How to prepare training data?

[samehraban]

Could you please elaborate on how to prepare training data? What are the contents of different directories you mentioned in the readme?

[samehraban]

I converted wider dataset to pascal format, then I had run pascalvoc_to_tfrecords.py. When I run train_model.py I get this error:

ValueError: Dimensions must be equal, but are 520400 and 136400 for 'ssd_losses/or1' (op: 'LogicalOr') with input shapes: [520400], [136400].

[samehraban]

I had set batch_size to 16 and setting it back to 1 solved the problem. But how can I train with greater batch sizes?

[EricZgw]

@samehraban I forgot to associate some parameters with batchsize, and I will modify it later. You can change the code here first. 25600->25600 * 16( your batchsize ) , 6400->6400 * 16( your batchsize ).The error should be caused by this. I hope you can get a good training result.

[samehraban]

@EricZgw Thank you. I also tried to modify image size and got error on that too. Does it make sense to decrease image size and still use current vgg16 checkpoint?

[EricZgw]

@samehraban The image size in the original paper is 640*640. The pretrained vgg16 checkpoint may not be very useful, I just think it is multi-scale training in a sense. You can try to train from scratch. And if you change the input size of the image in training, remember to also change the size of anchors.

[wwqstevenash]

I change the code with your advise (nets/ssd.py: 25600->25600 * 16( your batchsize ) , 6400->6400 * 16( your batchsize )), but I get the incorrect training results. Is there anything else about batchsize 16 modification?

[JuheonYi]

@samehraban Hi, did you manage to prepare the training data successfully? Could you share how you managed it?

Thanks!