Request a Personalized Demo

Learn how to accurately convert your designs with Scan2CAD

  • This field is for validation purposes and should be left unchanged.
  • Join thousands of happy customers worldwide

Train the neural network

 

1. Select Train Menu > Train Neural Net.
2. A dialog appears. Enter the target percentage recognition accuracy you want training to achieve – 95% is a typical value. It is unrealistic to expect 100% accuracy.
3. Click OK. Scan2CAD trains the neural network.
4. Training stops when the target accuracy has been reached. Training also stops if the target has not been reached after 3 million iterations.

When training stops, an information box displaying the accuracy that was achieved and the time it took appears.

5. Click OK to close the information box.

 

Stopping and starting training

What happens during neural network training

 

Stopping and starting training

You can interrupt training whenever you want by pressing the Esc key.

When you press the Esc key, Scan2CAD finishes the current 100 iterations and then stops. Save the neural network using Train Menu > Save Neural Net As.

 

You can continue training at any time as follows:

1. Load the neural network using Train Menu > Load Neural Net.
2. Load the associated training set using Train Menu > Load Training Set.
3. Select Train Menu > Train Neural Net.

 

Scan2CAD continues training at the point where you pressed Esc.

 

What happens during neural network training

During neural network training the example characters in the training set are presented to the neural network for learning, in random order.

Each time an example character is presented to the neural network, this and the resultant calculations made by Scan2CAD are termed an “iteration”. When all the example characters in the training set have been presented to the neural network, this is called a “learning cycle”. Learning cycles are repeated until Scan2CAD has made 3 million iterations or until the target recognition accuracy has been reached.

 

At the beginning of learning, the neural network has 0% recognition accuracy – that is, it can’t recognize any of the characters in the training set.

The progression of learning from 0% accuracy to 90%+ can be visualized as a stretched “S” shape (sigmoid curve). Initially learning is slow, but once accuracy reaches around 20% it speeds up until it reaches the high 80s, when it slows down again.

 

As the neural network is trained, the Progress Bar at the bottom of the screen acts like a window over the stretched “S” learning curve, which is displayed in magenta. At the beginning of learning you will not see anything in the Progress Bar, but the magenta display will then appear as a line at the bottom of the bar. Towards the end of training it will fill most of the Progress Bar:

 

 

Note that the surface of the magenta learning curve will not be smooth – although learning follows a general curve the recognition accuracy fluctuates as learning progresses:

 

 

Learning progress is also displayed in the Iteration Display at the bottom right of the screen. This shows the percentage recognition accuracy so far and the number of iterations. You will see the recognition accuracy fluctuating as learning progresses.

 

When training stops, an information box displaying the accuracy that was achieved and the time it took appears.

 

Have questions on this topic? Talk to us