Stats of SPNs

[yannickl96]

Hello,

does libspn-keras provide a function to directly obtain structural stats of SPNs (i.e. number of addition, multiplication and leaf nodes)? I only found the .summary() method which provides me the output shapes of each layer and the number of parameters in the examples. Any information would be appreciated!

Best, Yannick

[jostosh]

Hi Yannick,

There is no direct method for showing these stats. But it's not that complicated to do with a simple script. Assuming that you have an spn as an instance of tf.keras.models.Sequential or libspn_keras.models.SequentialSumProductNetwork you can use the following short script to print the number of sums, products and leaves:

import numpy as np

num_sums = 0
num_products = 0
num_leaves = 0

for layer in spn.layers:
  if isinstance(layer, spnk.layers.DenseSum):  # all sum-layers inherit from spnk.DenseSum
    if isinstance(layer, spnk.layers.RootSum):
      num_sums += 1
    else:
      num_sums += np.prod(layer.output_shape[1:])
  if isinstance(layer, spnk.layers.BaseLeaf):
    num_leaves += np.prod(layer.output_shape[1:])
  if isinstance(layer, (spnk.layers.DenseProduct, spnk.layers.ReduceProduct, spnk.layers.TemporalDenseProduct, spnk.layers.Conv2DProduct)):
    num_products += np.prod(layer.output_shape[1:])

print("num sums", num_sums)
print("num products", num_products)
print("num leaves", num_leaves)

[yannickl96]

Hi Jos,

works like a charm! Thank you!