Add distributed training support with dask

Also fixes #87 (closed)

assigned to @amohammadi and unassigned @tiago.pereira

the builds are failing because the tensorflow version is old.

added 2 commits

• cbb7d345 - allow changing batch norm params
• f1fa89e9 - Add a large-batch friendly resnet 50

Compare with previous version

added 6 commits

• 2b79cb39 - explicitly use nccl
• 44d5075f - Allow turning batch norm scale off
• 2e56ab7b - The whole arcface layer is 32bit in insightface pytorch code
• 9aba8dd6 - remove large batch resnet50 (was not used)
• ed1cfca0 - Change debug logging to info logging
• 452ba28d - Add iresnet models based on insightface arcface pytorch code

Compare with previous version

added 3 commits

• 082fa660 - improve intellisense
• 91613d6f - implement get_config for arcface layers
• 8f1b8d37 - Add a mirrored strategy

Compare with previous version

added 21 commits

• 8f1b8d37...f420d1b3 - 5 commits from branch master
• 26bda0a1 - [generator] remove use of deprecated arguments
• 9713d207 - multioutput models can be used instead of embedding validation models
• 0c69180a - Add distributed training support with dask
• 19eebcbb - allow changing batch norm params
• 5134f865 - Add a large-batch friendly resnet 50
• 3d12e501 - explicitly use nccl
• d6236129 - Allow turning batch norm scale off
• c7e244ad - The whole arcface layer is 32bit in insightface pytorch code
• 5fde2bb8 - remove large batch resnet50 (was not used)
• c637ed8f - Change debug logging to info logging
• 5ac8a643 - Add iresnet models based on insightface arcface pytorch code
• 6fba84cd - improve intellisense
• 2c76437d - implement get_config for arcface layers
• 48e384a5 - Add a mirrored strategy
• c7412ca2 - Fix arcface tests
• 795c81a5 - [docs] less picky sphinx settings

Compare with previous version

changed the description

requested review from @tiago.pereira

merged

mentioned in commit dc4c3981