I'm visualizing some samples generated by your approach through the edit.py script, and I have a question. My understanding is that this script generates samples by:
Starting from a latent code, call it z, a direction in that space, refer to it by a unit vector n, starting and finishing magnitudes, call them s and f, respectively, and
Generating a sequence of latent codes by making a linear interpolation from vector v1 = z - sn to vector v2 = z + fn.
I understand this is done in the function linear_interpolatehere, that is used here, and whose product is stored in the variable interpolations. From this understanding of the code, I would expect that all the codes saved in the interpolations variable are at an L2 distance of, at most, max(s, f); and, in particular, there should be an interpolation for which the distance iss, and another interpolation for which the distance isf. However, when I check this in the code by running
np.linalg.norm(interpolations - latent_codes[sample_id:sample_id + 1], axis=1) , I get other results.
I think this is either because I'm misunderstanding something, or because there is a (small) bug in the code. Such bug would probably be inconsequential, but I thought I should report it. I think the bug itself has to do with this line in particular, in the linear_interpolate function. Specifically, I'm unable to understand why the computation latent_code.dot(boundary.T) is performed. The boundary variable is a direction, rather than an actual boundary, right? As there is no presence of the bias term to determine the actual side of the hyperplane on which latent_code is falling. Further, I see that no such analogous computation is performed for the case of latent codes in the W+ space (see here).
I think that particular line is the root of the problem I observe. This is because, if I simply comment that line and run the code, the results are as expected.
Hi there,
Thanks for the great work!
I'm visualizing some samples generated by your approach through the edit.py script, and I have a question. My understanding is that this script generates samples by:
I understand this is done in the function
linear_interpolatehere, that is used here, and whose product is stored in the variableinterpolations. From this understanding of the code, I would expect that all the codes saved in theinterpolationsvariable are at an L2 distance of, at most, max(s, f); and, in particular, there should be an interpolation for which the distance is s, and another interpolation for which the distance is f. However, when I check this in the code by runningnp.linalg.norm(interpolations - latent_codes[sample_id:sample_id + 1], axis=1), I get other results.I think this is either because I'm misunderstanding something, or because there is a (small) bug in the code. Such bug would probably be inconsequential, but I thought I should report it. I think the bug itself has to do with this line in particular, in the
linear_interpolatefunction. Specifically, I'm unable to understand why the computationlatent_code.dot(boundary.T)is performed. Theboundaryvariable is a direction, rather than an actual boundary, right? As there is no presence of the bias term to determine the actual side of the hyperplane on whichlatent_codeis falling. Further, I see that no such analogous computation is performed for the case of latent codes in the W+ space (see here).I think that particular line is the root of the problem I observe. This is because, if I simply comment that line and run the code, the results are as expected.
Could you please take a look into my claims?
Thank you!