github.com:Go-Ab1/Split_DL.git

2. Create and activate a Python virtual environment[Recommended]

 python3 -m venv pytorch_env
 source pytorch_env/bin/activate 

3. Install dependencies:

# The packages listed are used for testing
# Adjust the PyTorch installation to match your CUDA version.
pip install -r requirements.txt

Usage

Once the dependencies are installed, run the scripts to reproduce the experiments.

# ==========================================
#               PROJECT USAGE
#   (Training, Testing, Quantization )
# ==========================================

# ------------------------------------------
# 1. TRAINING
# ------------------------------------------
# Trains 
# - Downloads dataset on first run
# - Logs training/validation loss & accuracy
# - Saves training log: media/
# - Saves trained model: models/

python3 Scripts/main.py train

# Visualize logs after training:
# - Plots loss curves
# - Plots accuracy curves

python3 Scripts/plot_log.py

# ------------------------------------------
# 2. TESTING
# ------------------------------------------
# Evaluates the trained model
# Produces:
# - Overall accuracy
# - Classification
# - Confusion matrix
# - Inference latency (batch & per-image)

python3 Scripts/main.py test

# ------------------------------------------
# 3. POST-TRAINING QUANTIZATION (PTQ)
# ------------------------------------------
# Runs quantization on the trained model.
# Produces:
# - Quantized model
# - Model size comparison
# - Latency comparison (FP32 vs INT8)
# - Accuracy comparison

python3 Scripts/main.py quant

# ------------------------------------------
# 4. CHAINED WORKFLOWS
# ------------------------------------------

# Train → Test
# (Complete pipeline for performance evaluation)

python3 Scripts/main.py train test


# Train → Quantize
# (Complete pipeline for optimizing deployed models)

python3 Scripts/main.py train quant

Configuration Parameters

All parameters can be modified in config.yaml and can be adjusted as required.

data_dir: Dataset
batch_size: 64
num_epochs: 100
learning_rate: 0.01
alpha: 1.0
media_log_dir: media
model_log: training_log.txt
test_log: test_log.txt
val_split: 0.1
num_workers: 4
models_dir: models
compare_models: true
quant_backend: fbgemm
num_runs_latency: 5
comparison_log_name: comparison_log.txt
trained_model_name: final_model.pth
quantized_model_name: quantized_model.pth

Results

Training Curve (Accuracy & Loss)

The training and validation performance of the model is shown below.

Training/Validation Accuracy

Training/Validation Loss

Sample Predictions & Confusion Matrix

Below are example predictions and the confusion matrix for CIFAR-10 test images.

Sample Predictions

Confusion Matrix

Model Comparison Summary After Quantization

Metric	Full-Precision Model	Quantized (PTQ) Model	Change
Accuracy	0.9139	0.9125	↓ 0.15%
Model Size (MB)	9.2820	2.6434	↓ 71.52%
Latency / Batch(32) (s)	0.5283	0.2276	↓ 56.92%
Latency / Image (s)	0.0165	0.0071	↓ 56.92%

Nov, 2025!!!
@Goitom