thomelane commented on issue #9034: Understanding the correlation layer URL: https://github.com/apache/incubator-mxnet/issues/9034#issuecomment-404578327 `kernel_size` must be odd number, otherwise strange things happen! It's mentioned in the docs string, but I think it would be better for this to be added as strict validation to avoid confusion like this in the future. With regards to the stride (once a odd `kernel_size` is used), `stride1` relates to the striding on the `data1` feature map 1 and `stride2` relates to the striding within the 'neighborhood' defined on feature map 2. Without a neighborhood this method would compare all patches from feature map 1 with all patches from feature map 2. With large feature maps, the number of comparisons becomes too large, so we restrict to a region in feature map 2 centered around the corresponding location of the patch in feature map 1. And `stride2` strides within this neighborhood region. I recommend taking a look the ['FlowNet: Learning Optical Flow with Convolutional Networks'](https://arxiv.org/pdf/1504.06852.pdf ) paper that defined this type of layer.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
