Weights and Biases support package for Neuron

NeuronRemoteLogger

Package that connects the Neuron package to various remote loggers.

Supported Loggers

Loggers OS Support External reqs
Weights and Biases macOS python 3, wandb package, numpy package, pillow package

Protocol

You can implement your own Remote logger by conforming to the below protocol.

public enum Remote {
  case wandb
}

public protocol RemoteLogger {
  associatedtype LogPayload
  associatedtype InitPayload
  
  var type: Remote { get }
  
  init?(payload: InitPayload)
  func setup() throws
  func log(payload: LogPayload) throws
}

Usage

Initialize the logger of your choice. Below is the Wandb example

Note: On first run, if you havnt signed into your Wandb account using wandb login in the CLI, the LLDB will ask you for your login API key through the debugger. Follow the instructions


let epochs = 10
let lr = 0.01

let payload = Wandb.InitializePayload(projectName: "NeuronTest",
                                      config: ["learning_rate": lr.pythonObject,
                                                "epochs": epochs.pythonObject])
guard let wandb = Wandb(payload: payload) else {
  fatalError()
  return
}
      

Call setup on your logger

do {
  try wandb.setup()
} catch {
  fatalError(error.localizedDescription)
  return
}

Start logging your data

let offset = Double.random(in: 0...1) / 5

for epoch in 0..<epochs {
  let e = Double(epoch)
  let acc = 1 - pow(2, -e - Double.random(in: 0...1) / e - offset)
  let loss = pow(2, -e + Double.random(in: 0...1) / e + offset)
  
  // this payload can contain any arbitrary tracking data as long as it's convertable to a PythonObject
  // This is specific to wandb. Other loggers will define their own payload.
  let payload = ["accuracy": acc.pythonObject, "loss": loss.pythonObject]
  
  do {
    try wandb.log(payload: payload)
  } catch {
    XCTFail(error.localizedDescription)
  }
}

Some of the loggers might support image data upload. For example Wandb has a function for building a compatable image object from a 3D array from a Tensor

func buildImage(data: [[[Float]]], name: String) -> PythonObject?

You can then log this in the payload by just adding it as a key:

payload["examples"] = buildImage(data: imageData, name: "Generator Image")

For Wandb login to your dashboard to see the results

GitHub

View Github