Writing Python (Lasagne + Theano library) code for respresenting model uncertainty in deep learning. Based on the following:
- 2016: Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning, Dropout as a Bayesian Approximation: Appendix
Training takes about less than 3 hours (for 3000 epochs) minutes with GPU; thanks National Supercomputing Centre (NSCC) Singapore!
The main implementation is in mnist_mc_dropout.py which uses helper functions from helpers.py, and of course the dataset mnist.pkl.gz. For plotting training/validation errors see plot_error.py. All the outputs are saved/pickled in the output folder.
Run/theano settings: THEANO_FLAGS='mode=FAST_RUN, device=gpu, floatX=float32' python mnist_mc_dropout.py
