Benchmarking on MNIST:
The following mentioned model definition files are under the folder: models/mnist/ .
DLBENCH_ROOT="path to the root directory of this benchmark"
TensorFlow uses a variant of LeNet and its network structure is shown as follows:
Run TensorFlow with its default MNIST setting:
cd $DLBENCH_ROOT/models/mnist/tensorflow/ python mnist_deep.py
The Training Time, Testing Time and Accuracy will appear after completion.
Similarly, the NN network structure of Caffe is shown as follows:
It is also a variant of LeNet.
Run Caffe with its default MNIST setting:
cd $DLBENCH_ROOT/models/mnist/caffe caffe train -solver lenet_solver.prototxt > log.txt 2>&1
The Training Time, Testing Time and Accuracy can be extracted from the log.txt file.
Run on CPU:
cd $DLBENCH_ROOT/models/mnist/theano THEANO_FLAGS=device=cpu python convolutional_mlp.py