[WIP] Multitask Training With mmdet and yolov5 Models

[rsomani95]

This PR is the culmination of a long ongoing dicussion centered around multitask training on the icevision Discord forums. See this channel for more info.

This PR brings together a few things:

Common

• ImageClassificationHead - A flexible single layer classification head that can work with either traditional classification models (where features are pooled) or FPN outputs in a flexible way
• HybridSingleStageDetector - An mmdet -> SingleStageDetector extension that allows for classification with an arbitrary number of ImageClassificationHead heads
• HybridAugmentationsRecordDataset - A dataset that allows for different augmentations for specific tasks. In practice, the model will do multiple forward passes for groups of tasks, followed by a collective backward pass. Useful when augmentations for a particular task destroy information for a different task
• Tutorial with ExDark dataset
• Standardise documentation style - it's currently structured, but not in a consistent way

MMDet - Single Stage

• Hybrid single stage model - can be used for RetinaNet and FCOS out of the box
• Model multi augmentation forward function
• Model freezing interface
• Model unfreezing interface
• dataloaders for single image datasets
• dataloaders for multi image augmentation datasets
• higher level train_dl, valid_dl dataloader wrappers
• prediction code
• Fastai adapter
• PyTorch Lightning adapter
• Icevision style model loading API
  • RetinaNet loader
  • FCOS loader
• Tests

Ultralytics - YOLO

• Model implementation
• Model multi augmentation forward function
• Model freezing interface
• Model unfreezing interface
• dataloaders for single image datasets
• dataloaders for multi image augmentation datasets
• higher level train_dl, valid_dl dataloader wrappers
• prediction code
• Fastai adapter
• PyTorch Lightning adapter
• Icevision style model loading API
• Tutorial for exporting? Should this be a module?
• Tests

[codecov]

Codecov Report

Merging #856 (d14a453) into master (38d46bf) will decrease coverage by 15.20%.
The diff coverage is 1.22%.

@@             Coverage Diff             @@
##           master     #856       +/-   ##
===========================================
- Coverage   87.60%   72.40%   -15.21%     
===========================================
  Files         231      259       +28     
  Lines        4899     6066     +1167     
===========================================
+ Hits         4292     4392      +100     
- Misses        607     1674     +1067     

Flag	Coverage Δ
unittests	72.40% <1.22%> (-15.21%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
.../models/multitask/classification_heads/__init__.py	0.00% <0.00%> (ø)
...n/models/multitask/classification_heads/builder.py	0.00% <0.00%> (ø)
...sion/models/multitask/classification_heads/head.py	0.00% <0.00%> (ø)
...evision/models/multitask/data/dataloading_utils.py	0.00% <0.00%> (ø)
icevision/models/multitask/data/dataset.py	0.00% <0.00%> (ø)
...ion/models/multitask/engines/lightning/__init__.py	0.00% <0.00%> (ø)
...itask/engines/lightning/lightning_model_adapter.py	0.00% <0.00%> (ø)
icevision/models/multitask/mmdet/dataloaders.py	0.00% <0.00%> (ø)
icevision/models/multitask/mmdet/dtypes.py	0.00% <0.00%> (ø)
icevision/models/multitask/mmdet/pl_adapter.py	0.00% <0.00%> (ø)
... and 68 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 38d46bf...d14a453. Read the comment docs.

[FraPochetti]

I think we shall close this one.
@ai-fast-track ?

[rsomani95]

@FraPochetti @ai-fast-track converting this to a draft for now. Hope to get back to it soon!