xuzheyuan624 / yolov3-pytorch Goto Github PK

我对你yolov3代码里的loss部分的边框回归稍做了修改，根据ground truth的大小对权重系数进行修正：self.lambda_xy * (2 -ghgw)，以及self.lambda_wh (2-gh*gw)这样对于尺度 较小的boxes其权重系数会更大一些，为了进一步放大这个权重，还乘了1.5作为系数。大神请看看这样改进有没有毛病

thanks for your work, like the title, is there any improvements?

Heya @xuzheyuan624 -continuing our discussion from the other repository -
I have extended your code to support varying aspect ratios and letterboxing. Doing this I was able to get an mAP of:

DONE (t=5.06s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.324
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.572
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.331
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.190
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.352
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.421
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.277
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.414
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.432
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.276
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.458
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.547

Which is almost the same as the original Darknet implementation of YOLOv3 - and equivalent to ultralytics's implementation.

I switched to testing your code due to what seems to be working training code + multi GPU support. I will update if/when I get something converged (I've got 9 1080ti's for this purpose).

Thanks for this repository. I'm in the process of trying to adapt it to my own dataset and needs. I was wondering if the weights file supplied was converted from the Darknet-trained weights file or was achieved by training using this code?