ObjectDetection-OneStageDet

ObjectDetection-OneStageDet copied to clipboard

YOLOLoss is slow

1

[https://github.com/TencentYoutuResearch/ObjectDetection-OneStageDet/blob/master/yolo/vedanet/loss/_yololoss.py#L211](url) yolo loss is very slow, how can we speed up the calculating?

guanfuchen

as title or will you consider making this project can be run for pytorch 1.1 version?

lunasdejavu

Hi, thanks for your great code implementation, but I have some questions about the data processing in the test. The resize operation in 'Letterbox' of '_preprocess.py' seems to be slightly...

DrugRui

feature to support snap resume

1

to save the epoch num and backup weights for resuming the training

guanfuchen

HI，when i train my own dataset i got Class always 0?can you tell me

1

AVG IOU 0.4056, Class 0.0000, Obj 0.3534, No obj 0.0001, .5R 0.4000, .75R 0.0000, Cur obj 5, All obj 6, Coord xy 0.1274, Coord wh 0.6270 2019-06-09 18:58:00,455:INFO:Head 1 AVG...

shenghsiaowong

Any Plan for focal loss?

1

It is a great work for yolo training, I am wondering whether it has the plan for focal loss, or anyone wants to develop together? here is the paper [Focal...

chenguanfu511

python examples/test.py 有没有方法从命令行代码开始调试呢 谢谢！

renlicheng

在_voc_train.py中有如下的代码： ` rf = data.transform.RandomFlip(flip) rc = data.transform.RandomCropLetterbox(self, jitter) hsv = data.transform.HSVShift(hue, sat, val) ` 其中 ` rc = data.transform.RandomCropLetterbox(self, jitter) ` 在_preprocessing.py 中有RandomCropLetterbox的定义。这个类具体的实现逻辑是什么？是将原来lightnet的 RandomCrop和Letterbox相互结合吗？ 下面这两行代码具体是什么任务呢？ ``` new_ar = float(orig_w...

pzxdd

1210264601

Why use the anchors clustered on Coco dataset to train VOC?

1

Thanks for your excellent work. But a little question here. The default anchor sizes are clustered on Coco dataset according to the paper. Do they also work well for the...

apxlwl