Experimental PyTorch Generator Distributions

PyTorch Generator Test

Outline: The point of this test was to check whether the PyTorch and Tensorflow generator produce the same images for the same seed. Moreover, I want to see the impact of generating new images on the distribution of p_hat_cnn.

To the first point, the two models produce the same images for the same seed ! I plot the p_hat_cnn distribution of the PyTorch StyleGAN2 port for which I have transferred weights. Here are a few important points:

• I am taking a specific seed (6000 - 8000) in the GANs latent space and generate with \(\psi= 0.5\) which truncates the latent distribution and produces what the GAN has learned as the “average” face
• I am also generating the same seeds for \(\psi= 0.8\) (note that for \(\psi< 0.5\) it produces the same results as \(\psi= 0.5\) as this is a lower bound)
• I am comparing the density of p_hat_cnn from our entire validation set to that of the 2,000 generated images for varying \(\psi\) . Here we can see the difference between the full distribution of real mugshots in red compared to what the GAN generates as made up mugshots in blue and green.

The main takeaway for me is that we may need to be careful how to set \(\psi\) when we generate images so as to not skew our editing. We could potentially be missing important features depending on how far we move from the baseline distribution.