Deeplearning4j - Convolution Layer

Hi all,

I have some questions regarding the Convolution Layer Node of the deeplearning4j integration.


First of all, I don't understand why the kernel size in the options is two dimensional. Is the node recognizing if the images are color images (3 dim) or gray ones (1 dim) ?

I also don't understand why in the AlexNet metanode the first layer has a kernel size of 8x8? Since they using 227x227 images, the kernel size mentioned in the paper ( is 11x11.

(W−F+2P)/S+1 => (227 - 11)/4 + 1 = 55

(227 - 8)/4 + 1 = 55,75 and not an int ?!

Also regarding the stride mentioned in the paper is 4. The stride in the node is a two dim 2x2. How do they compare ?



Hi Flo,

for convolutional networks the node will know about the dimensionality of the images because you need to specify the image size in the learner node dialog. In DL4J the kernel is always 2-dim. Now if the image has more than one channel this will be handled by the image size option.

You are right, the AlexNet metanode does not have exactly the same configuration as in the paper. We might want to update that in the future.

Regarding your calculations: if the division does not result in an integer than this will be an border case which is handled by the DL4J library internally. As you can use a lot of different image sizes for AlexNet (not only the exact same size from the paper) these border cases might happen, however, they shouldn't be a problem.

Regarding the stride: 2x2 just means stride in x and y direction. Again, you can use a different stride for x and y. The 4 in the paper just means stride x is 4 and stride y is 4.

I hope that helps.



Thanks David that clarified a lot for me :-)


Since the last update I have a problem feeding the learner node with images.

The old node (in the public examples) gives me the error message: "Execute failed: No converter for data cell of type Image available"

The new one is telling me that there are no input colums selected. Images are of float type and normalized to 0…1 

Did you update to the latest version of Deeplearning4J Integration - Image Processing? Then it should work again. Currently, that is not really clear in the examples. We will update the workflows shortly.

Ah my fault. It's working fantastic now :-)  Thanks a lot.

This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.