Torch Tutorial for PlantVillage Challenge.
I have trained the alexnet model with the above augmentation technique for 120 epochs. On a machine with Nvidia Titan X, each epoch took about 3/4 minutes. It took about 7 hrs to train the network. On my Macbook Air, each epoch takes this much time.
I get a F-Score of 0.99 with this network :).
I also have trained resnet-34 (ResNet is the winner of the latest ImageNet challenge) and got almost the same F-Score.
You can train these yourself using accompanying repo. Feel free to mail me if you have found any bugs.
git clone firstname.lastname@example.org:chsasank/plantvillage-challenge.git cd plantvillage-challenge # train alexnet th main.lua -save . -data path/to/train-val # evaluate alexnet th submission.lua alexnet_120.h5 path/to/test > alexnet_submission.csv # train resnet th main.lua -model resnet -depth 34 -learningRate 0.1 # evaluate alexnet th submission.lua resnet_120.h5 path/to/test > resnet_submission.csv
This article may also be seen on the author's blog