Summary:
[export_model.py](https://github.com/facebookresearch/detectron2/blob/master/tools/deploy/export_model.py)
 requires to process sample input, and current implementation requires whole dataset to be imported, it is not a easy way for custom models, so instead of dataset, user can pass a sample image.

Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/2931

Reviewed By: alexander-kirillov

Differential Revision: D27903407

Pulled By: ppwwyyxx

fbshipit-source-id: 1e5d9b5770d2a8200abc46aad01121f0012cf5e7

Summary: Add Confidence Based CSE data samplers (coarse segm only for the moment)

Reviewed By: vkhalidov

Differential Revision: D27188563

fbshipit-source-id: f220448e87c0f31e25c1c3f98d7061c16aa68b6c

Summary: Adapt CSE Bootstrapping to use ground truth category data

Reviewed By: vkhalidov

Differential Revision: D27045679

fbshipit-source-id: 7130931a4cb66e8a9a015309814410f5f8ded171

Summary: Create CSE Base and Uniform Data Samplers

Reviewed By: vkhalidov

Differential Revision: D26691586

fbshipit-source-id: 2688698d612220b4047322ecd91df24186f2efb5

Summary:
Refactor image transform so it does nothing else than the resizing.

Before it was a bit confusing because it was resizing AND assuming that the input data was RGB and in NHWC format, and applying corresponding transforms, which depends on the input dataset

Reviewed By: vkhalidov

Differential Revision: D27081805

fbshipit-source-id: e12a23ff843e24f86feccf9f0dfd00cf628a5a5b

Summary: Correct Image List Dataset to use one image with 3 channels instead of 3 images with 1 channel

Reviewed By: vkhalidov

Differential Revision: D27080295

fbshipit-source-id: 4b90f02302858af800d6e0514972a002f92a3a3d

Summary:
In the bootstrapping datasets, we might not have annotated images, but we often know very well to which animal categories they belong. This diff aims to pass down this information.

For the details of the Instagram Datasets and `build_fb`, check the following diff D27045763

Category to class mapping should look like that and aims to convert "layman's terms" categories, used in the datasets, to category classes (and then to mesh names with CLASS_TO_MESH_NAME_MAPPING)

{F506421983}

Reviewed By: vkhalidov

Differential Revision: D27045669

fbshipit-source-id: 3865c36e622813b7591a3a969a2fa58295a48744

Summary:
Hi~
We propose a new object detection method: Sparse R-CNN. The code is based on detectron2.
I will appreciate it if detectron2 could include it.

Thanks very much.

Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/2911

Reviewed By: alexander-kirillov

Differential Revision: D27899736

Pulled By: ppwwyyxx

fbshipit-source-id: 6b2032f6870e6b0643513fe4829aef1397a4bd02

Summary: This diff fixes a bug in the visualization framework, which image_inference and video_inference debugging tools utilize to visualize RCNN model outputs.

Reviewed By: sampepose

Differential Revision: D27875273

fbshipit-source-id: 4afc88b11b22c902a8956a7d6600a0dbf2550cde

Summary:
Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/2932

Pull Request resolved: https://github.com/fairinternal/detectron2/pull/533

Actually only the one between backbone and proposal generator is redundant unless
we implement a quantized version for roi_align (currently roi align is expecting
a float32 input, so we have to dequantize it in the code)

Reviewed By: vkuzo, ppwwyyxx

Differential Revision: D27837648

fbshipit-source-id: 74fd87b31d6f701dd2917e90efad558e9d5c5b21

Summary: Pull Request resolved: https://github.com/fairinternal/detectron2/pull/534

Reviewed By: alexander-kirillov

Differential Revision: D27877352

Pulled By: ppwwyyxx

fbshipit-source-id: 03e6aa99385115014b02952aa9f97282ed152752

Reviewed By: alexander-kirillov

Differential Revision: D27809203

fbshipit-source-id: 93daeaffbd0ebb0923191b810f78572338918cf9

Reviewed By: ppwwyyxx

Differential Revision: D27823003

fbshipit-source-id: 67e6905f3c5c7bb1f593ee004160b195925f6d39

Summary: A structure to represent masks in RoIs.

Reviewed By: rbgirshick

Differential Revision: D27810570

fbshipit-source-id: 53e4772c91a16119650d44f7bd895f2ce60d82cb

Reviewed By: alexander-kirillov

Differential Revision: D27771715

fbshipit-source-id: afad994dcc7ef1e578b59fd8de9273323e42560d

Summary: This is easier to config: just use [30, 60, 90, 120] instead of two separate configs

Reviewed By: theschnitz

Differential Revision: D27108999

fbshipit-source-id: f8ad25c62656c429d30ea29b7fa4dcbe960540a5

Reviewed By: alexander-kirillov

Differential Revision: D27770785

fbshipit-source-id: 0128bdcc64ad22862b072946dabcc7b261204a2f

Summary:
The new system no longer has pre-defined structure, so using a DefaultTrainer(cfg) makes it confusing what cfg keys are needed.

The code in DefaultTrainer is short enough to live in train_net so users can see directly what config is needed. This way we also avoid reinventing a trainer that takes new configs. We'll be able to remove newconfig/defaults.py then (or keep it private for reference)

Reviewed By: theschnitz, alexander-kirillov

Differential Revision: D27716057

fbshipit-source-id: 248446d19bfaf601653bb25bb09765f99e3a8432

Summary: Early exit from loss iterations / methods in case where no valid points are provided.

Reviewed By: MarcSzafraniec

Differential Revision: D27519651

fbshipit-source-id: e78df3da2e9186d98247c1c306a769641e4ddfa0

Summary: Others to be added later

Reviewed By: alexander-kirillov

Differential Revision: D27759645

fbshipit-source-id: 0b41b434eb6302fbe310fa2a8cf45cef17b5682f

Reviewed By: wat3rBro

Differential Revision: D27752054

fbshipit-source-id: 6828d65f09cf922fd61cff1860a5e0b2ed644e5a

Reviewed By: sujitoc, ppwwyyxx

Differential Revision: D27721286

fbshipit-source-id: 2869610272308314ce913a1d94c78cf809f0d485

Reviewed By: alexander-kirillov

Differential Revision: D27742322

fbshipit-source-id: 6cb6f3607b3baecf897ec6e824118f143e715712

Summary: expose more modules to higher level

Reviewed By: alexander-kirillov

Differential Revision: D27730955

fbshipit-source-id: f8912959e382c4b56b8012d0bd2a4540f75740d4

Summary:
Trainer can now checkpoint its states about:
1. optimizer & grad_scaler
2. current iter
3. hooks (now including scheduler)

Reviewed By: alexander-kirillov

Differential Revision: D27715660

fbshipit-source-id: 20c4a0a613c7182fe484e5c99775e4d525137bac

Summary: simplify DefaultTrainer

Reviewed By: alexander-kirillov

Differential Revision: D27710586

fbshipit-source-id: 77e9d8e6e2b40bf91af16af6cd236dbcbf86eca7

Reviewed By: alexander-kirillov

Differential Revision: D27692660

fbshipit-source-id: a7019f9b97b9b3918bcd129ed02717d77aa0bd68