A number recognition AI using Multi-Layer Preceptron (MLP) built from scratch without using TensorFlow or any other pre-built frameworks. Training images, which include 48,000 images for training and 28,000 images for testing, can be found in HERE

If you only would like to try out the application, download latest export in release page then follow Usage

1. Clone or download the project
2. Open the project with Visual Studio Code
3. Unzip train_images.zip and test_images.zip
   • Note that train_images contain subfolder 0 to 9 which contains approximately 4,000 images for each of the subfolder (48,000 images in total) and test_images contain unlabeled 28,000 images.
   • Unzipping these files may take quite amount of time.
4. After unzipping both folders, place them in the root directory. Make sure to name those folders as "train_images" (should have 0 to 9 subfolders) and "test_images" (should have 28,000 unlabeled images).
5. Follow along train_model.ipynb under learning folder.

Once you complete train_model.ipynb, you will have a trained model in HDF5 file, called trained_model_2.h5 (you may also use the one that already exists)

To compile and run the code locally, do one of the followings:

• Click run button at the top right in your editor.
• In the VS Code terminal, navigate to application directory then run the following code:

> & your/python/path/python.exe path/where/the/appPy/exists/app.py

either one will create small window.

1. From the GUI windows, click "Import Model."
2. Select an HDF5 (*.h5) file.
3. Once a canvas appears at the bottom of the screen, start drawing.
4. Click "Clear" button to erase the entire canvas.

• v1.0.0

  • Initial application

• v1.0.1

  • Execution file with example model in HDF5

• Expected script runtime: 40 minutes
  • (This may vary between PC's specification)
• sidenotes.ipynb includes problems I have faced, and how I managed to solve it.
• Running the entire script for training_model.ipynb may take long. Hence, in training_model.ipynb, press CTRL + F and search for "########## CHECK POINT ##########". This way, skip the training part that generates "trained_model.h5"
  • However, PLEASE at least read through the previous code since the comments explains what's really happening under the hood.