Caffe lets you define custom layers with Python. Python layers are obviously much slower than their C++ counterparts, but are easy to write and test quickly since using them doesn't require recompilation. DIGITS moves some files around for you so that you can specify different Python layers for each network.
The BVLC and DIGITS documentation online has detailed information and tutorials about how to use these layers. Some of that information has been re-created here for your convenience.
First, you'll need to create a Python file which defines one or more layers. As an example, here is a layer which re-creates the functionality of the standard Accuracy layer using JSON-encoded parameters:
import caffe
import json
class AccuracyLayer(caffe.Layer):
def setup(self, bottom, top):
if hasattr(self, 'param_str') and self.param_str:
params = json.loads(self.param_str)
else:
params = {}
self.top_k = params.get('top_k', 1)
def reshape(self, bottom, top):
top[0].reshape(1)
def forward(self, bottom, top):
# Renaming for clarity
predictions = bottom[0].data
ground_truth = bottom[1].data
num_correct = 0.0
# NumPy magic - get top K predictions for each datum
top_predictions = (-predictions).argsort()[:, :self.top_k]
for batch_index, predictions in enumerate(top_predictions):
if ground_truth[batch_index] in predictions:
num_correct += 1
# Accuracy is averaged over the batch
top[0].data[0] = num_correct / len(ground_truth)
def backward(self, top, propagate_down, bottom):
pass
NOTE: The uploaded file will be renamed to digits_python_layers.py, discarding the original filename.
To add the layer to your network, include this prototxt snippet in your custom network definition:
layer {
name: "accuracy"
type: "Python"
bottom: "pred" # These blob names may differ in your network
bottom: "label"
top: "accuracy"
python_param {
module: "digits_python_layers" # File name
layer: "AccuracyLayer" # Class name
param_str: "{\"top_k\": 1}"
}
include { stage: "val" }
}