CIFAR-100

Benchmarking on CIFAR-100:

The following mentioned model definition files are under the folder: models/cifar100/ .

Pre-setting:

DLBENCH_ROOT="path to the root directory of this benchmark"

TensorFlow:

Run TensorFlow with its default MNIST setting:

cd $DLBENCH_ROOT/models/cifar100/tensorflow/
python cifar100_train.py > train_log.txt 2>&1

After the completion of training, run the following command to test the tranined model:

python cifar100_eval.py > test_log.txt 2>&1

The Accuracy will appear after completion of cifar100_eval.py. And the Training Time and Testing Time can be extracted from the train_log.txt and test_log.txt.

Caffe:

Similarly, the NN network structure of Caffe is shown as follows:

Run Caffe with its default setting:

cd $DLBENCH_ROOT/models/cifar100/caffe
./train_quick.sh > log.txt 2>&1

The Training Time, Testing Time and Accuracy can be extracted from the log.txt file.

Torch:

cd $DLBENCH_ROOT/models/cifar100/torch

Run on CPU:

th train-on-cifar-100.lua

Theano:

Note: The implementation for Theano on CIFAR-100 derived from Reslab Theano tutorial (10 February 2015)

cd $DLBENCH_ROOT/models/cifar100/theano
THEANO_FLAGS=device=cpu python convolutional_mlp.py