How expression-FID is calculated?

[Darius-H]

In paper Sec4.1 it says: "E-FID leverages the Inception network’s features to critically evaluate the authenticity of the generated images, providing a more nuanced gauge of image fidelity. First, E-FID employs face reconstruction method, as detailed in the [2], to extract expression parameters. Then, it calculates the FID of these extracted parameters to quantitatively assess the disparity between the facial expressions present in the generated videos and those found in the GT dataset."

It confused me, as expression parameters is in shape (B,64). Did you use inception network to extract the features of the expression parameters before calculating the FID or did you calculate the FID directly from the expression parameters, can you provide the test code?