Adaptation

[0x00b1]

Adds the following adaptation features described in https://arxiv.org/abs/1803.03243:

• Gradient Reversal layer (https://arxiv.org/abs/1505.07818)
• Image and instance-level domain adaptation losses (https://arxiv.org/abs/1803.03243)
• Changes to keras_rcnn.models.RCNN to permit domain adaptation
• Changes to keras_rcnn.preprocessing.ObjectDetectionGenerator to permit domain adaptation

From a user’s perspective, the only noteworthy change is the addition of the “adaptive” parameter in the RCNN constructor. When the user uses:

RCNN(…, adaptive=True)

They will provide an addition input to the model for target images.

[codecov-io]

Codecov Report

Merging #199 into master will decrease coverage by 1.18%. The diff coverage is 45.45%.

@@            Coverage Diff             @@
##           master     #199      +/-   ##
==========================================
- Coverage   71.42%   70.24%   -1.19%     
==========================================
  Files          28       29       +1     
  Lines        1106     1173      +67     
==========================================
+ Hits          790      824      +34     
- Misses        316      349      +33

Impacted Files	Coverage Δ
keras_rcnn/backend/cntk_backend.py	0% <0%> (ø)	:arrow_up:
keras_rcnn/models/_rcnn.py	0% <0%> (ø)	:arrow_up:
keras_rcnn/backend/theano_backend.py	0% <0%> (ø)	:arrow_up:
keras_rcnn/layers/__init__.py	100% <100%> (ø)	:arrow_up:
keras_rcnn/backend/tensorflow_backend.py	95.77% <100%> (+0.85%)	:arrow_up:
keras_rcnn/layers/losses/__init__.py	84% <44.82%> (-16%)	:arrow_down:
keras_rcnn/layers/_gradient_reversal.py	57.14% <57.14%> (ø)

---

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 46ce5f4...cdabf64. Read the comment docs.