Default FP16 TensorRT export (#6798)

* Assert engine precision #6777

* Default to FP32 inputs for TensorRT engines

* Default to FP16 TensorRT exports #6777

* Remove wrong line #6777

* Automatically adjust detect.py input precision #6777

* Automatically adjust val.py input precision #6777

* Add missing colon

* Cleanup

* Cleanup

* Remove default trt_fp16_input definition

* Experiment

* Reorder detect.py if statement to after half checks

* Update common.py

* Update export.py

* Cleanup

Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

detect.py

@@ -97,6 +97,10 @@ def run(weights=ROOT / 'yolov5s.pt',  # model.pt path(s)
     half &= (pt or jit or onnx or engine) and device.type != 'cpu'  # FP16 supported on limited backends with CUDA
     if pt or jit:
         model.model.half() if half else model.model.float()
+    elif engine and model.trt_fp16_input != half:
+        LOGGER.info('model ' + (
+            'requires' if model.trt_fp16_input else 'incompatible with') + ' --half. Adjusting automatically.')
+        half = model.trt_fp16_input
 
     # Dataloader
     if webcam: