Support for mixed precision training

Add support for mixed precision training via nvidia's apex library.

Bugfixes

Merge pull request #192 from inferno-pytorch/fix-setup-requirements

fix setup.py requirements

Remove variable

Remove deprecated pytorch constructs like Variable and volatile.
This breaks compatibility with pytorch < 0.4.

Also fixes some bugs:

• Remove deprecated scipy.imsave
• Fix import in setup.py

Some bugfixes

Fixes the following issues:

• Invalid default values in TensorboardLogger
• Delayed Keyboard interrupt not in main thread
• Use of deprecated scipy.misc.imsave

Additional functionality and bugfixes

New functionality:

• gradient logging callback
• minimal logging callback
• flexible number of channels in U-Net
• more noise transforms

Bugfixes:

• Multi-channel hdf5 loader
• Sum layer

Support networkx 2

Adds support for networkx >= 2

Support for pytorch 1.0

New release, supports pytorch >=0.3;<=1.0.
This will be the last release supporting pytorch 0.3.
Additional changes:

• fix unittest and activate unittests in travis
• fix support for tensorboardx 1.14

and more small additions

v0.2.0: Merge pull request #136 from inferno-pytorch/super-dev

New release includes:

• support for pytorch 0.4.1 and legacy support up to 0.3
• models with u-net implementation
• tensorboard integration supports histograms
• package on conda-forge
• many bugfixes

First tentative pre-release

v0.1.7

Update setup.cfg