I have six image layers with 32-bit float. I created a CNN model using the following parameters:

• Sample patch size: 32
• Number of image layers: 6
• Model classes: Class-1, Class2
• Number of hidden layers: 3
• Use batch normalization: Yes
• Hidden layer 1: Kernel size 7, Number of feature maps 20, Max pooling No
• Hidden layer 2: Kernel size 3, Number of feature maps 12, Max pooling No
• Hidden layer31: Kernel size 3, Number of feature maps 12, Max pooling No

Model training was successful. When I executed the CNN model, though, I've got an error message "Invalid argument: input depth must be evenly divisible by filter depth: 5 vs 6 [[{{node BiasAdd_8}}]]"

This very same parameters worked alright when I used a 4-band, 8-bit unsigned integer image. Has anyone experienced with this error message and came up with a fix?