Repository navigation
Training Controls
The training controls determine how the network advances through training.
The main playback controls live in the application header. They use compact icon buttons for reset, play/pause, step, epoch, and playback speed.
Step advances one visual computation step.
It is useful for carefully inspecting:
- Individual terms in the forward pass.
- Node-level computations.
- Layer-level progress.
- Gradient propagation during backpropagation.
Epoch runs one full epoch using the same visual step mechanism as Step.
Instead of updating the network instantly, it animates the full pass until the epoch is complete. At the end of the epoch, gradients are applied and the loss chart is updated.
Play continuously runs animated epochs.
When one epoch finishes, the network updates and the next epoch begins automatically.
Use Pause to stop continuous training.
The speed control changes how quickly Play advances through the visual training steps.
Available speeds:
0.5x1x2x4x
Click the speed control to cycle through the available speeds. Lower values are useful when inspecting formulas and node updates. Higher values are useful when observing longer training behavior.
Reset rebuilds the network using the current configuration.
It also clears:
- Loss history.
- Generated table data.
- Training state.
If the initializer is custom, weights and biases return to 0.
If the initializer is he or xavier, weights are reinitialized and biases are randomized.
The learning rate controls the size of each parameter update.
Higher values make learning faster but can make training unstable.
Lower values make learning slower but can make updates smoother.