‹ projects

cluster-rnn

a distributed Torch7 RNN cluster over MPI
Log | Files | Refs | README

commit 4ccfa17fc00882ad3994249ce098b9ed2c83a901
parent 71251f886db6cb5bd4d1e4cde1685fa359565708
Author: umhau <umhau@users.noreply.github.com>
Date:   Thu, 16 Feb 2017 12:24:58 -0500

troubleshooting uncreated initial files

Diffstat:
Mmlaunch.lua | 11+++++++----
Mutil/CharSplitLMMinibatchLoader.lua | 2++
2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/mlaunch.lua b/mlaunch.lua @@ -14,6 +14,9 @@ core on each available machine in the cluster - so the 'ranks' below will range from 0-7 if there are two machines with 4 CPU cores each. I don't have any GPUs, so I can't speak to how those are presented. +The paper refers to some of the parameters with greek symbols. These are noted +as relevant in the variables section below. + --]] -- VARIABLES ------------------------------------------------------------------ @@ -23,10 +26,10 @@ local torchfile = 'train.lua' -- name of torch file to run with MPI local iterations = 10 -- i.e., epochs. don't need that many for testing. -- advanced parameters -local communicationPeriod = 64 -local movingRateAlpha = 0.001 --0.9/6 -local learningRate = 5e-3 -local momentum = 0.99 +local communicationPeriod = 64 -- \tau +local movingRateAlpha = 0.001 --0.9/6 -- \alpha +local learningRate = 5e-3 --\eta +local momentum = 0.99 -- \delta -- very advanced parameters local learningRateDecay = 1e-4 diff --git a/util/CharSplitLMMinibatchLoader.lua b/util/CharSplitLMMinibatchLoader.lua @@ -22,6 +22,8 @@ function CharSplitLMMinibatchLoader.create(data_dir, batch_size, seq_length, spl -- fetch file attributes to determine if we need to rerun preprocessing local run_prepro = false + print(vocab_file) + print(tensor_file) if not (path.exists(vocab_file) or path.exists(tensor_file)) then -- prepro files do not exist, generate them print('vocab.t7 and data.t7 do not exist. Running preprocessing...')