Yolov3: Did anybody train voc with this code?

[Brizel]

I've trained coco with this code and the result is impressive. So I want to try voc on this code . I made train_list.txt as : cls_name x_center y_center width height. But the result is not as good as I prospected. Could anybody who successfully trained voc give me some suggestion?

[nirbenz]

If you successfully trained COCO with this code I'd love to hear more about how you did it, what parameters did you use and what (if any) changes you made for this to work. And obviously - did you get the same mAP as the original paper? :-)

[Brizel]

Actually I haven't run mAP yet. I just put some pics into SAMPLE and run detect.py. The result seems fine.

[zhaoyang10]

@nirbenz Have you trained COCO with this code? I trained several other data sets from scratch, all of them got low recall, as like around 0.6-0.7.

[tianqiaotiemo]

I‘ve trained voc with this code，and the mAP is around 0.5-0.6。However，I haven't get the same mAP as original paper by training COCO!

[ming71]

I‘ve trained voc with this code，and the mAP is around 0.5-0.6。However，I haven't get the same mAP as original paper by training COCO!

How can you train your dataset on this repo? I wonder wether it impossible to train without flag --report?As is setted in model.py , the p_boxes will always be None without flag --report!!!! Thus no parameters were delivered into build_targets(),and no prediction generated such as tx,ty,th,tw,etc.

[Peterisfar]

I‘ve trained voc with this code，and the mAP is around 0.5-0.6。However，I haven't get the same mAP as original paper by training COCO!

i also make a similiar result!
can you describe more detailed？ cfg is what? how about anchor size? how about hyp in loss function? test set and train set are what?

[ruoruo6]

I‘ve trained voc with this code，and the mAP is around 0.5-0.6。However，I haven't get the same mAP as original paper by training COCO!

i also make a similiar result!
can you describe more detailed？ cfg is what? how about anchor size? how about hyp in loss function? test set and train set are what?

excuse me. i've trained voc with batchsize equals 2 and epoches equals 88 but the map is only around 0.4. Why ? batchsize and epoches are too small? Hope reply soon.

[Peterisfar]

I‘ve trained voc with this code，and the mAP is around 0.5-0.6。However，I haven't get the same mAP as original paper by training COCO!

i also make a similiar result!
can you describe more detailed？ cfg is what? how about anchor size? how about hyp in loss function? test set and train set are what?

excuse me. i've trained voc with batchsize equals 2 and epoches equals 88 but the map is only around 0.4. Why ? batchsize and epoches are too small? Hope reply soon.

batchsize is 2? do you use how many GPUs, epochs about 50 the mAP is stable

[ruoruo6]

1050ti GPU. When the epoch equals 50 the map is around 0.4. Is this normal?

------------------ 原始邮件 ------------------
发件人: "leon"[e-mail];
发送时间: 2019年5月23日(星期四) 晚上8:11
收件人: "ultralytics/yolov3"[e-mail];
抄送: "郁蔚"[e-mail];"Comment"[e-mail];
主题: Re: [ultralytics/yolov3] Did anybody train voc with this code? (#28)

I‘ve trained voc with this code，and the mAP is around 0.5-0.6。However，I haven't get the same mAP as original paper by training COCO!

i also make a similiar result!
can you describe more detailed？ cfg is what? how about anchor size? how about hyp in loss function? test set and train set are what?

excuse me. i've trained voc with batchsize equals 2 and epoches equals 88 but the map is only around 0.4. Why ? batchsize and epoches are too small? Hope reply soon.

batchsize is 2? do you use how many GPUs, epochs about 50 the mAP is stable

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub, or mute the thread.

[glenn-jocher]

Closing for inactivity.