search

openai / guided-diffusion

MIT License
6.11k stars 813 forks source link

mixed_6/conv:0 layer in pytorch structure #109

Open wtliao opened 1 year ago

wtliao commented 1 year ago

Thanks for sharing the nice code. I want to use pytorch code for calculating the metrics. However, I dont know which layer in pytorch InceptionV3 is the one corresponding to the mixed_6/conv:0 layer in the tensorflow structure. Anyone has the idea? Thanks!

wtliao commented 1 year ago

I find the corresponding layer in the pytorch InceptionV3, it is the first 192 channels of mixed_6d.

But however, I found that, whatever the size of input image, the spatial_pred features are always (batch_size, 17, 17, 7). It seems that any input image is resized into 299x299? I am confused that, whether this resize operation can reflect the quality of generated image not in 299x299 originally?