Reviewed By: rbgirshick

Differential Revision: D6848494

fbshipit-source-id: 1bc4605443060891872ec1cc7c78216980440003

Summary:
Generalizations to enable building forward passes and param update
operations separately.

Reviewed By: rbgirshick

Differential Revision: D6827027

fbshipit-source-id: 9d2ace349b4ebdf8993baa2f2c1529e2c5ce751d

Reviewed By: rbgirshick

Differential Revision: D6845192

fbshipit-source-id: cf172ab7c9fa7137d35839b721b517576994eb35

Summary:
Under a distributed setting (multiple nodes), the loss scale may need
to account for the number of nodes participating in the algorithm. This method
provides a reasonable way for providing this config.

Reviewed By: rbgirshick

Differential Revision: D6826976

fbshipit-source-id: d224200710b0f3c3245851874c9f4f3ab871710f

Summary:
The `AffineChannel` op wrapper did not use the officially sanctioned
version of parameter creation, namely using `ModelHelper.create_param()`
Presumably this was because of the `share_with` option where you could share
parameters.

While doing this though, the parameters were never initialized! This probably
doesn't matter for the algorithm (those params get overwritten pretty quickly)
but it's a bug nevertheless. I decided to kill the `share_with` option since
this isn't used anyway and resort to the standard way of creating weights and
biases.

Reviewed By: rbgirshick

Differential Revision: D6826914

fbshipit-source-id: 65628c59b085b9ab160006b003dd40dbefa2f7c1

Summary:
This diff starts a series of diffs to refactor the detectron codebase
so it can be used in an elastic data parallel context (see
https://our.intern.facebook.com/intern/dex/caffe2/elastic-data-parallel-model-for-distributed-training/)
without any forks.

Specifically, this diff does the following:

 - Splits out the `TrainingStats` class so it can be reused / composed
 - Slightly refactor of `initialize_weights_from_file()`
 - Reduces the verbosity of some logs

Hopefully, nothing controversial here :)

Reviewed By: rbgirshick

Differential Revision: D6826820

fbshipit-source-id: fc15209a0ff50e5d09281e36173198c77aa77a12

Summary: Revise print_net to show the full forward model when there are multiple losses (which is always the case).

Reviewed By: rbgirshick

Differential Revision: D6829753

fbshipit-source-id: 53c1d63d95a78f83a0275b90cad1517315786e7d

Reviewed By: KaimingHe

Differential Revision: D6842986

fbshipit-source-id: c8370e2f831239b5489dc723738c6bdf5708a40d

Summary:
Added a script to convert pkl model to pb.
- Supported Faster R-CNN detection model.
- Supported converting to CPU and GPU model.
- Fused AffineChannel to Conv.
- Verified model after conversion.

Reviewed By: rbgirshick

Differential Revision: D6783633

fbshipit-source-id: 19706d3074a4a784a2161a695d7c534e014ebb3f

Reviewed By: rbgirshick

Differential Revision: D6829499

fbshipit-source-id: 66fe40ef1b2c8c560ce769faa27d107c3c7be841

Reviewed By: rbgirshick

Differential Revision: D6829350

fbshipit-source-id: a5fbaa8900fab28602d7a993d63206ca67b24975

Reviewed By: rbgirshick

Differential Revision: D6829206

fbshipit-source-id: 06e8917366a68e4c2e44637d784b019ba2d6fddf

Reviewed By: rbgirshick

Differential Revision: D6829095

fbshipit-source-id: 1d36fd7adf1596350017c437a4365d360a21dccd

Reviewed By: rbgirshick

Differential Revision: D6828915

fbshipit-source-id: dfc8bd970c0555d56bfd7ad313550b0a12c154f2

Reviewed By: rbgirshick

Differential Revision: D6793771

fbshipit-source-id: 4716102f41f12801c6f715b6f7ad7e438591b295

Reviewed By: rbgirshick

Differential Revision: D6793159

fbshipit-source-id: 8d8f12915809238d3dcf8066a94446e0feb4cc46

Reviewed By: rbgirshick

Differential Revision: D6784702

fbshipit-source-id: b5f81d858f9e9d33c048c39f849d0e29e8db3cb1

Differential Revision: D6782696

fbshipit-source-id: a932e0f75138be23cee09495bf28b99f06b920eb

fbshipit-source-id: d6798bb3ead07e6e3da5edebc53b946e6cfa0807