Neon comes with the nvis tool for easy visualization.

nvis reads hdf5 data written out by neon callbacks and generates Bokeh visualizations embedded in html.

Note that Bokeh and nvis are not installed by default (due to the size of the Bokeh package). To enable these be sure to first install via make -e VIS=true (or edit the Makefile to set the VIS variable). If you’ve already installed neon without enabling visualization dependencies you’ll need to touch vis_requirements.txt prior to the make -e VIS=true call to ensure virtualenv python dependencies get triggered.

Cost visualization

Visualization of training and validation set cost on an epoch or minibatch axis is currently supported. The following example saves training cost data from neon and creates a visualization with nvis:

neon examples/mnist_mlp.yaml -o data.h5
nvis -i data.h5 -o .

If validation cost is computed in neon, it will also be saved and visualized by nvis:

neon examples/mnist_mlp.yaml -o data.h5 --eval_freq 1
nvis -i data.h5 -o .

Here’s an example of the visualized output:


Script examples not directly using the neon command can still dump visualization data:

./examples/ -o data.h5
nvis -i data.h5 -o .

Layer deconvolution visualization

Guided Backprop based visualization of convolutional layer activations is also supported via the DeconvCallback class or add_deconv_callback() function. Note that unlike in the paper, :py:meth:`add_deconv_callback() does not support networks that utilize max pooling.

For example, one can provide the –deconv argument to examples/imagenet/ to ensure that the callback gets added:

examples/imagenet/ -o data.h5 --deconv
nvis -i data.h5 -o .

Here’s an example of the output.


Layer histogram visualization

Per layer value histograms are also available via the HistCallback class or add_hist_callback() function. To use this, create Callbacks and call add_hist_callback(). For example, edit the examples/ script by adding the following line after callbacks is initialized:


Then, at each minibatch or epoch, the model weights for every minibatch. We run the model then generate the visualization:

./ -o data.h5 -e 1
nvis -i data.h5 -o .

Here’s an example output for the first two convolution layers:


By default, the callback only saves the Tensor W from every layer that has that attribute (e.g. all layers that inherit from ParameterLayer). To build a histogram for your specific attribute, pass a list to filter_key.

callbacks.add_hist_callback(plot_per_mini=True, filter_key = ['W','dW','myParameter'])

Then, all layers that have self.W, self.dW, or self.myParameter will have those weights saved for visualization in the specified output file.