neuro-src

Neural network

Image classification using a neural network. Works with 28x28 black-white images

You can change:

• amount of layers (minimum amount - 2: input and output layer)
• activation function for hidden layers
• separate activation function for output layer
• batch size
• learning norm
• amount of epoch
• amount of neurons for each layer

You can see two text inputs in central panel:

• Input file - file name with input image (28x28). Program will update classification after changes appears in file. For now you need to hover over program to see new classification result (due to some egui behaviour. Maybe will be fixed later)

• Training data load/save path:

  • if given path is a directory and you clicked Load: it is treated as a directory with training data. Each training image need to have name as label_*. Also directory should contain train.json with all labels in array. Index of label in array is equal to activated neuron for label.

  Example of train.json : ["auto","heli","plane","ship"]

  • if given path is a file and you clicked Load: it is treated as saved trained neural network and it is loaded in app. You can't change loaded neural network.
  • if given path is a file and you clicked Save: it will save your neural network to file (JSON format).

Drop button just drops current neural network from app, so you can create new one.

(in the example gif - classification of types of vehicles)

You can transform all images in IMG_DIR into training sample with this python script:

python import numpy as np
from PIL import Image
import cv2
import os 
IMG_DIR = ‘img’ 
for img in os.listdir(IMG_DIR):
    img_array = cv2.imread(os.path.join(IMG_DIR,img), cv2.IMREAD_GRAYSCALE) 
    img_pil = Image.fromarray(img_array)
    img_28x28 = np.array(img_pil.resize((28, 28), Image.LANCZOS)) 
    data = Image.fromarray(img_28x28) 
    data.save(img)