Training on own dataset

[ammaradel]

Is it possible to train the network on my own dataset?

[inFreedom92]

Yes, I think so. If you could make dataset and dataloader, which outputs same format as VOC or COCO in this repository then it might work.

[nigyiii]

just organize your dataset in VOC style or COCO style, then change num_classes in train.py and dataset path in voc0712.py / coco.py

[lgl603]

How to set the value of num_classes? Is the num_classes equals the real classes of objects, or num_classes = "real classes of object +1"?

[inFreedom92]

@lgl603

How to set the value of num_classes? Is the num_classes equals the real classes of objects, or num_classes = "real classes of object +1"?

it's "real classes of object +1" as I know

[alen-mask]

@lgl603

How to set the value of num_classes? Is the num_classes equals the real classes of objects, or num_classes = "real classes of object +1"?

it's "real classes of object +1" as I know

hi @inFreedom92 ,

what about class names and thresholds_for_classes? i guess class name is just a label for model, but the thresholds is totally different for different objects and image sources.

[inFreedom92]

@alen-mask Do you mean class name something like this? VOC_CLASSES = ( # always index 0 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor') I made my own CUSTOM_CLASSES but I don't use it now. Because I use my own transform... Last, I didn't change anything about thresholds in training. I don't know how it works... I just use thresholds in demo.py to see how the model inference.

I'm sorry that I'm not helpful.

[alen-mask]

@alen-mask Do you mean class name something like this? VOC_CLASSES = ( # always index 0 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor') I made my own CUSTOM_CLASSES but I don't use it now. Because I use my own transform... Last, I didn't change anything about thresholds in training. I don't know how it works... I just use thresholds in demo.py to see how the model inference.

I'm sorry that I'm not helpful.

Thanks very much for your reply. According previous mrcnn finetuning experience, the thresholds does not need to be changed. Maybe the model is not trained enough yet .

[lgl603]

@lgl603

How to set the value of num_classes? Is the num_classes equals the real classes of objects, or num_classes = "real classes of object +1"?

it's "real classes of object +1" as I know

The num_classes in "datasets/coco.py" is 80 for COCO datasets, so it should be "real calsses of object". However, the problem "no boxes to NMS" occurs when inferencing my own trained model.

[lgl603]

@lgl603

How to set the value of num_classes? Is the num_classes equals the real classes of objects, or num_classes = "real classes of object +1"?

it's "real classes of object +1" as I know

hi @inFreedom92 ,

what about class names and thresholds_for_classes? i guess class name is just a label for model, but the thresholds is totally different for different objects and image sources.

Thanks very much for your reply! The num_classes in "datasets/coco.py" is 80 for COCO datasets, so it should be "real calsses of object". However, the problem "no boxes to NMS" occurs when inferencing my own trained model.

[wanglaotou]

i trained my own dataset like voc datasets, but after trained 65 epochs its map is 0.0 and no boxes to nms, have you met this problem?

[CraigWang1]

@wanglaotou I have also trained on my own dataset, and it also says that the map is 0.0 and there are no boxes to nms. Currently looking for a solution

[yushi-okuinishi]

@wanglaotou @CraigWang1 I also got that the map is 0.0 XD

[CraigWang1]

@yushi-okuinishi Lol I solved it by switching over to the signatrix efficientdet implementation that you can also find on Github. (no disrespect or offense intended to toandaominh, it is extremely difficult to code a huge project like this)

[yushi-okuinishi]

@CraigWang1 Thank you. I try it

[NAcvat]

The signatrix efficientdet implementation worked for you?

[CraigWang1]

@NAcvat Yes for effiicientdet-d0, although efficientdet by itself is an unproven toddler compared to the proven models like Yolov3, so I'm iffy on using it.