My environment could be contributing to the trouble I’m running into, so let me be clear: running Ubuntu 16.04, CUDA 8.0, cuDNN 5.11, Anaconda 2 (4.4.0), and relying on Keras 2.0.2 + Theano 0.9.0, and doing all of this atop an intel based machine with a GTX 1080 TI attached. I’ve been able to make modifications to the original utils.py and vgg16.py files to make them work with Keras 2.0.2, but I consistently run into a segmentation fault when the vgg16.py calls the following:
model.add(Conv2D(filters, (3, 3), activation='relu'))
The code I’m using is only lightly adjusted for Keras 2.0.2 (mostly based on feedback others have provided here). I’ve posted all the code I’m using here: https://gist.github.com/pavgup/e6e22eb6b6c3261837881b1cf8276e96
And just in case, I also ran this in gdb to grab some additional information on the fault which I dropped into a comment on that gist.
Help! I’d love to make this work correctly, but I’m unsure what the segmentation fault really means…