Questions on GANSynth training

Hi, thanks for the great work. I have some questions about GANSynth training from the paper. I figured out some of them from the codes but couldn’t find everything. (cc’ @jesseengel @chrisdonahue)

• Q 1-1. From Section D, On concatenating (x, minibatch std), I guess it means - if the x= the tensor size of (batch_size, 2, 16, 256), the since the std channel size is (None, 2, 16, 1), the std is computed for each pixel & over channel & over mini batch items? I.e., x.std(axis=0, keepdims=True).std(axis=3, keepdims=True)?
• Q 1-2. Could you tell me why did we need this?

I’m also confused by (Section D)

measure the maximum range over 100 examples and independently shift and scale the log-magnitudes and phases to [-0.8, 0.8]

, especially by ‘measure… over 100 examples’ vs ‘independently’. Also, unlike section D in the paper, it says in Section 2.2 that

…then scale the magnitudes to be between -1 and 1 to… …the phase angle is also scaled to between -1 and 1 … as phase models.

Also later,

…unwrap the phase angle and take the finite difference … IF

To summarize,

• Q 2-1. What is happening to log magnitude spectrograms? is it [-0.8, 0.8] or [-1, 1] (independently? how are the information from 100 examples are used?)
• Q 2-2. What is happening to the phases in Phase model?
• Q 2-3. Same for IF model?

I would very appreciate some concrete explanation on it. Thanks!

[jesseengel]

Also, fwiw, I followed np.unwrap closely when writing the tensorflow version.

*From: *Chris Donahue [e-mail] *Date: *Mon, May 13, 2019 at 8:46 PM *To: *tensorflow/magenta *Cc: *Jesse Engel, Mention

Ah right. It depends on what you set as the maximum discontinuity values

for np.unwrap. But with the default configuration of np.pi (which we also use), you are correct that the range should be [-pi, pi]. If you use a larger discontinuity > np.pi, the range will increase accordingly.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/tensorflow/magenta/issues/1524?email_source=notifications&email_token=AANFCCPUX6HDQPSUYZKB7EDPVIYZ3A5CNFSM4HLMU6YKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODVKFX5Q#issuecomment-492067830, or mute the thread https://github.com/notifications/unsubscribe-auth/AANFCCIJULXZQNZ4V7ELMJDPVIYZ3ANCNFSM4HLMU6YA .

[chrisdonahue]

Ah right. It depends on what you set as the maximum discontinuity values for np.unwrap. But with the default configuration of np.pi (which we also use), you are correct that the range should be [-pi, pi]. If you use a larger discontinuity > np.pi, the range will increase accordingly.