Todo.md

Todo

• LSTM

• Neuron

• NEAT connections

  • Execute
  • Gradients
  • Optimize
    • Various methods
    • Evolution methods

• adaptative learning rates

  • return cost while training

• reinforcement learning cost function

  • GAE in GPU CPU function that given a value function estimator NN and rewards, computes GAE
  • Proximal Policy optimization cost function

• server socket

  • Create log file
  • Create destructor for socket interpreter -> NN manager
  • Use poll instead of select
  • Improve security
    1. Bind to a file for security
    2. Use script for server init
       • Create file
       • Set chmod for socket file
       • Set max open fd for server
       • Start server
       • with flag --strict-security or -s
         • reboot

• client_socket

• socket functions

  • construct -- destruct
  • Get ID of pointer to a NN you pass a id and returns another id that is a reference to a existing NN it has its own execution data to train in parallel
  • training execute
  • training functions
  • save & load
  • evolution methods
  • Inference
  • delete memory

• Create logging options for training in csv format

  • Add python script for plotting it

• Paralelize layer derivative calculation calls

Bugs

• device array being used as host array in NeuronLayer and LSTMLayer remove neuron
• bug in add neuron
• constrain evolution
  • if removing neuron in a layer with 1 neuron remove layer (now it won't delete the neuron)
• constrain weights and biases to prevent nans (reset NaNs)

Not to do until all Layer constructors are made:

• Tensorflow-like class constructor
• Save, Load, Cloning
• Crossover
• Evolution
• Modify kernel launches to have more capacity of neurons (current max 1024)