I want to know how to train on my own dataset?I didn't see the code for train

[Jamescallow]

Oh,I'm so

stupid！！！ hahaha,I found it

[ananddb90]

@XingJun940207 Could you help me to find out too how to train for your own data set ?

[ananddb90]

and also could you help me to know the possible reason behind #6 issue. That would be really appreciated. thank you

[RobertCsordas]

Training on your own data that is different from MS COCO is not trivial. You need to write your own data loader. You can use the Dataset/CocoDataset.py as an example/reference. As long as you maintain interface compatibility with the original CocoDataset class, you shuld be fine. Maybe you need some additional hyper paramtere tuning to make the training stable if your data is very different.

[Jamescallow]

Hi,I want to know how many Iterations did you train? I trained on COCO dataset for about 180000 Iterations,but It seems got a bad result.The loss is around 1.5 and the detection result is not as good as yours. Thank you very much.

[RobertCsordas]

It was trained for 2.2M iterations. After 100-200k iters enable training of the base network (trainFrom argument, you can experiment with different values). After 1.5M iterations decrease the learning rate by factor of 10 (1e-5).

[zhangleiedu]

hi, i want to train on voc datasets, but i cannot figure out how to read the xml data and feed to this project, can you give me a demo if possible, thanks!

[qiningonline]

To train your own dataset on this repo:

Training Stage : [1] replace the CocoDataset.py by your own script, to read the images, load the coordinates, and the class label/index [2] Update the total number of categories here [3] Update the main.py, import script in [1]

Inference Stage [1] in test.py, update the category definition

[RobertCsordas]

Thank you qiningonline! I was very busy recently and not able to support the issue.

[obendidi]

@qiningonline or you could just make a script that reformat your data to coco format, I believe it's easier that doing step 1 , it's just parsing the values in a json file following this format (in section 4.1 )

step 2 remain the same of course , and there will be no need for step 3 :)

cheers !

[Qiuchenggit]

After my training(only vehicle ang background), the mAP was just only 0.773. When i training the dataset, the params unchanged, except the categories(set 1) and used the pretrained model(inception_resnet_v2_2016_08_30.ckpt). Can you tell me how can i improve the mAP???

[MaskVulcan]

@XingJun940207 I got the same problem.I change the trainFrom from -1 to 0 after 40000 iterations but my loss is still 1.5. And result is bad. How do you deal with this?

[MaskVulcan]

@xdever I trained for 5 epochs and the loss is always around 1.5(I have changed the trainFrom to 0).Will loss decrease after more epochs?How is your result? Will The loss decrease all the time? Or will loss suddenly decrease at some time?

[audience3]

@qiningonline could you please give me an example how to rewrite a script to replace the CocoDataset.py. Or

[henbucuoshanghai]

how to train my own data?????????????