Updated MultiProcessDatasetPredictor so that it runs on one GPU (allows for async dataset processing). Also properly checks the number of GPUs if CUDA_VISIBLE_DEVICES is not defined.

* fix dp extraction bug for visualize().

* gt_masks is actually not needed here.

* simplify GAN usage

* move implementation to class

* small rename

* Use last k observations instead of last k epochs.

* rewriting in a more concise way

* Fix valid_for_nccl: nccl supports fp16 now

* Check TF version first for valid_for_nccl

update docs

update docs

* start the showcase of exporting models from tensorpack

* mobile support

* showcase loading a frozen and pruned graph

* fix travis issues

* add docs

* add example image

* classes to functions

* typo

* typos

* some more fixes

* update docs

* rename the example