GitLab

* Add rough implementation of RetinaNet.

* Move AnchorGenerator to a seperate file.

* Move box similarity to Matcher.

* Expose extra blocks in FPN.

* Expose retinanet in __init__.py.

* Use P6 and P7 in FPN for retinanet.

* Use parameters from retinanet for anchor generation.

* General fixes for retinanet model.

* Implement loss for retinanet heads.

* Output reshaped outputs from retinanet heads.

* Add postprocessing of detections.

* Small fixes.

* Remove unused argument.

* Remove python2 invocation of super.

* Add postprocessing for additional outputs.

* Add missing import of ImageList.

* Remove redundant import.

* Simplify class correction.

* Fix pylint warnings.

* Remove the label adjustment for background class.

* Set default score threshold to 0.05.

* Add weight initialization for regression layer.

* Allow training on images with no annotations.

* Use smooth_l1_loss with beta value.

* Add more typehints for TorchScript conversions.

* Fix linting issues.

* Fix type hints in postprocess_detections.

* Fix type annotations for TorchScript.

* Fix inconsistency with matched_idxs.

* Add retinanet model test.

* Add missing JIT annotations.

* Remove redundant model construction

Make tests pass

* Fix bugs during training on newer PyTorch and unused params in DDP

Needs cleanup and to add back support for images with no annotations

* Cleanup resnet_fpn_backbone

* Use L1 loss for regression

Gives 1mAP improvement over smooth l1

* Disable support for images with no annotations

Need to fix distributed first

* Fix retinanet tests

Need to deduplicate those box checks

* Fix Lint

* Add pretrained model

* Add training info for retinanet
Co-authored-by: Hans Gaiser <hansg91@gmail.com>
Co-authored-by: Hans Gaiser <hans.gaiser@robovalley.com>
Co-authored-by: Hans Gaiser <hans.gaiser@robohouse.com>

Co-authored-by: vfdev <vfdev.5@gmail.com>

Change the default anchor generator settings to default Faster RCNN version(without FPN), so it is hopefully less misleading. Also add an assert to grad_anchors functions to make sure people use it as it is aspected.

* Readme for video models.

* Adressing @fmassa comments
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Replace nonzero by where, now that it works with just a condition

* partially enable mypy for .models

* fix existing errors

* ignore error instead of using Union

* fix type hints and move degenerate boxes to a function in torchvision.models.detection.generalized_rcnn

* format code

* format code

* changed to static method

* revert imports

* changed to method

* revert procedure for degenerating boxes

* Update transform.py

* Update transform.py

* Try remove eager scripting calls

* remove script call
Co-authored-by: eellison <eellison@fb.com>
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* Remove interpolate in favor of PyTorch's implementation

* Bugfix

* Bugfix

* Add norm_layer to MobileNetV2

* Add simple test case

* Small fix

Co-authored-by: eellison <eellison@fb.com>

* add layer freezing param to maskrcnn_resnet50_fpn

* freeze ayer param in keypointrcnn_resnet50_fpn

* layer freeze tests for mask and keypoint rcnn

* correct linting errors

* correct linting errors.

* correct linting errors

* Fix #2221, DenseNet issue with gradient checkpoints (memory_efficient=True)

* Add grad/param count test for mem_efficient densenet

* Fixing nms on boxes when no detection

* test

* Fix for scale_factor computation

* remove newline

* Fix for mask_rcnn dynanmic axes

* Clean up

* Update transform.py

* Fix for torchscript

* Fix scripting errors

* Fix annotation

* Fix lint

* Fix annotation

* Fix for interpolate scripting

* Fix for scripting

* refactoring

* refactor the code

* Fix annotation

* Fixed annotations

* Added test for resize

* lint

* format

* bump ORT

* ort-nightly version

* Going to ort 1.1.0

* remove version

* install typing-extension

* Export model for images with no detection

* Upgrade ort nightly

* update ORT

* Update test_onnx.py

* updated tests

* Updated tests

* merge

* Update transforms.py

* Update cityscapes.py

* Update celeba.py

* Update caltech.py

* Update pkg_helpers.bash

* Clean up

* Clean up for dynamic split

* Remove extra casts

* flake8

* Fix for mask rcnn no detection export

* clean up

* Enable mask rcnn tests

* Added test

* update ORT

* Update .travis.yml

* fix annotation

* Clean up roi_heads

* clean up

* clean up misc ops

Another instance of integer division using the division operator. In this case line 266 already shows the correct formulation, so line 185 only needs the update.

Integer division using the div operator is deprecated and will throw a RuntimeError in PyTorch 1.6 (and on PyTorch Master very soon). Running a test build with a recent Torchvision commit and integer division using div ('/') disabled revealed this integer division. 

I'll re-run the tests once this is fixed in case it's masking additional issues.

* Make copy of targets in GeneralizedRCNNTransform

* Fix flake8

Co-authored-by: eellison <eellison@fb.com>

* freeze layers only if pretrained backbone is used

If pretrained backbone is not used and one intends to train the entire network from scratch, no layers should be frozen.

* function argument to control the trainable features

Depending on the size of dataset one might want to control the number of tunable parameters in the backbone, and this parameter in hyper parameter optimization for the dataset. It would be nice to have this function support this.

* ensuring tunable layer argument is valid

* backbone freezing in fasterrcnn_resnet50_fpn

Handle backbone freezing in fasterrcnn_resnet50_fpn function rather than the resnet_fpn_backbone function that it uses to get the backbone.

* remove layer freezing code

layer freezing code has been moved to fasterrcnn_resnet50_fpn function that consumes resnet_fpn_backbone function.

* correcting linting errors

* correcting linting errors

* move freezing logic to resnet_fpn_backbone

Moved layer freezing logic to resnet_fpn_backbone with an additional parameter.

* remove layer freezing from fasterrcnn_resnet50_fpn

Layer freezing logic has been moved to resnet_fpn_backbone. This function only ensures that the all layers are made trainable if pretrained models are not used.

* update example resnet_fpn_backbone docs

* correct typo in var name

* correct indentation

* adding test case for layer freezing in faster rcnn

This PR adds functionality to specify the number of trainable layers while initializing the faster rcnn using fasterrcnn_resnet50_fpn function. This commits adds a test case to test this functionality.

* updating layer freezing condition for clarity

More information in PR

* remove linting errors

* removing linting errors

* removing linting errors

* Check target boxes input on generalized_rcnn.py

* Fix target box validation in generalized_rcnn.py

* Add tests for input validation of detection models

* Fix mypy type annotations

* follow torchscript Tuple type

* redefine torch_choice output type

* change the type in cached_grid_anchors

* minor bug
Co-authored-by: Guanheng Zhang <zhangguanheng@devfair0197.h2.fair>
Co-authored-by: Guanheng Zhang <zhangguanheng@learnfair0341.h2.fair>

* Fixing nms on boxes when no detection

* test

* Fix for scale_factor computation

* remove newline

* Fix for mask_rcnn dynanmic axes

* Clean up

* Update transform.py

* Fix for torchscript

* Fix scripting errors

* Fix annotation

* Fix lint

* Fix annotation

* Fix for interpolate scripting

* Fix for scripting

* refactoring

* refactor the code

* Fix annotation

* Fixed annotations

* Added test for resize

* lint

* format

* bump ORT

* ort-nightly version

* Going to ort 1.1.0

* remove version

* install typing-extension

* Export model for images with no detection

* Upgrade ort nightly

* update ORT

* Update test_onnx.py

* updated tests

* Updated tests

* merge

* Update transforms.py

* Update cityscapes.py

* Update celeba.py

* Update caltech.py

* Update pkg_helpers.bash

* Clean up

* Clean up for dynamic split

* Remove extra casts

* flake8

At the moment in the ASPP-Layer the number of output channels are predefined as a constant,
which is good for DeepLab but not necessairly in other projects, where another out-channel Nr. is required.

Also the number of "atrous rates" is fixed to three, which also could be sometimes more or less depending on the notwork-arch.
Again these fixed values may make sense in DeepLab-Model but not necessarily in other type of models.

This pull-req. contains the needed changes to make ASPP-Layer generic.

* For detection model, replace L1 loss with beta smooth L1 loss to achieve better performance.

* Add type annotations for torchscript

* Resolve E226

* Fixing nms on boxes when no detection

* test

* Fix for scale_factor computation

* remove newline

* Fix for mask_rcnn dynanmic axes

* Clean up

* Update transform.py

* Fix for torchscript

* Fix scripting errors

* Fix annotation

* Fix lint

* Fix annotation

* Fix for interpolate scripting

* Fix for scripting

* refactoring

* refactor the code

* Fix annotation

* Fixed annotations

* Added test for resize

* lint

* format

* bump ORT

* ort-nightly version

* Going to ort 1.1.0

* remove version

* install typing-extension