Contribute to tflearn about tensorflow-fcn HOT 6 CLOSED

@MarvinTeichmann that indeed looks really great! It would definitely be useful for our users. Do not hesitate to send us a pull request :) Just my suggestion is to move the 'semantic segmentation loss' to 'objectives.py' file instead of 'losses.py' file, and change 'logits'/'labels' arguments name to 'y_pred'/'y_true', for consistency with other objective functions.

@FreakTheMighty You can go ahead, thanks!

from tensorflow-fcn.

Thanks for the hint and good idea. I have implemented the upscore_layer and the semantic segmentation loss to tflearn. Can you have a short look at it? I have no experience with tflearn. Needed to do some changed to fit in to the framework but have cannot test whether it works now.

Thanks in advance! If you approve my code, I will open a pull request.

from tensorflow-fcn.

@MarvinTeichmann wow, that was fast :D. I can take a closer look tonight, but I think it makes sense to open a pull request sooner rather than later. The maintainer, @aymericdamien, of tflearn will be able to give the most substantive feedback.

I only get a few hours a day to work on these things, but I'd like to contribute an example that recreates the FCN using tflearn and your new layers and loss.

from tensorflow-fcn.

@aymericdamien I've started working on porting @MarvinTeichmann networks here. MarvinTeichmann/tflearn#1 I'm sure I will get stuck along the way, but I think it will be good experience.

One thing I noticed about the weak_cross_entropy_2d_loss objective is that it takes more than two arguments where tflearn objectives are passed as strings into the solver. How would that interface work in tflearn?

from tensorflow-fcn.

Yes, because of the extra argument, it cannot directly be called as a string argument in regression layer, however, loss argument can be a function, so it could be something like:

def my_loss(y_pred, y_true):
  return tflearn.weak_cross_entropy_2d_loss(y_pred, y_true, num_classes=10)
net = ...
net = tflearn.regression(net, loss=my_loss)

Another approach would be to remove num_classes argument and calculate it automatically by checking dimension of y_true. So passing just

net = ...
net = tflearn.regression(net, loss='weak_cross_entropy_2d_loss')

will work

from tensorflow-fcn.

Merged in tflearn/tflearn#137.

from tensorflow-fcn.