Allow all constituent object detection losses to be logged

[AdeelH]

Overview

This PR removes Learner.build_metric_names(). This means that all metrics returned by Learner.train_step() and Learner.validation_step() can be logged without needing to be pre-specified in Learner.build_metric_names(), which means losses from different object detection models (which can have different names) can be logged without needing to override the default ObjectDetectionLearner.

After the changes, training logs with a FasterRCNN model look like so:

{'epoch': 2,
 'loss_classifier': 0.06613687425851822,
 'loss_box_reg': 0.06521709263324738,
 'loss_objectness': 0.01976802386343479,
 'loss_rpn_box_reg': 0.0028790205251425505,
 'train_loss': 0.15400101244449615,
 'train_time': '0:00:12.718091',
 'mAP': 0.007138107431842703,
 'mAP50': 0.047601330107510685,
 'valid_time': '0:00:12.937088'}

With an external model (SSD):

{'epoch': 4,
 'bbox_regression': 2.0971059799194336,
 'classification': 2.8579580783843994,
 'train_loss': 4.955064296722412,
 'train_time': '0:00:19.800291',
 'mAP': 0.0027229238176515428,
 'mAP50': 0.011754015726541719,
 'valid_time': '0:00:11.715102'}

Other changes:

• Metric aggregation has been refactored slightly.
• New unit tests for metric aggregation and log writing.

Checklist

• Added needs-backport label if PR is bug fix that applies to previous minor release
• Ran scripts/format_code and committed any changes
• Documentation updated if needed
• PR has a name that won't get you publicly shamed for vagueness

Notes

N/A

Testing Instructions

• Run an object detection job and check the logs.
• See new unit tests.

[codecov]

Codecov Report

Merging #1716 (97ad7e9) into master (8e8a20c) will increase coverage by 0.00%.
The diff coverage is 91.66%.

@@           Coverage Diff           @@
##           master    #1716   +/-   ##
=======================================
  Coverage   75.74%   75.75%           
=======================================
  Files         192      192           
  Lines        9319     9304   -15     
=======================================
- Hits         7059     7048   -11     
+ Misses       2260     2256    -4     

Impacted Files	Coverage Δ
...ervision/pytorch_learner/object_detection_utils.py	77.22% <ø> (-0.13%)	⬇️
...rastervision/pytorch_learner/regression_learner.py	36.25% <0.00%> (-3.75%)	⬇️
...ch_learner/rastervision/pytorch_learner/learner.py	69.92% <90.00%> (-0.77%)	⬇️
...ervision/pytorch_learner/classification_learner.py	100.00% <100.00%> (ø)
...vision/pytorch_learner/object_detection_learner.py	92.00% <100.00%> (-0.21%)	⬇️
...pytorch_learner/object_detection_learner_config.py	84.78% <100.00%> (ø)
...n/pytorch_learner/semantic_segmentation_learner.py	98.11% <100.00%> (-0.04%)	⬇️
...earner/rastervision/pytorch_learner/utils/utils.py	94.15% <100.00%> (+0.81%)	⬆️
...er/dataset/visualizer/classification_visualizer.py	100.00% <0.00%> (+2.63%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.