* 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

* 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

* Add RetinetNet parameters in cfg.

* hot fix.

* Add the retinanet head module now.

* Add the function to generate the anchors for RetinaNet.

* Add the SigmoidFocalLoss cuda operator.

* Fix the bug in the extra layers.

* Change the normalizer for SigmoidFocalLoss

* Support multiscale in training.

* Add retinannet  training script.

* Add the inference part of RetinaNet.

* Fix the bug when building the extra layers in retinanet.
Update the matching part in retinanet_loss.

* Add the first version of the inference of RetinaNet.
Need to check it again to see if is there any room for speed
improvement.

* Remove the  retinanet_R-50-FPN_2x.yaml first.

* Optimize the retinanet postprocessing.

* quick fix.

* Add script for training RetinaNet with ResNet101 backbone.

* Move cfg.RETINANET to cfg.MODEL.RETINANET

* Remove the variables which are not used.

* revert boxlist_ops.
Generate Empty BoxLists instead of [] in retinanet_infer

* Remove the not used commented lines.
Add NUM_DETECTIONS_PER_IMAGE

* remove the not used codes.

* Move retinanet related files under Modeling/rpn/retinanet

* Add retinanet_X_101_32x8d_FPN_1x.yaml script.
This model is not fully validated. I only trained it around 5000
iterations and everything is fine.

* set RETINANET.PRE_NMS_TOP_N as 0 in level5 (p7), because previous setting may generate zero detections and could cause
the program break.
This part is used in original Detectron setting.

* Fix the rpn only bug when the training ends.

* Minor improvements

* Comments and add Python-only implementation

* Bugfix and remove commented code

* keep the generalized_rcnn same.
Move the build_retinanet inside build_rpn.

* Add USE_C5 in the MODEL.RETINANET

* Add two configs using P5 to generate P6.

* fix the bug when loading the Caffe2 ImageNet pretrained model.

* Reduce the code depulication of RPN loss and RetinaNet loss.

* Remove the comment which is not used.

* Remove the hard coded number of classes.

* share the foward part of rpn inference.

* fix the bug in rpn inference.

* Remove the conditional part in the inference.

* Bug fix: add the utils file for permute and flatten of the box
prediction layers.

* Update the comment.

* quick fix. Adding import cat.

* quick fix: forget including import.

* Adjust the normalization part according to Detectron's setting.

* Use the bbox reg normalization term.

* Clean the code according to recent review.

* Using CUDA version for training now. And the python version for training
on cpu.

* rename the directory to retinanet.

* Make the train and val datasets are consistent with mask r-cnn setting.

* add comment.

* Add RPN config files

* Add more RPN models

* [WIP] Keypoints inference on C2 models work

* Training seems to work

Still gives slightly worse results

* e2e training works but gives 3 and 5 mAP less

* Add modification proposed by @ChangErgou

Improves mAP by 1.5 points, to 0.514 and 0.609

* Keypoints reproduce expected results

* Clean coco.py

* Linter + remove unnecessary code

* Merge criteria for empty bboxes in has_valid_annotation

* Remove trailing print

* Add demo support for keypoints

Still need further cleanups and improvements, like adding fields support for the other ops in Keypoints

* More cleanups and misc improvements

* Fixes after rebase

* Add information to the readme

* Fix md formatting

* 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

* add force json option

* fix the same issue as #185

* bug fix

* cityscapes config

* update paths catalog

* discard config change

* organize code for more-datasets

* use better representation for coco-style datasets

* rename coco-style config

* remove import

* chmod 644

* make the config more verbose

* update readme

* rename

* chmod

* add support for pascal voc dataset and evaluate

* optimization for adding voc dataset

* make inference.py dataset-agnostic; add use_difficult option to voc dataset

* handle voc difficult objects correctly

* Remove dependency on lxml plus minor improvements

* More cleanups

* More comments and improvements

* Lint fix

* Move configs to their own folder