Live Training Simulation

This module performs actual live training directly in your browser. We are using a lightweight CNN to regress the 4 corner coordinates of a projected 21x21 grid.

Ground Truth vs. Prediction

Every frame generated by our data factory comes with "Ground Truth" labels—the absolute mathematical location of the corners. The network makes a "Prediction," and the difference between them is the Error (Loss).

In the visualizer, notice how the green prediction dots eventually "snap" onto the red ground truth targets as the weights optimize.

Performance Note: We are running a small model (64x64 input) with a simplified regression head to ensure it runs smoothly without a GPU.

Observing Convergence

Watch the Loss (MSE). It starts high and should drop exponentially. If it plateaus, remember the strategies from the next module: check your learning rate or consider switching to Huber loss.

Model: Ready Iteration: 0
Ground Truth
Predict
Current MSE
---
Sub-pixel Drift
---