This adds an example of how to use the YOLO object detection to go alongside the documentation provided for it. The need for this example to be simple also prompted changes to dataset reading for object detection. These changes include refactoring the existing object detection-exclusive code in the loaders and adding a new function for handling dataset loading plus the automatic patching.
Meanwhile, some bugs were found in compute_full_test_accuracy (particularly when doing semantic segmentation). These caused it to crash when trying to 'fix' variables that weren't made (due to creation being guarded in an if-expression) and when trying to concatenate semantic segmentation losses due to how the losses were returned. These bugs were fixed with a change to those if-statements and some of the test operations in the graph.
All of the supported problem types (including the new object detection example) have been tested to work and the tests still pass.
This adds an example of how to use the YOLO object detection to go alongside the documentation provided for it. The need for this example to be simple also prompted changes to dataset reading for object detection. These changes include refactoring the existing object detection-exclusive code in the loaders and adding a new function for handling dataset loading plus the automatic patching.
Meanwhile, some bugs were found in compute_full_test_accuracy (particularly when doing semantic segmentation). These caused it to crash when trying to 'fix' variables that weren't made (due to creation being guarded in an if-expression) and when trying to concatenate semantic segmentation losses due to how the losses were returned. These bugs were fixed with a change to those if-statements and some of the test operations in the graph.
All of the supported problem types (including the new object detection example) have been tested to work and the tests still pass.