DOODLE NUMBERZ

Video Demo:

https://youtu.be/jVBaQSjgU5c

Description:

• website that lets you draw a number in a field
  • neural network (python) guessing the number after every stroke
  • show all possibilities for 0-9
• ability to train the AI
  • by clicking correct answer (no matter if guess was right or wrong)
  • app.py can be changed to always start with the better of initial/trained model (### comments)
• ability to compare initial model with trained one
  • show loss functions by clicking eval button
  • initial model depends on app.py > stupid, genius, trained

Design choices:

In the 'shipped' version i decided to always load the most stupid model (initialized with random parameters).

Like the heading 'AI TRAINER 3000' suggests, you are now able to train a model that has never seen numbers, to a model that recognizes them after a few trainings.

I am aware that one can always click the training buttons, no matter if they are wrong, or there isn't even a number drawn. But if a user decides to do this, he/she will just have a more stupid model and would be responsible for it.

(i tried out different settings to really get the feeling that bad training results in a bad network, even though 'good data' from mnist gets included in every train - depends on weight of user-inputs, sample size of mnist, number of training epochs)

The training always takes all weighted user samples since site refresh and n random samples from the mnist data set. Although i am aware that shuffling mnist data set on every train takes much time, this is the result of experimenting with different training sets and i think the current version gives a good sense of training with small sample sizes for the user.

I kept the styling (CSS) simple and am also aware that the site might not be scalable but my focus was on the following lessons:

• sending data from frontend to backend and reverse
• single site application (no reloading / subsites)
• using, training, evaluating neural network
• making the site robust (i.e. handling inputs, while training)

Parameters:

• initial model
• user input weight
• mnist size on training
• training epochs

Files:

app.py

• flask app

model_init.ipynb

• model built and config

/models

• model saves

/static & /templates

• JavaScript, CSS, icon, html

Toolset:

• python
  • numpy
  • flask
  • tensorflow keras
  • jupyter notebook
• JavaScript
  • asynchronous, fetch API
  • JSON
• HTML
  • templating (JINJA)
• CSS