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 448x448. give the result as follows

	Original (darknet)	Original (darknet)	abeardear/pytorch-YOLO-v1	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)
ARCH	YOLOv1	FastYOLOv1	ResNet_YOLOv1	YOLOv1(S=14)	FastYOLOv1(S=14)	YOLOv1	FastYOLOv1
VOC AP[IoU=0.50]	63.4	52.7	66.5	71.71	60.38	66.85	52.89

After this update, the implementation of zjykzj/YOLOv1 has completely surpassed the training results of the paper

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)	Original (darknet)	abeardear/pytorch-YOLO-v1	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)
ARCH	YOLOv1	FastYOLOv1	ResNet_YOLOv1	YOLOv1(S=14)	FastYOLOv1(S=14)	YOLOv1	FastYOLOv1
VOC AP[IoU=0.50]	63.4	52.7	66.5	66.06	54.81	63.06	49.89

SURPASS

1. Adding the use of IGNORE_THRESH in YOLOv1Loss.
2. Reset lambda_ *, refer to YOLOv1 paper settings.

In this version of the implementation, the training results on the VOC dataset have exceeded the YOLOv1 paper implementation.

	Original (darknet)	Original (darknet)	abeardear/pytorch-YOLO-v1	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)
arch	YOLOv1	FastYOLOv1	ResNet_YOLOv1	YOLOv1	FastYOLOv1
train	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval
val	VOC2007 Test	VOC2007 Test	VOC2007 Test	VOC2007 Test	VOC2007 Test
VOC AP[IoU=0.50]	63.4	52.7	66.5	67.21	54.93
conf_thre	/	/	0.5	0.005	0.005
nms_thre	/	/	/	0.45	0.45
input_size	448	448	448	448	448

OPTIMIZE

1. Cancel the last reduction of receptive field to optimize YOLOv1 benchmark network.
2. Use F.cross_entropy instead of F.mse_loss for pred_probs training.

	Original (darknet)	Original (darknet)	abeardear/pytorch-YOLO-v1	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)
arch	YOLOv1	FastYOLOv1	ResNet_YOLOv1	YOLOv1	FastYOLOv1
train	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval
val	VOC2007 Test	VOC2007 Test	VOC2007 Test	VOC2007 Test	VOC2007 Test
VOC AP[IoU=0.50]	63.4	52.7	66.5	62.55	50.46
conf_thre	/	/	0.5	0.005	0.005
nms_thre	/	/	/	0.45	0.45
input_size	448	448	448	448	448

UPDATE

Based on zjykzj/YOLOv2, the entire project has been restructured, including training modules, data modules, and inference modules. In addition, the loss function YOLOv1Loss is reimplemented, which is completely defined by reference to YOLOv1 paper.

	Original (darknet)	Original (darknet)	abeardear/pytorch-YOLO-v1	zjykzj/YOLOv1(This)	zjykzj/YOLOv1(This)
arch	YOLOv1	FastYOLOv1	ResNet_YOLOv1	YOLOv1	FastYOLOv1
train	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval	VOC07+12 trainval
val	VOC2007 Test	VOC2007 Test	VOC2007 Test	VOC2007 Test	VOC2007 Test
VOC AP[IoU=0.50]	63.4	52.7	66.5	52.95	43.65
conf_thre	/	/	0.5	0.005	0.005
nms_thre	/	/	/	0.45	0.45
input_size	448	448	448	448	448

INIT

In the initial version, the basic training and loss function are implemented with reference to aberdear/pythoch-YOLO-v1. In addition, YOLOv1 and FastYOLOv1 model definitions are implemented based on ChatGPT.

At present, there is no effective evaluation code implemented, and only the loss is output during the training process. From the results, the effect is not ideal.