Deep Learning for Computer Vision with Caffe and cuDNN

Originally published at:

Deep learning models are making great strides in research papers and industrial deployments alike, but it’s helpful to have a guide and toolkit to join this frontier. This post serves to orient researchers, engineers, and machine learning practitioners on how to incorporate deep learning into their own work. This orientation pairs an introduction to model structure…


Is this work with Jetson TK1

Caffe runs on the Jetson TK1. See this blog post guide by Pete Warden:

What is the difference between the first set of grayscale looking gabors and the second set of color looking gabors in Figure 3? Is the color supposed to represent sensitivity of those filters for those color difference? Does the same apply to higher visualization of the network, or does color in higher layers encode orientation sensitivity?

The feature extraction and visualization link appears to be dead

Figure 3 shows the filter weights for the first layer so these filters are interpretable as image patches representing what the filter responds to most whether oriented edges, colors, or so on. In this model certain layers' filters are learned in two separate groups and as it happens this makes the first layer learn sets of grayscale edge features and color features. The higher layers are not so interpretable due to the nonlinearity of the network but research continues. See for a summary of some approaches.

Fixed now. Thanks!

how I can know the kernels formula or the kernels that computed for the trained vesion.

Lets say I am just interested in trained filtering kernels, is there anyways that I can have them?