Home

Awesome

##Training an Object Classifier in Torch-7 on multiple GPUs over ImageNet

In this concise example (1200 lines including a general-purpose and highly scalable data loader for images), we showcase:

Requirements

Data processing

The images dont need to be preprocessed or packaged in any database. It is preferred to keep the dataset on an SSD but we have used the data loader comfortably over NFS without loss in speed. We just use a simple convention: SubFolderName == ClassName. So, for example: if you have classes {cat,dog}, cat images go into the folder dataset/cat and dog images go into dataset/dog

The training images for imagenet are already in appropriate subfolders (like n07579787, n07880968). You need to get the validation groundtruth and move the validation images into appropriate subfolders. To do this, download ILSVRC2012_img_train.tar ILSVRC2012_img_val.tar and use the following commands:

# extract train data
mkdir train && mv ILSVRC2012_img_train.tar train/ && cd train
tar -xvf ILSVRC2012_img_train.tar && rm -f ILSVRC2012_img_train.tar
find . -name "*.tar" | while read NAME ; do mkdir -p "${NAME%.tar}"; tar -xvf "${NAME}" -C "${NAME%.tar}"; rm -f "${NAME}"; done
# extract validation data
cd ../ && mkdir val && mv ILSVRC2012_img_val.tar val/ && cd val && tar -xvf ILSVRC2012_img_val.tar
wget -qO- https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh | bash

Now you are all set!

If your imagenet dataset is on HDD or a slow SSD, run this command to resize all the images such that the smaller dimension is 256 and the aspect ratio is intact. This helps with loading the data from disk faster.

find . -name "*.JPEG" | xargs -I {} convert {} -resize "256^>" {}

Running

The training scripts come with several options which can be listed by running the script with the flag --help

th main.lua --help

To run the training, simply run main.lua By default, the script runs 1-GPU AlexNet with the CuDNN backend and 2 data-loader threads.

th main.lua -data [imagenet-folder with train and val folders]

For 2-GPU model parallel AlexNet + CuDNN, you can run it this way:

th main.lua -data [imagenet-folder with train and val folders] -nGPU 2 -backend cudnn -netType alexnet

Similarly, you can switch the backends to 'cunn' to use a different set of CUDA kernels.

You can also alternatively train OverFeat using this following command:

th main.lua -data [imagenet-folder with train and val folders] -netType overfeat

# multi-GPU overfeat (let's say 2-GPU)
th main.lua -data [imagenet-folder with train and val folders] -netType overfeat -nGPU 2

The training script prints the current Top-1 and Top-5 error as well as the objective loss at every mini-batch. We hard-coded a learning rate schedule so that AlexNet converges to an error of 42.5% at the end of 53 epochs.

At the end of every epoch, the model is saved to disk (as model_[xx].t7 where xx is the epoch number). You can reload this model into torch at any time using torch.load

model = torch.load('model_10.t7') -- loading back a saved model

Similarly, if you would like to test your model on a new image, you can use testHook from line 103 in donkey.lua to load your image, and send it through the model for predictions. For example:

dofile('donkey.lua')
img = testHook({loadSize}, 'test.jpg')
model = torch.load('model_10.t7')
if img:dim() == 3 then
  img = img:view(1, img:size(1), img:size(2), img:size(3))
end
predictions = model:forward(img:cuda())

If you ever want to reuse this example, and debug your scripts, it is suggested to debug and develop in the single-threaded mode, so that stack traces are printed fully.

th main.lua -nDonkeys 0 [...options...]

Code Description