a problem meet during training my own data

[jiawensong]

a problem meet during training my own data

[HiKapok]

@jiawensong try to filter images those have zero bbox, by the way, can you train PASCAL VOC successfully?

[franklu323]

I think we met the same problem. I use my own dataset which followed by the VOC format, and convert was fine (I think, cause I ran 'inspect' after, and train samples, val samples it matches). But when I train the model, the error occurs. I also received this 'InvalidArgumentError (see above for traceback): flat indices[30, :] = [30, -1] does not index into param (shape: [32,6792]).' If you solve this, could you please help me.

[HiKapok]

@franklu323 did you change your batch_size? Hope you can check if there is images that have no bbox with them, And also thanks for you issue report.

[jiawensong]

@HiKapok i can train ,test,eval,pascal successfully. I have checked all my own data, the minimum size is 3kb, contains more than one boundingbox. i wonder if the problem is that i haven't create negative boundingbox in my own data.

[franklu323]

@HiKapok I do change my batch size to 4, cause my GPU is GTX1060 and I also change the number of classes to 5 (cause I only have 5 classes). BTW, all my dataset contains object
And I saw #12 , I also change scafford to None, but it seems got same error

[HiKapok]

@franklu323 did you also used tensorflow1.4?

[HiKapok]

@jiawensong what did you mean by 'haven't create negative boundingbox', did you have one class for background?

[jiawensong]

@HiKapok i have just create two class, number and chinese character, i haven't create class for background

[HiKapok]

@jiawensong @franklu323 both of your problems should caused by this line of code when there is image's 'batch_n_neg_select' equals 0, this only can be caused by the zero positives, so I guess there maybe some image have no bbox with them

[HiKapok]

@jiawensong background class is needed in any detection task

[jiawensong]

@HiKapok oh, maybe the problem is that i set number class with 0 label, then the program take number class as negative example, when some xml files without chinese character taken as input , the number of positive example is set to 0 and batch_n_neg_select equals 0.

[franklu323]

@HiKapok So do I need to label the background or just in data_common.py, set background as 0 and the rest of them are my own data classes??

[HiKapok]

@franklu323 @jiawensong you needn't label background class, just follow the original pascal dataset's setting, leave one placeholder of value '0' for background class, thanks for your issue report