Training

• Download the data from www.cs.toronto.edu/~nitish/data/mnist.h5 [390Mb]
• Set the data_dir in all *_data.pbtxt so that it points to the directory where the data was downloaded.
• Set the checkpoint directory in net.pbtxt. This is where the model, error stats, logs etc will be written. Make sure this diretory has been created.

Run:

$ train_convnet --board=<board-id> --model=net.pbtxt --train=train_data.pbtxt --val=val_data.pbtxt

or

$ train_convnet --board=<board-id> --model=net.pbtxt --train=train_plus_val_data.pbtxt --val=test_data.pbtxt

Toronto users- Make sure the board is locked before running this.

This will train the net for 100,000 updates and write the model out in the checkpoint directory. This takes 4-5 min (on NVIDIA Titan, may vary depending on GPU). Training on train_plus_val_data.pbtxt should get around 1.2% test error by then.

Extracting features

The representation at different layers of the learned model can be extracted:

$ extract_representation --board=<board-id> --model=<model-file> --feature-config=<feature-config-file>

For example,

$ extract_representation --board=0 --model=checkpoint_dir/mnist_net_20140627130044.pbtxt --feature-config=feature_config.pbtxt

The feature_config.pbtxt file describes what data to input, which layers to extract features from and where to write the output.

Looking at the performance

The *.log files in the checkpoint directory contain the performance metrics for the training and validation sets. Run

$ python ../../apps/show_plots.py ./checkpoint_dir

Draw the net.

Draw the network graph.

$ python ../../apps/pbtxt2dot.py net.pbtxt net.dot
$ dot -Tpng net.dot -o net.png