Computation of FID at different scales - MSPIE about mmgeneration

[nbei]

Hi @vglsd , thanks for your interest in our work and MMGEN.

I suggest that you may visualize the generated samples first. Since FID=19 indicates a poor generation quality, you can check whether the output samples indeed show any artifacts.

[entavelis]

Thanks for the prompt reply!

I realized the FID gets worse when I use truncation on the noise input. Did you use any truncation in the results you report?

Moreover, just to make sure I am not making any mistake on this:
python tools/utils/inception_stat.py --imgsdir ${path-to-1024-FFHQ} --pklname ${out-pkl-destination} --size ${size-of-output-images-to-compare-with} --num-samples 50000
Is this the right command to get the inception pkl?

[nbei]

There are some mistakes in your command:

• Firstly, you should explicitly resize the 1024-FFHQ dataset to 256-FFHQ dataset. Then, the --imgsdir should be set with the path for the 256-FFHQ.
• --size will set the size of input images to the inception network. In general, the input shape of the pre-trained inception network should be (299, 299), which is our default setting in tools/utils/inception_stat.py

We have already provided the 256-FFHQ inception stat, which you can download from here.

We did NOT use any truncation tricks in calculating quantitative metrics. In GAN literature, we merely use the truncation trick to perform visualization.

[entavelis]

Thanks for the help!