Alex Krizhevsky's cuda-convnet details the model definitions, parameters, and training procedure for good performance on CIFAR-10. This example reproduces his results in Caffe.

We will assume that you have Caffe successfully compiled. If not, please refer to the Installation page. In this tutorial, we will assume that your caffe installation is located at CAFFE_ROOT.

We thank @chyojn for the pull request that defined the model schemas and solver configurations.

This example is a work-in-progress. It would be nice to further explain details of the network and training choices and benchmark the full training.

You will first need to download and convert the data format from the CIFAR-10 website. To do this, simply run the following commands:

cd $CAFFE_ROOT
./data/cifar10/get_cifar10.sh
./examples/cifar10/create_cifar10.sh

If it complains that wget or gunzip are not installed, you need to install them respectively. After running the script there should be the dataset, ./cifar10-leveldb, and the data set image mean ./mean.binaryproto.

The CIFAR-10 model is a CNN that composes layers of convolution, pooling, rectified linear unit (ReLU) nonlinearities, and local contrast normalization with a linear classifier on top of it all. We have defined the model in the CAFFE_ROOT/examples/cifar10 directory's cifar10_quick_train_test.prototxt.

Training the model is simple after you have written the network definition protobuf and solver protobuf files (refer to MNIST Tutorial). Simply run train_quick.sh, or the following command directly:

cd $CAFFE_ROOT
./examples/cifar10/train_quick.sh

train_quick.sh is a simple script, so have a look inside. The main tool for training is caffe with the train action, and the solver protobuf text file as its argument.

When you run the code, you will see a lot of messages flying by like this:

I0317 21:52:48.945710 2008298256 net.cpp:74] Creating Layer conv1
I0317 21:52:48.945716 2008298256 net.cpp:84] conv1 <- data
I0317 21:52:48.945725 2008298256 net.cpp:110] conv1 -> conv1
I0317 21:52:49.298691 2008298256 net.cpp:125] Top shape: 100 32 32 32 (3276800)
I0317 21:52:49.298719 2008298256 net.cpp:151] conv1 needs backward computation.

These messages tell you the details about each layer, its connections and its output shape, which may be helpful in debugging. After the initialization, the training will start:

I0317 21:52:49.309370 2008298256 net.cpp:166] Network initialization done.
I0317 21:52:49.309376 2008298256 net.cpp:167] Memory required for Data 23790808
I0317 21:52:49.309422 2008298256 solver.cpp:36] Solver scaffolding done.
I0317 21:52:49.309447 2008298256 solver.cpp:47] Solving CIFAR10_quick_train

Based on the solver setting, we will print the training loss function every 100 iterations, and test the network every 500 iterations. You will see messages like this:

I0317 21:53:12.179772 2008298256 solver.cpp:208] Iteration 100, lr = 0.001
I0317 21:53:12.185698 2008298256 solver.cpp:65] Iteration 100, loss = 1.73643
...
I0317 21:54:41.150030 2008298256 solver.cpp:87] Iteration 500, Testing net
I0317 21:54:47.129461 2008298256 solver.cpp:114] Test score #0: 0.5504
I0317 21:54:47.129500 2008298256 solver.cpp:114] Test score #1: 1.27805

For each training iteration, lr is the learning rate of that iteration, and loss is the training function. For the output of the testing phase, score 0 is the accuracy, and score 1 is the testing loss function.

And after making yourself a cup of coffee, you are done!

I0317 22:12:19.666914 2008298256 solver.cpp:87] Iteration 5000, Testing net
I0317 22:12:25.580330 2008298256 solver.cpp:114] Test score #0: 0.7533
I0317 22:12:25.580379 2008298256 solver.cpp:114] Test score #1: 0.739837
I0317 22:12:25.587262 2008298256 solver.cpp:130] Snapshotting to cifar10_quick_iter_5000
I0317 22:12:25.590215 2008298256 solver.cpp:137] Snapshotting solver state to cifar10_quick_iter_5000.solverstate
I0317 22:12:25.592813 2008298256 solver.cpp:81] Optimization Done.

Our model achieved ~75% test accuracy. The model parameters are stored in binary protobuf format in

cifar10_quick_iter_5000

which is ready-to-deploy in CPU or GPU mode! Refer to the CAFFE_ROOT/examples/cifar10/cifar10_quick.prototxt for the deployment model definition that can be called on new data.

CIFAR-10, while still small, has enough data to make GPU training attractive.

To compare CPU vs. GPU training speed, simply change one line in all the cifar*solver.prototxt:

# solver mode: CPU or GPU
solver_mode: CPU

and you will be using CPU for training.