Caffe

Github repo

• Caffe source

Tutorials

• Anatomy of a Caffe Model

Layers

• Layers

Specifying a custom Caffe network

Classification

• The Deploy network must contain a Softmax layer. This should produce the only network output.
• If any InnerProduct layers have left num_output unset, DIGITS will fill it in for you automatically with the number of classes in your dataset.
• If any Accuracy layer has a top_k value that is smaller than the number of classes in your dataset, then that layer is removed.
• The network should include the necessary layers for all computation states - Train, Val and Deploy. When searching for layers, the network will be set to one of these three states:
  • phase: TRAIN stage: "train"
  • phase: TEST stage: "val"
  • phase: TEST stage: "deploy"
• Your layer's include and exclude rules should be set properly to be picked up by the desired computation state[s]. You can look at the standard networks for examples. In addition, since there are many ways to specify the rules, here's a table that you may find useful: