uncertainty-baselines
uncertainty-baselines copied to clipboard
google
High-quality implementations of standard and SOTA methods on a variety of tasks.
> @jereliu > Currently the ensemble.py implementations in edward2/baselines compiles model graph by loading model hyper-parameters from flags, and then load model weights from checkpoints. This creates challenges for both...
> @dustinvtran > This will also let us automatically run postprocessing-style models such as deep ensembles or greedy HP ensembles. https://github.com/google/edward2/issues/327
> @dustinvtran >> @murphyk >> Their experiments seem like a great application (and are open-source). >> >> M. Walmsley et al., “Galaxy Zoo: Probabilistic Morphology through Bayesian CNNs and Active...
> @dustinvtran > @GhassenJ @JeremyNixon > > We can get it from the internal experiment management interface. We could try to establish a link from each output directory to each...
> @dustinvtran > @dusenberrymw has a setup for plotting several metrics, compared across methods, and each method of which is averaged over multiple seeds. The plotting data is obtained directly...
> @dustinvtran > I'm thinking of efficientnet or assemble-resnet. Both have TF implementations: > > * https://github.com/clovaai/assembled-cnn > * https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet/tpu (which I believe is still in the process of converting...
> @dustinvtran > @ywen666 > > We should save the time elapsed number in checkpoints. This enables pre-empted jobs to add on top of the number instead of restarting from...
> @dustinvtran > From preliminary experiments, the LR schedule from the SWA papers (https://github.com/timgaripov/swa/blob/master/train.py#L94) seems to improve the baseline results (at least for deterministic and dropout). Upgrading to that one...
> @dustinvtran > We have the exact Bash commands to launch jobs with Google's cloud cluster in order to reproduce all results. There's nothing particularly notable about them as all...
> @dustinvtran > For cifar 10.1, add nll / accuracy / ce. For cifar c, add mean nll, mce, and mean ce. similarly cifar p. > > It looks like...