Skip to content
Alexey Kamenev edited this page Sep 7, 2017 · 1 revision

Neural network scripts

Our TrailNet DNN is based on the popular ResNet-18 architecture with a few modifications (refer to the paper for more details). The caffe_srelu_resnet.py script in ResNet directory generates a prototxt file which is a description of the DNN architecture that is used at the first stage of training using the original Trails dataset. At the second stage of training, the network is modified and a second "head" is added which is then trained on the translation dataset. Only the last layer of this second head is trained; all other layers (convolutional etc) are frozen. The srelu-resnet-dragonnet.prototxt file contains a network architecture for the second step. During training, both networks use custom data augmentation techniques as well as a custom loss function. This functionality is implemented as Caffe Python layers in the python-layers.py file. This file needs to be specified in DIGITS when launching a job.