twitter-archive / torch-distlearn

A set of distributed learning algorithms for Torch

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

DistLearn

Some common distributed learning algorithms built in Torch with the help of the the ipc library.

AllReduceSGD

Spreads the computation of gradients for mini-batch of items across N processes. Uses AllReduce to quickly sum the gradients and distribute the total back out to every process.

local allReduceSGD = require 'distlearn.AllReduceSGD'(tree)
-- Make sure all the nodes start with the same parameter values
allReduceSGD.synchronizeParameters(params)
for _ = 1,epochs do
   for _ = 1,steps
      -- Compute your gradients as normal
      local grads = computeYourGrads(...)
      -- Sum and normalize them
      allReduceSGD.sumAndNormalizeGradients(grads)
      -- Do your SGD as normal
      SGD(params, grads)
   end
   -- Before validating we should make sure all nodes have
   -- the exact same parameter values
   allReduceSGD.synchronizeParameters(params)
   -- Validate...
end

When used in combination with Dataset you can quickly parallelize the processing of large datasets without a ton of effort. See the MNIST example for a complete working setup.

AllReduceEA

We also have a AllReduce based implementation of the Elastic Averaging algorithm as described in Deep learning with Elastic Averaging SGD. Its just as easy to add this to your training script, there are only two parameters required tau and alpha. Tau is how many steps to run before averaging the nodes and alpha is the weight used during the averaging step. You can read more about our implementation of AllReduceEA.

-- Use a tau of 10 and an alpha of 0.2
local allReduceEA = require 'distlearn.AllReduceEA'(tree, 10, 0.2)
-- Make sure all the nodes start with the same parameter values
allReduceEA.synchronizeParameters(params)
for _ = 1,epochs do
   for _ = 1,steps
      -- Compute your gradients as normal
      local grads = computeYourGrads(...)
      -- Do your SGD as normal
      SGD(params, grads)
      -- Average the params
      allReduceEA.averageParameters(params)
   end
   -- Make sure the center's haven't drifted too far due to
   -- floating point precision error build up
   allReduceEA.synchronizeCenter(params)
   -- Validate...
end

See a complete working example of EA and MNIST

License

Licensed under the Apache License, Version 2.0. See LICENSE file.

About

A set of distributed learning algorithms for Torch

License:Apache License 2.0


Languages

Language:Lua 100.0%