Skip to content

Latest commit

 

History

History
44 lines (31 loc) · 2.28 KB

flow.md

File metadata and controls

44 lines (31 loc) · 2.28 KB
Raw

Start the Erlang Shell for Master and Slaves

  • Run rebar3 shell --sname master@localhost to start the master.
  • Run rebar3 shell --sname slave1@localhost and rebar3 shell --sname slave2@localhost to start two slaves.

Initialization of the Master Node

  • Create two Python processes: one for the model and the other for the UI.
  • Initialize the Python communication handler for messaging with the Erlang master.
  • Create an Erlang process to perform a looping operation and handle all incoming messages.

Initialization of Slave Nodes

  • Spawn a process on each node found on the specified network.
  • Each node will spawn a Python process and register its message handler.
  • The node enters a loop for message handling.

Loading the Dataset

  • The master sends a message to all the Erlang nodes, requesting them to load the dataset.
  • Once the Python process loads the dataset, the node will send an acknowledgment (ack).
  • When all acks are collected, the Erlang master will continue its loop.

Distributing the Model

  • The master asks the Python model to provide the model.
  • The master sends a message to all the Erlang nodes, requesting them to load the provided TensorFlow model.
  • Once the Python process loads the model, the node will send an ack.
  • When all acks are collected, the Erlang master will continue its loop.

Distributing Weights

  • The master asks the Python model to provide the weights of the network.
  • The master sends a message to all the Erlang nodes, requesting them to load the provided weights.
  • Once the Python process loads the weights, the node will send an ack.
  • When all acks are collected, the Erlang master will continue its loop.

Training

  • The master sends a message to all the Erlang nodes, requesting them to train their local models for a specified number of epochs.
  • Once the Python process finishes training, the node will send an ack.
  • When all acks are collected, the Erlang master will send the collected weights to the Python model, which will perform a weighted average and update its weights.
  • I ask the python model to provide the weight of the new network and distribute them across different nodes.
  • When the nodes sends back an ack and there are still some epochs to be trained, the Erlang master will restart this sequence of steps.