Encode only coarse dlatents

[jakeelwes]

Hey @pbaylies thanks for this repo! :)
You suggested to predict pose I could train a resnet to predict just the coarse dlatents (Puzer#15).

Would it be something to do with this reshape?

stylegan-encoder/train_resnet.py

Line 162 in b5ddcd7

x = Reshape((model_scale, 512))(x) # train against all dlatent values

Or changing the size of W?

stylegan-encoder/train_resnet.py

Lines 55 to 60 in b5ddcd7

	W = Gs.components.mapping.run(Z, None, minibatch_size=minibatch_size) # Use mapping network to get unique dlatents for more variation.
	dlatent_avg = Gs.get_var('dlatent_avg') # [component]
	W = (W[np.newaxis] - dlatent_avg) * np.reshape([truncation, -truncation], [-1, 1, 1, 1]) + dlatent_avg # truncation trick and add negative image pair
	W = np.append(W[0], W[1], axis=0)
	W = W[:, :mod_r]
	W = W.reshape((n*2, model_scale, 512))

Thanks for your help, sorry I'm not from a ML background.

[pbaylies]

Hi @jakeelwes -- for a simpler example, take a look at this file from issue #1 -- it only predicts a single 512-wide vector for everything, so that includes the pose. This is also what StyleGAN does while training.

I'm not originally from an ML background either, I've just been programming for a long time!