Variables name and scope
I hope you're doing fine in California ! I'm now working with your bob.learn.tensorflow package (which is great btw, it really simplifies my life) and I have some questions regarding variable names and scopes ...
Since I'm implementing GANs, I have several networks and hence I have to modify the way to "identify" variables (since the name and the scope is now somehow hardcoded, https://gitlab.idiap.ch/bob/bob.learn.tensorflow/blob/master/bob/learn/tensorflow/layers/FullyConnected.py#L79).
I can easily work around that by giving a proper name to each layer (in both graph) to be sure that there won't be any "collisions" on variable names, but I don't find it that elegant. Also, in what you did, the scope and the name of each variable are the same, and I find it a bit strange (but I'm pretty sure there's a reason for this, am I right ?)
Anyway, I was planning to implement something in the line of this:
- having a scope containing the network name and the layer name
- having the variable name automatically append to this
As a result, I would ideally have, for instance, a variable like this: discriminator/conv2d_1/weights (instead of something like W_d_conv2d_1/W_d_conv2d_1).
Now the problem is that I did not experiment with model saving / loading (and summaries) yet, so I was wondering if it would be ok ... As far as I understand, you made it so because of the bug you mentioned here https://gitlab.idiap.ch/bob/bob.learn.tensorflow/blob/master/bob/learn/tensorflow/layers/Layer.py#L124 (which is now a closed issue).
Do you have any suggestions / comments on this ?
Thanks a lot !