Andrew Hundt

cool looks nice! I think your generator mechanism is much better than what I had been doing. I also saw your contribution to Keras-FCN which was great! Oh did you...

no data augmentation at the moment I assume? Sometimes it isn't necessary if the dataset is sufficiently large anyway, perhaps coco is big enough. With this update have you been...

Good points! I also have an update: François Chollet, Keras' author, said he is interested in directly incorporating dense prediction/FCN into the Keras API, so I'm seeking suggestions/feedback at https://github.com/fchollet/keras/issues/6538

I looked through your updates and it is looking great! It would be cool to create a pull request to incorporate some of your tools that are already set up...

>> Here is my idea for the generate_samples_from_disk API: > LGTM (I always wanted to say that 😛). Cool I'll post that API proposal on the keras dense prediction issue....

>> What kind of image rate are you getting now? 5-10 per second or something? > Yes, something like that, about a second at 4x512x512xdim or slightly faster at 8x256x256xdim....

Sadly so far my mscoco results in Keras-FCN were no good. The loss function went negative with the full image segmentation and one_hot encoding, then I ran out of time...

There has been a bugfix in densenet that solved some problems so it might work better now! https://github.com/farizrahman4u/keras-contrib/blob/master/keras_contrib/applications/densenet.py

@jmtatsch Sorry my post is totally irrelevant I must have mixed up tabs on my browser or something.

Keras spends a lot of time with an empty gpu. There are collectively quite a few reasons, some of which are discussed in https://github.com/fchollet/keras/issues/6928. Putting things into a tfrecord, using...