Integrate yolov5, refactor YOLOv2/YOLOv2-Fast and YOLOv2Loss

1. Fully integrated with YOLOv5 v7.0, including training, evaluation, and prediction;
2. Referring to model and loss function implementation, we have reimplemented YOLOv2/YOLOv2-Fast and YOLOv2Loss;
3. We trained the corresponding classification model based on the YOLOv2 configuration file. Please refer to the specific results in docs/

Update ultralytics/yolov5 Transforms

1. update ultralytics/yolov5(485da42) transforms, add Mosaic/Perspective and so on.

• Train using the VOC07+12 trainval dataset and test using the VOC2007 Test dataset with an input size of 416x416. give the result as follows

	Original (darknet)	tztztztztz/yolov2.pytorch	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)
ARCH	YOLOv2	YOLOv2	YOLOv2+Darknet53	YOLOv2	YOLOv2-tiny
VOC AP[IoU=0.50]	76.8	72.7	74.95/76.33(v0.2.1)	73.27	65.44

• Train using the COCO train2017 dataset and test using the COCO val2017 dataset with an input size of 416x416. give the result as follows (Note: The results of the original paper were evaluated on the COCO test-dev2015 dataset)

	Original (darknet)	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)
ARCH	YOLOv2	YOLOv2+Darknet53	YOLOv2
COCO AP[IoU=0.50:0.95]	21.6	25.86	22.84
COCO AP[IoU=0.50]	44.0	48.40	43.95

From the training results, it can be seen that the pretraining configuration of yolov5 can effectively improve the performance of yolov2. However, it can also be observed that the improvement is not very high, and the performance of YOLOv2+Darknet53 is not as good as the previous results. This indicates that there are different optimal training configurations for different network architectures.

Refactor Data Module

1. Refactor data module;
2. Fix demo.py;
3. Update latest training results.

• Train using the VOC07+12 trainval dataset and test using the VOC2007 Test dataset with an input size of 448x448. give the result as follows

	Original (darknet)	tztztztztz/yolov2.pytorch	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)
ARCH	YOLOv2	YOLOv2	YOLOv2+Darknet53	YOLOv2	YOLOv2-tiny
VOC AP[IoU=0.50]	76.8	72.7	76.33	72.00	64.36

• Train using the COCO train2017 dataset and test using the COCO val2017 dataset with an input size of 416x416. give the result as follows (Note: The results of the original paper were evaluated on the COCO test-dev2015 dataset)

	Original (darknet)	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)
ARCH	YOLOv2	YOLOv2+Darknet53	YOLOv2
COCO AP[IoU=0.50:0.95]	21.6	24.98	22.01
COCO AP[IoU=0.50]	44.0	46.85	42.70

UPGRADE

1. Reconstructed loss function implementation, consistent with the original paper implementation. 2bd6bae
2. Add darknet53 network as a backbone. 2c772aa
3. Fix cocoevaluator. abd5760

• Train using the VOC07+12 trainval dataset and test using the VOC2007 Test dataset with an input size of 416x416. give the result as follows

	Original (darknet)	tztztztztz/yolov2.pytorch	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)
ARCH	YOLOv2	YOLOv2	YOLOv2+Darknet53	YOLOv2	YOLOv2-tiny
VOC AP[IoU=0.50]	76.8	72.7	76.27	71.65	64.19

• Train using the COCO train2017 dataset and test using the COCO val2017 dataset with an input size of 416x416. give the result as follows (Note: The results of the original paper were evaluated on the COCO test-dev2015 dataset)

	Original (darknet)	tztztztztz/yolov2.pytorch	zjykzj/YOLOv2(This)
ARCH	YOLOv2	YOLOv2+Darknet53	YOLOv2
COCO AP[IoU=0.50:0.95]	21.6	25.33	21.96
COCO AP[IoU=0.50]	44.0	47.24	42.65

UPDATE

1. Increase YOLOv2 training results for COCO dataset
2. Update VOC dataset results for YOLOv2 and YOLOv2-tiny
3. IGNORE_THRESH: 0.5/0.7 search #400b0f8
4. YOLOv2 loss function search #129ba72e
5. Comparing YOLOv2 Training of Darknet 224x224 and 448x448 pretrained Models #e4b7399

	Original (darknet)	tztztztztz/yolov2.pytorch	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)
arch	YOLOv2	YOLOv2	YOLOv2	YOLOv2-tiny
train	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval
val	VOC2007 Test	VOC2007 Test	VOC2007 Test	VOC2007 Test
VOC AP[IoU=0.50]	76.8	72.7	70.39	63.96
conf_thre	/	0.005	0.005	0.005
nms_thre	/	0.45	0.45	0.45
input_size	416	416	416	416

v0.1.1

1. Fix target transform error label2yolobox. #029f796
2. Add multi-GPUs training and update yolov2_voc.cfg and yolov2-tiny_voc.cfg training results. #6632949
3. Integrated COCO dataset and evaluator. #5b19877 #e1ba3ba

	Original (darknet)	tztztztztz/yolov2.pytorch	zjykzj/YOLOv2(This)	zjykzj/YOLOv2(This)
arch	YOLOv2	YOLOv2	YOLOv2	YOLOv2-tiny
train	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval
val	VOC2007 Test	VOC2007 Test	VOC2007 Test	VOC2007 Test
VOC AP[IoU=0.50]	76.8	72.7	70.39	63.77
conf_thre	/	0.005	0.005	0.005
nms_thre	/	0.45	0.45	0.45
input_size	416	416	416	416

v0.1.0

1. Realized the YOLOv2 network and YOLOv2 loss function definition;
2. Implemented complete training/evaluation and prediction code;
3. Darknet19 and FastDarknet19 use ImageNet for pretraining, with an input size of 224x224;
4. Train using voc07+12 training data, evaluate based on voc2007 test.

	Original (darknet)	tztztztztz/yolov2.pytorch	zjykzj/YOLOv2(This)
train	VOC2007+2012 trainval	VOC2007+2012 trainval	VOC2007+2012 trainval
val	VOC2007 test	VOC2007 test	VOC2007 test
VOC AP[IoU=0.50], inference	76.8	72.7	69.69
conf_thre	/	0.005	0.005
nms_thre	/	0.45	0.45
input_size	416	416	416