DEV Community: Axel

🧠 CNN from Scratch

Axel — Fri, 08 Aug 2025 14:38:37 +0000

This project is a simple Convolutional Neural Network (CNN) implemented entirely from scratch using only low-level libraries like NumPy, PIL, and SciPy—no deep learning frameworks (e.g., TensorFlow or PyTorch) are used. It includes image preprocessing, convolution and pooling operations, ReLU and softmax activations, forward/backward propagation, and a fully connected classifier.

📦 Releases

Version	Latest	Stable	Test a trained model
0.1.1	✅	✅	Test
0.1.0	❌	✅	Test

🚀 Features

Manual image preprocessing (RGB separation, resizing, normalization)
Handcrafted convolution and max-pooling operations
Fully connected layers (L1, L2, and output)
Softmax + Cross-Entropy Loss
Mini-batch gradient descent with backpropagation
Model saving/loading using pickle
Class prediction on new images
Realtime training visualization using matplotlib

🖼 Dataset Structure

Make sure your dataset folder is structured like this:

data/
├── class1/
│   ├── image1.png
│   ├── image2.png
├── class2/
│   ├── image1.png
│   ├── image2.png
├── class../
│   ├── ..
..

Each subfolder represents a class (e.g., cat, dog), and contains sample images.

To help you get started, we’ve included a starter data folder with example class directories.

🧪 How It Works

Image Preprocessing:

Each image is resized to a fixed size and normalized.
Filters (e.g., sharpening, edge detection) are applied using 2D convolution.
ReLU activation and 2×2 max-pooling reduce spatial dimensions.

Feature Vector:

Flattened pooled feature maps are fed into fully connected layers.

Feedforward + Softmax:

Dense layers compute activations followed by a softmax for classification.

Backpropagation:

Gradients are computed layer-by-layer.
Weights and biases are updated using basic gradient descent.

🛠 Setup

pip install git+https://github.com/77AXEL/CNN-FS.git

✅ Training

Update and run the training block:

from cnnfs.model import CNN

model = CNN()
model.init(
    image_size=64,
    batch_size=32,
    h1=128,
    h2=64,
    learning_rate=0.001,
    epochs=400,
    dataset_path="data", # Your dataset folder path
    max_image=200, # If not specified, the model will load all images for each class
    filters=[
        [[0, -1, 0], [-1, 5, -1], [0, -1, 0]],
        [[1, 0, -1], [1, 0, -1], [1, 0, -1]],
        [[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]]
    ] # If not specified, the model will use its own default filters
)
model.load_dataset() # Processes all images for each class to prepare them for later use in training
model.train_model(visualize=True) # Starts model training based on the classes in your dataset with optional visualization support
model.save_model() # Stores the trained model's weights and biases in a model.bin file

🔍 Predicting New Images

model.load_model("model.bin") # Load the trained model
prediction = model.predict("test_images/mycat.png") # Applies the trained model to classify the input image
print("Predicted class:", prediction)

💡 Example Filters Used

[ [0, -1,  0],   Sharpen
  [-1, 5, -1],
  [0, -1,  0] ]

[ [1,  0, -1],   Edge detection
  [1,  0, -1],
  [1,  0, -1] ]

[[-1, -1, -1],   Laplacian
 [-1,  8, -1],
 [-1, -1, -1] ]

📊 Performance

Metric	Value (example)
Accuracy	~90% (binary class)
Epochs	10–50
Dataset	Custom / ~8000 imgs

* Note that a larger dataset and more training epochs typically lead to higher accuracy.

🧠 Concepts Demonstrated

CNNs without frameworks
Data vectorization
Forward and backward propagation
Optimization from scratch
One-hot encoding for multi-class classification

📦 Dependencies

📜 License

MIT License — feel free to use, modify, and share.

🤝 Contributing

PRs are welcome! You can help:

Add evaluation functions
Improve filter design
Extend to grayscale or multi-channel separately
Parallelize dataset loading

. Github repo: https://github.com/77AXEL/CNN-FS
. Github page: https://77axel.github.io/CNN-FS

Web2APK

Axel — Thu, 27 Jun 2024 15:05:40 +0000

Presentation:

Our GitHub repository houses a transformative project that automates the conversion of HTML, CSS, and JavaScript front-end projects into Android applications. This tool streamlines the process, enabling developers to port their web projects to Android without extensive manual effort, enhancing cross-platform development efficiency. #### Installation:
If Git is not installed, you can obtain the tool by clicking the Download button
If Git is already installed, you can utilize this command:

git clone https://github.com/77AXEL/Web2APK

Use

To use the tool, follow these steps:
1) Develop a front-end project similar to this example:

2) Compress the project folder into a ZIP file:

3) Navigate to the Web2APK directory and run this cammand:

python web2apk.py -zip path_to_your_zip_file -icon path_to_your_desired_icon -name your_desired_app_name

Once you run this command, the tool will start compiling and building the APK file. After compiling, you will get output like this:

Finally, you will find the compiled APK in the dist directory:

Note:

Using the WebP image format for the app icon is recommended.

If you encounter any problem or issue with the tool, you can check the build.log and sign.log files located in the log folder

Using this tool requires having the JAVA JDK and ANDROID SDK installed, with their paths, JAVA_HOME and ANDROID_HOME, set in your environment path

If you don't have them installed yet, follow those links: Java JDK Android SDK

Platforms

Supported Platform : Windows, Mac-OS, Ubuntu/Debian/Kali/Parrot/Arch Linux

If you like this project, star or sponsor our repo on github from here Web2APK

NetDisco

Axel — Fri, 21 Jun 2024 15:53:22 +0000

Presentation:

A versatile app designed for both Windows and Linux platforms. This tool efficiently scans the LAN network, displaying detailed information such as IP addresses, MAC addresses, and host names of devices connected to the same Wi-Fi network as the host PC. Its capability to execute seamlessly across different operating systems enhances user accessibility and provides valuable insights into network connectivity, making it an invaluable tool for both home and professional use.

Use:

When you run the app, it initiates a scan of your local LAN network, systematically searching for connected devices. The interface displays ongoing progress as it identifies each device, showcasing detailed information such as IP addresses, MAC addresses, and host names. This real-time visibility into your network environment provides a comprehensive view of all devices connected to the same Wi-Fi network as your PC, ensuring you stay informed about your network's status and security
Captures:

If you want to initiates another scann press Update button:

Download:

You can download the app by pressing Download
On Windows, you can launch the executable directly. However, on Linux-based systems, you'll need to use an emulator like Wine to run the application:

wine NetDisco.exe

This setup allows the app to function seamlessly across different operating environments, ensuring compatibility and accessibility for users on varying platforms

URLs:

You can also view the project on GitHub: https://github.com/77AXEL/NetDisco/

DEV Community: Axel

🧠 CNN from Scratch

📦 Releases

🚀 Features

🖼 Dataset Structure

To help you get started, we’ve included a starter data folder with example class directories.

🧪 How It Works

🛠 Setup

✅ Training

🔍 Predicting New Images

💡 Example Filters Used

📊 Performance

* Note that a larger dataset and more training epochs typically lead to higher accuracy.

🧠 Concepts Demonstrated

📦 Dependencies

📜 License

🤝 Contributing

Web2APK

Presentation:

Use

Platforms

NetDisco

Presentation:

Use:

Download:

URLs:

To help you get started, we’ve included a starter `data` folder with example class directories.