[Issue] Incompatible shapes during training of own dataset

Nobody has claimed this yet.

Hello again,

I successfully managed to use my dataset but the problem apparently is with the shapes inbetween layers...

I get the following error:

InvalidArgumentError (see above for traceback): Incompatible shapes: [8,28,28] vs. [8,512,512]
[[Node: mrcnn_mask_loss/logistic_loss/mul = Mul[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"](mrcnn_mask_loss/Log, mrcnn_mask_loss/Gather_1)]]

I did change the batch size to see if there were any changes, and the shapes switch to [8,28,28] to [6,28,28] when I change the batch size from 16 to 8.

Here are screens of the configuration and the error (partly):


I resized the images to be (512,512) as well. The problem seems to come from the "mul" gradient... But that's all I can even understand...

Thanks for your help.

No source file, test, or entry point is named. Start by reproducing training on the reported custom dataset with the shown configuration, then trace the mask-loss inputs associated with mrcnn_mask_loss/logistic_loss/mul. Done means identifying the source of the incompatible [8,28,28] and [8,512,512] shapes and confirming training proceeds with matching mask dimensions.

Written by the indexing model from the issue text.