nick_feature_vis.py

import scipy.ndimage as nd
from scipy.ndimage import gaussian_filter
import lucid.optvis.transform as transform

def mean_alpha():
    def inner(T):
        input_t = T("input")
        return 1.0 - tf.reduce_mean(input_t[..., 3:])

    return objectives.Objective(inner)

def pretty_alpha(img, padding = 2):
  rgb = img[..., :3]
  white = np.ones(rgb.shape)
  alpha = img[..., 3:]

  percent_bg = 0.5
  bg = percent_bg * rgb + (1 - percent_bg) * white
  beautiful = (gaussian_filter(bg, sigma=[5, 5, 1])) * (1 - alpha) + rgb * alpha
  center = beautiful
  cropped = nd.zoom(
      center[padding:-padding, padding:-padding, :], order=2, zoom=[1.2, 1.2, 1]
  )
  return cropped

def render_neuron(model_name, layer, channel):  
  model = models.InceptionV1()
  
  w = 90
  param_f = lambda: param.image(w, alpha = True)
  
  obj = objectives.neuron(layer, channel)

  obj -= 1e2 * objectives.blur_alpha_each_step()
  obj *= mean_alpha()

  transforms = []
  transforms.append(transform.jitter(4))
  transforms.append(transform.jitter(4))
  transforms.append(transform.jitter(8))
  transforms.append(transform.jitter(8))
  transforms.append(transform.jitter(8))
  transforms.append(transform.random_scale([0.995**n for n in range(-5,80)] + [0.998**n for n in 2*list(range(20,40))]))
  transforms.append(transform.random_rotate(list(range(-20,20))+list(range(-10,10))+list(range(-5,5))+5*[0]))
  transforms.append(transform.jitter(2))
  transforms.append(transform.crop_or_pad_to(w, w))
  transforms.append(transform.collapse_alpha_random())

  result = render.render_vis(model, obj, param_f=param_f, transforms=transforms, thresholds=[1024])[-1]

  result = pretty_alpha(result[0])
  show(result)
  
  return result

Thanks @ncammarata! Everything works great. The only modification I needed to make to run as is was to load the graph (likely a typo in your gist!): model.load_graphdef()