Name

dnn_saliency - Find the saliency parts of an image that contribute to the activation of the specified channel.

Synopsis

dnn_saliency [-h] -net Net -layer Layer -chn Channel [Channel ...]
             -stim Stimulus [-meth Method] [-mode Mode] [-cmap Colormap]
             [-vmin Vmin] [-vmax Vmax] [-show] [-out Output]

Arguments

Required Arguments

Argument	Discription
net	a neural network name
layer	specify the layerE.g., ‘conv1’ represents the first convolution layer and ‘fc1’ represents the first full connection layer.
chn	Channel numbers used to specify which channels are used to find salience images
stim	a .stim.csv file which contains stimulus information

Optional Arguments

Argument	Discription
meth	the method used to generate the saliency image.choose from (‘guided’, ‘vanilla’). Default is ‘guided’
mode	Visualization mode of the saliency image.RGB: visualize derivatives directly;gray: retain the maximal magnitude of RGB channels for each pixel, and visualize as a gray image.Note: -cmap, -vmin and -vmax options are only valid at the gray mode.choose from (‘RGB’, ‘gray’). Default is ‘RGB’.
cmap	show salience images with the specified colormapchoose from matplotlib colormaps. Default is coolwarm.
vmin	The minimal value used in colormap is applied in all salience images.Default is the minimal value of each salience image for itself.
vmax	The maximal value used in colormap is applied in all salience images.Default is the maximal value of each salience image for itself.
show	If used, display stimuli and salience images in figures.
out	an output directory where the figures are saved

Outputs

Display or save out the figures which contain stimuli and their salience images corresponding each channel.

Examples

Display examples’ salience images corresponding to the 294th and 23rd channels in layer ‘fc3’.

dnn_saliency -net AlexNet -layer fc3 -chn 294 23 -stim examples.stim.csv -show

Save out to the current directory for examples’ salience images corresponding to the 294th and 23rd channels in layer ‘fc3’. Using gray mode and gray colormap.

dnn_saliency -net AlexNet -layer fc3 -chn 294 23 -stim examples.stim.csv -mode gray -cmap gray -out .