Recognize Train Your Own Model

Train Your Own Model

This is an isolated sandbox. It never affects the production recognizer.

1. Add a labeled sample

Draw a letter, then click the matching letter below to save it to your dataset.

2. Dataset

Sample counts per letter. More samples per letter generally means a more reliable model.

    3. Hyperparameters

    Controls how the network learns. Higher epochs train longer; higher learning rate adjusts faster but can overshoot.

    4. Training Log

    Live progress from your last training run, including final average loss.

    
      
    
    
    

    
    
      
    5. Test Your Model
    
      
    Draw a letter and predict with your trained model.
    
      
    
        
        
    
          
          
      
            
    • Inactive
      • 
            
    • Low
      • 
            
    • Medium
      • 
            
    • High
      • 
            
    • Very strong
      • 
          
    
        
    
        
      
    
      
    
        
      
    
      
    
    
    

    
    
      
    6. Model
    
      
    Name, export, or import your model as JSON. Export to back it up or use it elsewhere; import to continue training a saved model.