/10807

Primary LanguageLua

Installation

The code depends on Torch http://torch.ch. Follow instructions here.

Example

Standard convnet

th Train.lua "models/simple.lua CifarProcessor.lua -flip 0.5 -minCropPercent 0.8" /data/cifar10/trainval.txt out/output.t7 -val /data/cifar10/test.txt -valSize -1 -valEvery 1 -batchSize 128 -epochSize -1 -epochs 100 -learningRate 0.01 -LRDropEvery 30 -LRDropFactor 5
th Forward.lua "out/output.t7 CifarProcessor.lua" /data/cifar10/test.txt -batchSize 128

Weighted version that doesn't work

th Train.lua "models/DeepTreeWeighted.lua -channels 64 -preconv 3 -depth 3 -nodedepth 2 -pool 3,5 CifarDeepTreeWeightedProcessor.lua -flip 0.5 -minCropPercent 0.8" /data/cifar10/trainval.txt out/output.t7 -val /data/cifar10/test.txt -valSize -1 -valEvery 1 -batchSize 128 -epochSize -1 -epochs 100 -learningRate 0.01 -LRDropEvery 30 -LRDropFactor 5
th Forward.lua "out/output.t7 CifarDeepTreeWeightedProcessor.lua" /data/cifar10/test.txt -batchSize 128

With pretraining (spoilers: pretraining doesn't work)

th TrainDeepTreeWithPretraining.lua "models/DeepTreeRootOnly.lua -channels 64 -preconv 3 -depth 3 -nodedepth 2 -pool 3,5 CifarDeepTreeProcessor.lua -flip 0.5 -minCropPercent 0.8 [-greedy]" /data/cifar10/trainval.txt out/output.t7 -val /data/cifar10/test.txt -valSize -1 -valEvery 1 -batchSize 160 -epochSize -1 -pretrainEpochs 10 -epochs 100 -learningRate 0.01 -LRDropEvery 30 -LRDropFactor 5
th Forward.lua "out/output.t7 CifarDeepTreeProcessor.lua [-forwardWeighted]" /data/cifar10/test.txt -batchSize 128

Optimizing full tree from random initialization

th Train.lua "models/DeepTree.lua -channels 64 -preconv 3 -depth 3 -nodedepth 2 -pool 3,5 CifarDeepTreeProcessor.lua -flip 0.5 -minCropPercent 0.8 [-greedy]" /data/cifar10/trainval.txt out/output.t7 -val /data/cifar10/test.txt -valSize -1 -valEvery 1 -batchSize 160 -epochSize -1 -pretrainEpochs 10 -epochs 100 -learningRate 0.01 -LRDropEvery 30 -LRDropFactor 5
th Forward.lua "out/output.t7 CifarDeepTreeProcessor.lua [-forwardWeighted]" /data/cifar10/test.txt -batchSize 128

Use WRN-22-8

th Train.lua "models/DeepTree.lua -resnet -channels 128 -preconv 1 -depth 3 -nodedepth 3 -pool 3,4 CifarDeepTreeProcessor.lua -flip 0.5 -minCropPercent 0.8" /mnt/sdb1/jshen/cifar10/trainval.txt out/output.t7 -val /mnt/sdb1/jshen/cifar10/test.txt -valSize -1 -valEvery 1 -batchSize 160 -epochSize -1 -epochs 200 -learningRate 0.001 -LRDropEvery 60 -LRDropFactor 5
th Forward.lua "out/output.t7 CifarDeepTreeProcessor.lua [-forwardWeighted]" /data/cifar10/test.txt -batchSize 128