* h range clip range edit in hsv augmentation
h range is [0., 179,]
* Update datasets.py
reduced indexing operations and used inplace clip for hsv. Two clips are used unfortunately (double clip of axis 0), but the overall effect should be improved speed.
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>
* efficient calling test dataloader
efficient calling test dataloader
* efficient calling test dataloader during training
efficient calling test dataloader during training
* Update test.py
* Update train.py
* Update train.py
* changed the criteria for the best weight file
changed the criteria for the best weight file from loss to mAP
I trained the model on my custom dataset. But I failed to get a good results when I load the weight file that has the lowest loss on test dataset.
I thought that the loss used in YOLO is not proper criteria for detection performance. So I changed the criteria from loss to mAP.
what do you think of this?
* Update train.py
Yonghye Kwon
