Trying train neural network for image classifycation. When using LeNet model with 24x24 image size all work perfect. But when i select AlexNet or GoogLeNet models with 256x256 images poblems comming out. Network training seems to be normal, with no errors. But when i try to clacify image after training, the results are NAN%. In visualization and statistic all conv layer images are blue and all weights zero or NAN.
During training, what do the loss functions look like?
Did you modify any of the image settins from default, for example the “subtract average image” setting?
All settitng are default, only change batch size to 1 because my 1 GB GTX650 have not enough memory even for batch size 2.
And results are this:
This has been here for a long time with no answer. But I had the same question.
Try this: Set the “Crop Size” under “Data Transformations.”
I used 256 which matched the cropping dimensions that I used when I created the database.