The alpha and k values for LocalResponseNorm0 and LocalResponseNorm1 are incorrect. The code originally used to construct lucent/modelzoo/inceptionv1/InceptionV1.py has since been updated to correct this, but that change isn't reflected here.

In order to validate the correct values, one can do the following:

1. Download the original InceptionV1 protobuf file used in Lucid.

2. Open and inspect the contents:

import tensorflow as tf

graph_def = tf.compat.v1.GraphDef.FromString(open("path/to/saved_model.pb", "rb").read())
for x in graph_def.node[:]:
  if x.name not in ["localresponsenorm0", "localresponsenorm1"]:
    continue
  print(x)

and then the various attributes of the LocalResponseNorm can be examined. Example output for localresponsenorm1:

name: "localresponsenorm1"
op: "LRN"
input: "conv2d2"
device: "/cpu:0"
attr {
  key: "alpha"
  value {
    f: 9.999999747378752e-05
  }
}
attr {
  key: "beta"
  value {
    f: 0.5
  }
}
attr {
  key: "bias"
  value {
    f: 2.0
  }
}
attr {
  key: "depth_radius"
  value {
    i: 5
  }
}

LocalResponseNorm layers differ slightly in TensorFlow and in PyTorch:

• Size ends up simply being $2\times$ depth_radius
• TensorFlow simply uses $\alpha$ as is whereas PyTorch uses $\frac{\alpha}{n}$ where $n$ is the size which is just $10$ so to offset that the original $\alpha$ should be replaced with $10 \times \alpha$