* fixed vertical flip TTA

* fixed test

* add doc and one more check

* fixed docs

* Add cls weight in CrossEntropyLoss

* update unittes

* resolve comments

* fix unittest bug

* Implement NAS-FCOS

1. Supports NASFCOS
2. Added NASFCOS_FPN
3. Added NASFCOS_Head
4. Added r50 config

* Revert "Implement NAS-FCOS"

This reverts commit ebae11cbbdc5b4f3cc7226c763e9254c94a9c92b.

* Revert "Revert "Implement NAS-FCOS""

This reverts commit 32bca597859a447de62ddf0f2c2f482809730da9.

* Implement NASFCOS

1. Supports NASFCOS
2. Added NASFCOS_FPN
3. Added NASFCOS_Head
4. Added r50 config

* reformatted & use list to build convs in head

1. code refornatted
2. add docstring in nasfcos_head.py
3. use a list of conv config to build convs, rather than names

* cells merged

1. merge cells in nasfpn ans nasfcos_fpn
2. add nasfpn_cell_factory.py

* reformatted

* fixed

* fixed

* Update nasfpn_cell_factory.py

* pre-commit

* bug fixed

* renamed&moved

* not done

* fixed

* fixed

* Delete remove_parawise_cfg_nasfcos_r50_nashead_4x4_coco.py

* fixed

* Create 1x_bias_lr.py

* Revert "Create 1x_bias_lr.py"

This reverts commit 54602ded283a8f646aebfa0f216350fd2ce82e6d.

* fixed

* fixed

* fixed docstring

* init modified

* formatted

* fixed

* updated

* fixed

* fixed

* fixed

* fixed

* fixed

* fixed

* pass test

* fixed

* fixed

* fixed

* rename

* 123

* nasfcos_readme

* Update README.md

* rename & leave empty

* GN->GN-head in README

* fixed

* Add upsample_cfg support in FPN

* small fix

* Add multiple extra conv sources

* small logical fix

* Add neck tests for fpn

* Add neck tests for fpn

* fixed several typos

* resolved issues

* Removed extra_convs_source option

* added necks to apis.rst

* change according to comments

* reconfigured configs

* save

* add retinanet and fix some inconsistent

* fix registry

* fix typos

* fix unmap bug

* fix some bugs

* clean

* minor fix

* fix k b

* add more cfgs

* fix detach

* add some benchmarks

* add some comments

* add more comments

* change test rpn proposal num

* add more models

* modify accroding comments

* more fix

* add pape link, fix some grammar error

* modify according to comments

* add unit test

* add unit tests for head

* fix some comments, add pape link

* add pisa roi head unittest

* fix README.md

* fix year

* Add RegNet, docs/refactor TBD

* Support RegNet

* Add RegNets

* Modify docs and refactor code

* fix pre-train bug

* Performance ready

* Use public model weights

* Performance & settings done

* Fix unittest bug

* Try another way of test config

* Add unittest and refactor

* Reformat

* Finish refactor regnet

* merge tests

* fix conflicts

* Fix resnext CI bug

* rename parameter

* Add ref link & clean _delete_

* tweaks of doc

* try resnext doctest

* add expansion in block

* remove doctest of resnext

* add res2net backbone

* add readme for res2net

* simplify res2net file

* remove some init in bottle2neck

* optimize forward code

* change init

* update results of res2net

* update config and readme

* add test and fix small issues

* fix ms rcnn show

* fix show error when empty det

* compute pure inference time instead

* fix batch size

* fix batch size bug

* use img metas

* Delete file

* reformat

* fix yapl

* fix comments

* Changed the default num_pos to make fsaf loss without pos samples more robust

* moved avg_factor out of the for loop in fsaf_head

Co-authored-by: wangxinjiang <wangxinjiang@sensetime.com>

* Allow different stride in x and y direction

* Fix bugs, pass tests

* Added new test

* Added extra test

* Correct docs

* Update docs, only square SSD

* pass tests

* Rewrite test

* rename anchor_heads to dense_heads

* update docs

* update docs

* update doc with code block

* add api doc

* update mock modules and fix requirements

* fix autodoc modules

* remove tmp file

* add release version in docs/conf.py

* fix docstring styles

* move shared_heads, mask_heads, bbox_heads under roi_heads

* move roi_extractor

* remove widlcard import

* remove ... import

* starting to merge fsaf

* changed configs

* changed bbox_coder

* get_targets fun in AnchorHead class is modified to enable additional returns from _get_target_single

* effective area assigner and tblr bbox_coder are implemented

* added fsaf_head and config

* added more fsaf configs

* fixed the nan bbox_loss by clamping bbox_pred with min=1e-4

* fixed the nan bbox_loss by clamping bbox_pred with min=1e-4

* changed config

* formatter

* changed docstring

* Added ignored pixel-class pair for FSAF

* More elegant assignment for overlapping objects

* more documents in effective area assigner

* fixed a bug in object overlay

* Added no gt_bboxes case and pytest for assigner

* Added no gt_bboxes case and pytest for assigner

* changed default config

* fixed small bugs in FSAF head doc example

* resolved pr review comments.
1. Formats of comments and docstrings are modified
2. Obsolete arguments are removed from FSAFHead
3. AssignResults class is revised to be more compatible with old version

* Added more comments to FSAFHead

* resolved pr review comments

* resolved pr review comments
1. More flexible CenterRegionAssigner bbox priority
2. More flexible tblr_coder by adding normalize_by_wh flag
3. README modifications with model link
4. More docstrings and comments
5. Reformatted
6. Change of default normalizer in TBLRBBoxCoder from 1.0 to 4.0

* Corrected FSAF model path in readme

* Fixed the bug of bbox weights in fsaf head

* trying to add fsafhead test

* Added FSAFHead test and supported zero gt in FSAFHead

Co-authored-by: wangxinjiang <wangxinjiang@sensetime.com>

* fixed webcam demo

* fixed show=True

* add warning when show=False and out_file=None

* fix bbox_loss in fovea_head.py when num_pos=0

* Remove PIL from .isort.cfg

* Make loss_bbox scalar in fovea_head.py when num_pos==0

* Sum based reduction for loss_bbox in fovea_head

* reverting to scalar implementation

* migrate to mmcv.cnn.bricks

* remove conv_module, scale, upsample

* update mmdet version and mmcv requirement

* fixed fsaf warning

* starting to merge fsaf

* changed configs

* changed bbox_coder

* get_targets fun in AnchorHead class is modified to enable additional returns from _get_target_single

* effective area assigner and tblr bbox_coder are implemented

* added fsaf_head and config

* added more fsaf configs

* fixed the nan bbox_loss by clamping bbox_pred with min=1e-4

* fixed the nan bbox_loss by clamping bbox_pred with min=1e-4

* changed config

* formatter

* changed docstring

* Added ignored pixel-class pair for FSAF

* More elegant assignment for overlapping objects

* more documents in effective area assigner

* fixed a bug in object overlay

* Added no gt_bboxes case and pytest for assigner

* Added no gt_bboxes case and pytest for assigner

* changed default config

* fixed small bugs in FSAF head doc example

* resolved pr review comments.
1. Formats of comments and docstrings are modified
2. Obsolete arguments are removed from FSAFHead
3. AssignResults class is revised to be more compatible with old version

* Added more comments to FSAFHead

* resolved pr review comments

* resolved pr review comments
1. More flexible CenterRegionAssigner bbox priority
2. More flexible tblr_coder by adding normalize_by_wh flag
3. README modifications with model link
4. More docstrings and comments
5. Reformatted
6. Change of default normalizer in TBLRBBoxCoder from 1.0 to 4.0

* Corrected FSAF model path in readme

* Fixed the bug of bbox weights in fsaf head

Co-authored-by: wangxinjiang <wangxinjiang@sensetime.com>

* fixed iou_loss

* add weight=None check

* add head-iou

* fixed comment

* add shape

* change view to asset

* Use apis/inference in detectors show_result

* Add images_out_dir arg to test apis

* Include --images_out_dir in assertion

* Fix single class parsing

* Remove single class hack

* Move import

* Add example of saving results

* Fix list number

* Refactor show_result

* Update docs

* Update __init__

* Fix CLASSES reference

* Remove unnecessar assert for custom classes

* Use - instead of _

* Use - instead of _

* Rename images_out_dir -> out_dir

* Remove unnecessary model.module

* Update show_result with all params and docstring

* Fix missing comma

* Set rescale always to True

* Remove outdated restriction from docs

* Drop pathlib

* Rename out_dir -> show_dir

* Update docstring

* More explicit code. Fix out_dir

* Update docstrings

* Flake8

Co-authored-by: mmeendez8 <miguelmndez@gmail.com>
Co-authored-by: sbugallo <sbugallo@gradiant.org>

* fix unmap

* make it more clear

* Fix bugs in regression test

* Fix test_loading

* Change .format to f-string

* Resolve comments

* Fix missing change

* replace AT_CHECK with TORCH_CHECK to support pytorch1.5

* fix compile warning, delete grid sample op

* fix ci pytorch version

* delete python3.5

* update version of pytorch and torchvision in ci

* change is_cuda to device().is_cuda()

* specify TORCH_CUDA_ARCH_LIST=6.0,7.0

* fix typo

* delete affine grid

* fix typos

* fix setup.py

* remove redundant comment

* use mmcv Registry instead of mmdet

* merged registry and builder

* merge dataset builder registry

* remove circulate import in bbox

* rename registry to builder