* Initial multi-precision training

Adds fp16 support via apex.amp
Also switches communication to apex.DistributedDataParallel

* Add Apex install to dockerfile

* Fixes from @fmassa review

Added support to tools/test_net.py
SOLVER.MIXED_PRECISION -> DTYPE \in {float32, float16}
apex.amp not installed now raises ImportError

* Remove extraneous apex DDP import

* Move to new amp API

* proposals from RPN per image during training

* README

* Update README for setting FPN_POST_NMS_TOP_N_TRAIN

* Update README.md

* removing extra space change

* add color jitter augmentation

* fix spelling

* make pixel indexes 0-based for bounding box in pascal voc dataset

* replacing all instances of torch.distributed.deprecated with torch.distributed

* replacing all instances of torch.distributed.deprecated with torch.distributed

* add GroupNorm

* add GroupNorm -- sort out yaml files

* use torch.nn.GroupNorm instead, replace 'use_gn' with 'conv_block' and use 'BaseStem'&'Bottleneck' to simply codes

* modification on 'group_norm' and 'conv_with_kaiming_uniform' function

* modification on yaml files in configs/gn_baselines/ and reduce the amount of indentation and code duplication

* use 'kaiming_uniform' to initialize resnet, disable gn after fc layer, and add dilation into ResNetHead

* agnostic-regression for bbox

* please set 'STRIDE_IN_1X1' to be 'False' when backbone use GN

* add README.md for GN

* add dcn from mmdetection

https://gist.github.com/botcs/a59f3f59e22e5df93e3e5e4f86718af3

* fix py2

* fix py2

* support RLE and binary mask

* do not convert to numpy

* be consistent with Detectron

* delete wrong comment

* [WIP] add tests for segmentation_mask

* update tests

* minor change

* Refactored segmentation_mask.py

* Add unit test for segmentation_mask.py

* Add RLE support for BinaryMaskList

* PEP8 black formatting

* Minor patch

* Use internal  that handles 0 channels

* Fix polygon slicing

add tqdm in line32 : RUN pip install ninja yacs cython matplotlib opencv-python tqdm

Fix a bug.
Romove the echo command in line 36: RUN conda install pytorch-nightly cudatoolkit=${CUDA} -c pytorch 
To enable conda installation of pytorch-nightly.

* add a FORCE_CUDA flag

Following discussion [here](https://github.com/facebookresearch/maskrcnn-benchmark/issues/167), this seemed the best solution

* Update Dockerfile

* Update setup.py

* add FORCE_CUDA as an ARG

* 	modified:   docker/Dockerfile
	modified:   setup.py

* small fix to readme of demo

* remove test print

* keep ARG_CUDA

* remove env value and use the one from ARG

* keep same formatting as source

* change proposed by @miguelvr

* Update INSTALL.md

This reverts commit f0318794.

* fixes to dockerfile

* replaces local installation by git clone

* Merge branch 'master' of /home/braincreator/projects/maskrcnn-benchmark with conflicts.

* rolls back the breaking AT dispatch changes (#555)

* revert accidental docker changes

* revert accidental docker changes (2)

* support for any one cuda device

* Revert "support for any one cuda device"

This reverts commit 0197e4e2ef18ec41cc155f3ae2a0face5b77e1e9.

* support runnning for anyone cuda device

* using safe CUDAGuard rather than intrinsic CUDASetDevice

* supplement a header dependency (test passed)

* Support for  arbitrary GPU device.

* Support for arbitrary GPU device.

* add docs for two method to control devices

Fix dispatch breakage

* Added a timer to benchmark model inference time in addition to total runtime.

* Updated FBNet configs and included some baselines benchmark results.

* Added a unit test for detectors.

* Add links to the models

* fix INSTALL.md

* fix PR

* Update INSTALL.md

Using existing get_world_size to prevent AttributeError 'torch.distributed' has no attribute 'is_initialized'. (#511)

* Add new section "Projects using maskrcnn-benchmark".

* Update README.md

update the format.

* Update README.md

* Add coco_2017_train and coco_2017_val

* Update README.md

Add the instructions about COCO_2017

* Update the pip install.
Adding tqdm which is used in engine/inference.py

* make pixel indexes 0-based for bounding box in pascal voc dataset

* replacing all instances of torch.distributed.deprecated with torch.distributed

* replacing all instances of torch.distributed.deprecated with torch.distributed

* add GroupNorm

* add GroupNorm -- sort out yaml files

* use torch.nn.GroupNorm instead, replace 'use_gn' with 'conv_block' and use 'BaseStem'&'Bottleneck' to simply codes

* modification on 'group_norm' and 'conv_with_kaiming_uniform' function

* modification on yaml files in configs/gn_baselines/ and reduce the amount of indentation and code duplication

* use 'kaiming_uniform' to initialize resnet, disable gn after fc layer, and add dilation into ResNetHead

* agnostic-regression for bbox

* please set 'STRIDE_IN_1X1' to be 'False' when backbone use GN

* add README.md for GN

There were two `RESNETS` sections, which overrode each other, leading to error messages like: 

```
RuntimeError: Error(s) in loading state_dict for GeneralizedRCNN:
	size mismatch for backbone.fpn.fpn_inner1.weight: copying a param with shape torch.Size([256, 256, 1, 1]) from checkpoint, the shape in current model is torch.Size([1024, 256, 1, 1]).
...
	size mismatch for roi_heads.mask.feature_extractor.mask_fcn1.weight: copying a param with shape torch.Size([256, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 1024, 3, 3]).
```

This just combines them back into a single section, while maintaining all param values. That got the model running again for me.

* Supported any feature map size for average pool.
* Different models may have different feature map size.

* Used registry to register keypoint and mask heads.

* Passing in/out channels between modules when creating the model.

Passing in/out channels between modules when creating the model. This simplifies the code to compute the input channels for feature extractors and makes the predictors independent of the backbone architectures.
* Passed in_channels to rpn and head builders.
* Set out_channels to model modules including backbone and feature extractors.
* Moved cfg.MODEL.BACKBONE.OUT_CHANNELS to cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS as it is not used by all architectures. Updated config files accordingly.

For new architecture modules, the return module needs to contain a field called 'out_channels' to indicate the output channel size.

* Added unit test for box_coder and nms.

* Added FBNet architecture.

* FBNet is a general architecture definition to support efficient architecture search and MaskRCNN2GO.
* Included various efficient building blocks (inverted residual, shuffle, separate dw conv, dw upsampling etc.)
* Supported building backbone, rpn, detection, keypoint and mask heads using efficient building blocks.
* Architecture could be defined in `fbnet_modeldef.py` or in `cfg.MODEL.FBNET.ARCH_DEF` directly.
* A few baseline architectures are included.

* Added various unit tests.

* build and run backbones.
* build and run feature extractors.
* build and run predictors.

* Added a unit test to verify all config files are loadable.

* make pixel indexes 0-based for bounding box in pascal voc dataset

* replacing all instances of torch.distributed.deprecated with torch.distributed

* replacing all instances of torch.distributed.deprecated with torch.distributed

* add GroupNorm

* add GroupNorm -- sort out yaml files

* use torch.nn.GroupNorm instead, replace 'use_gn' with 'conv_block' and use 'BaseStem'&'Bottleneck' to simply codes

* modification on 'group_norm' and 'conv_with_kaiming_uniform' function

* modification on yaml files in configs/gn_baselines/ and reduce the amount of indentation and code duplication

* use 'kaiming_uniform' to initialize resnet, disable gn after fc layer, and add dilation into ResNetHead

* agnostic-regression for bbox

* please set 'STRIDE_IN_1X1' to be 'False' when backbone use GN

I'm using pycharm to debug the code on a remote server, the remote debugging seems to be performed by pytest and it pops errors like:
train_net.py E
test setup failed
file /tmp/pycharm_project_269/tools/train_net.py, line 79
  def test(cfg, model, distributed):
E       fixture 'cfg' not found
>       available fixtures: cache, capfd, capfdbinary, caplog, capsys, capsysbinary, doctest_namespace, monkeypatch, pytestconfig, record_property, record_xml_attribute, recwarn, tmp_path, tmp_path_factory, tmpdir, tmpdir_factory
>       use 'pytest --fixtures [testpath]' for help on them.


it seems the function name ‘test()’ has come conflict with pytest, so it may be better use another name.

* Remove Detectron dependency

I have looked into the boxes.py to swap [these lines](https://github.com/facebookresearch/Detectron/blob/8170b25b425967f8f1c7d715bea3c5b8d9536cd8/detectron/utils/boxes.py#L51L52):
```
import detectron.utils.cython_bbox as cython_bbox
import detectron.utils.cython_nms as cython_nms
```

```
from maskrcnn_benchmark.structures.boxlist_ops import boxlist_iou
from maskrcnn_benchmark.structures.boxlist_ops import boxlist_nms
```
However some functions are missing from the `boxlist_ops` like the [`soft_nms`](https://github.com/facebookresearch/Detectron/blob/master/detectron/utils/cython_nms.pyx#L98L203) .

So I just tried to modify the `maskrcnn-benchmark/tools/cityscapes/convert_cityscapes_to_coco.py` script.
Here we have `polys_to_boxes` function from `segms.py` and I could not find its analogous in the maskrcnn_benchmark lib. It seems to me that the original function in `segms.py` is using pure lists so I just wrote two auxiliary functions reusing the boxList's convert method( https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/maskrcnn_benchmark/structures/bounding_box.py#L67L70 )
and Detectron's polys_to_boxes ( https://github.com/facebookresearch/Detectron/blob/b5dcc0fe1d091cb70f9243939258215dd63e3dfa/detectron/utils/segms.py#L135L140 ):
```
def poly_to_box(poly):
    """Convert a polygon into a tight bounding box."""
    x0 = min(min(p[::2]) for p in poly)
    x1 = max(max(p[::2]) for p in poly)
    y0 = min(min(p[1::2]) for p in poly)
    y1 = max(max(p[1::2]) for p in poly)
    box_from_poly = [x0, y0, x1, y1]

    return box_from_poly

def xyxy_to_xywh(xyxy_box):
    xmin, ymin, xmax, ymax = xyxy_box
    TO_REMOVE = 1
    xywh_box = (xmin, ymin, xmax - xmin + TO_REMOVE, ymax - ymin + TO_REMOVE)
    return xywh_box
```

* removed leftovers

* Update convert_cityscapes_to_coco.py

Finishing the clean up in https://github.com/facebookresearch/maskrcnn-benchmark/pull/455, unsetting the proper variable. 

In general, thanks for making this so easy to install! I'd run into all kinds of versioning issues (version of Ubuntu not playing nicely with versions of CUDA/pytorch/libraries) trying to install other libraries implementing these algorithms. I'm super impressed by the quality of support, and the easy install, for this library.