Name

dnn_rsa - Characterise DNN activation’s representation dissimilarity matrix (RDM).

Synopsis

dnn_rsa [-h] -act Activation [-layer Layer [Layer ...]]
        [-chn Channel [Channel ...]] [-dmask DnnMask] [-iteraxis Axis]
        [-cate Category] [-metric Metric] [-zscore] -out Output

Arguments

Required Arguments

Argument	Discription
act	DNN activation file
out	output filename with suffix as .rdm.h5

Optional Arguments

Argument	Discription
layer	layer names of interest
chn	channel numbers of interest Default using all channels of each layer specified by -layer.
dmask	a .dmask.csv file in which layers of interest are listed with their own channels, rows and columns of interest.
iteraxis	choices=(channel, row_col) Iterate along the specified axis. channel: Do rsa on each channel. row_col: Do rsa on each location (row_idx, col_idx). default: Do rsa on the whole layer.
cate	a .stim.csv file which contains category information (i.e. ‘label’ item) If used, do rsa category-wisely that average activation pattern before calculating the distance. And the row/column order of RDM is organized from small to big according to the ‘label’.
metric	Specify metric used to calculate distance. Default: euclidean
zscore	Standardize feature values for each sample by using zscore.

Outputs

The output is a .rdm.h5 file, which contains each layer’s RDM.

Examples

Calculate euclidean distance for each pair of stimuli using the activation pattern of each layer in test.act.h5.
Save results to out.rdm.h5

dnn_rsa -act test.act.h5 -out out.rdm.h5