Copy wandb param dict before training to avoid overwrites (#7317)

* Copy wandb param dict before training to avoid overwrites.

Copy the hyperparameter dict retrieved from wandb configuration before passing it to `train()`. Training overwrites parameters in the dictionary (eg scaling obj/box/cls gains), which causes the values reported in wandb to not match the input values. This is confusing as it makes it hard to reproduce a run, and also throws off wandb's Bayesian sweep algorithm.

* Cleanup

Co-authored-by: Glenn Jocher <[email protected]>

utils/loggers/wandb/sweep.py

@@ -16,8 +16,8 @@ from utils.torch_utils import select_device
 
 def sweep():
     wandb.init()
-    # Get hyp dict from sweep agent
-    hyp_dict = vars(wandb.config).get("_items")
+    # Get hyp dict from sweep agent. Copy because train() modifies parameters which confused wandb.
+    hyp_dict = vars(wandb.config).get("_items").copy()
 
     # Workaround: get necessary opt args
     opt = parse_opt(known=True)