Slerp through the BigGAN latent space

biggan_slerp

# to be used in conjunction with the functions defined here: 
#  https://colab.research.google.com/github/tensorflow/hub/blob/master/examples/colab/biggan_generation_with_tf_hub.ipynb

# party parrot transformation
noise_seed_A = 3 # right facing
noise_seed_B = 31 # left facing

num_interps = 14
truncation = 0.2
category = 14

def slerp(A, B, num_interps):
  # see https://en.wikipedia.org/wiki/Slerp
  alphas = np.linspace(-1.5, 2.5, num_interps) # each unit step tends to be a 90 degree rotation in high-D space, so this is ~360 degrees
  omega = np.zeros((A.shape[0],1))
  for i in range(A.shape[0]):
      tmp = np.dot(A[i],B[i])/(np.linalg.norm(A[i])*np.linalg.norm(B[i]))
      omega[i] = np.arccos(np.clip(tmp,0.0,1.0))+1e-9
  return np.array([(np.sin((1-a)*omega)/np.sin(omega))*A + (np.sin(a*omega)/np.sin(omega))*B for a in alphas])

def slerp_and_shape(A, B, num_interps):
  interps = slerp(A, B, num_interps)
  return (interps.transpose(1, 0, *range(2, len(interps.shape)))
                 .reshape(num_interps, *interps.shape[2:]))

z_A, z_B = [truncated_z_sample(num_samples, truncation, noise_seed)
            for noise_seed in [noise_seed_A, noise_seed_B]]
z_slerp = slerp_and_shape(z_A, z_B, num_interps)

ims_slerp = sample(sess, z_slerp, one_hot([category] * num_interps), truncation=truncation)
imshow(imgrid(ims_slerp, cols=num_interps))

# output video
import cv2
video_name = 'video.mov'
_, height, width, _ = ims_slerp.shape
video = cv2.VideoWriter(video_name, cv2.VideoWriter_fourcc('M','J','P','G'), fps=24, frameSize=(width,height))
for iter in range(0,ims_slerp.shape[0]):
    video.write(ims_slerp[iter,:,:,::-1])
cv2.destroyAllWindows()
video.release()