Skip to content

DESCRIPTION.md

Latest commit

 

History

History
46 lines (39 loc) · 3.45 KB

DESCRIPTION.md

File metadata and controls

46 lines (39 loc) · 3.45 KB
Raw

Description

This file explains all definitions and members important for the math of the network class.

Members

All Members of the Network.cpp/Network.h class:

  • layers_[][]: An array of all layers (which contain neurons) in the network. (double)
    • Note:
    • layers_[0] would target the input layer.
    • layers_[0][0] would target the first neuron's value (top one in topology) in the input layer.
  • biases_[][]: An array of biases for every neuron. (double)
    • Note:
    • biases_[0] would target the input layer's biases.
    • biases_[0][0] would target the bias for the first neuron (top one in topology) in the input layer. (Every layer but input layer has biases)
  • weights_[][][] An array of weights between neurons. (double)
    • Note:
    • weights_[0] would target the input layer's weights (or neurons in this case).
    • weights_[0][0] would target the weights going out from the first neuron (top one in topology) in the input layer.
    • weights_[0][0][0] would target the weight between the first neuron (top one in topology) in the input layer and the first neuron in the next layer (first hidden layer).
  • topology_ An object representing the network's topology (containing sizes & counts and weights). (network_topology)
  • learn_rate_ A value indicating the learn rate of the network. (double)
    • Note:
    • This can be a constant: 0.1 or be updated after every train: 1 / (train_times + 1).
    • The higher the learn rate, the faster the network learns - but be aware your network will get unaccurate
  • layers_count_ The total count of layers in this network (e.g.: Input + 3 Hidden + Output = 5 total). (int)
    • Note:
    • Mostly used to get last element in arrays. (layers_count_ - 1)

Definitions

All compile time (#define) Definitions inside the source/header files:

  • BrabeNetz/Network.h
    • LEARNING_RATE: The default learning rate if none is manually set (0.17)
    • STATE_FILE: The default filepath to the network's state file ("state.nn")
    • FORCE_MULTITHREADED: Force use of multithreading for Backwards-Propagation algorithm, only use on larger networks (hidden layers with >5 neurons) (Only in Release mode) (false)
    • ITERS_PER_THREAD: Amount of iterations in the Backwards-Propagation algorithm each thread is expected to do (ITERS_PER_THREAD * core_count are the minimum expected iterations of backprop to start multithreading instead of serial if FORCE_MULTITHREADED is false) (thread spawning takes ~270.000ns, the loop ~250ns) (600)
  • BrabeNetzConsole/Trainer.cpp
    • CONST_LEARN_RATE: Use a constant learn rate for training instead of using the thumb-rule formula (true)
    • PRINT_OUTPUT: Print the output of the training to console (use only for debug, I/O like printing is very slow, training will take ~0.02ms longer per iteration) (true)
    • UPDATE_STATUS: Update Titlebar of Console Window with status of iteration (i/total) (use only for debug, I/O like setting title is slow, training will take ~0.12ms longer per iteration) (true)
  • BrabeNetzConsole/BrabeNetzConsole.cpp:
    • LOAD_STATE: Load the network's state when starting the program instead of generating a new random network each time (false)
    • TRAIN_TIMES_EACH: Amount of times to train the network (depending on the training this has to scale, e.g. 4000 / 4 for XOR, so 1000 for each possibility) (4000)