{# Copyright (c) 2015, NVIDIA CORPORATION. All rights reserved. #}
Specifying a custom network
You should enter a train_val network here. Some cleanup is done for you automatically:
The num_output for each InnerProduct layer which is a network output gets set to the number of labels in the chosen dataset.
Any Accuracy layers with top_k >= num_labels is removed automatically.
The deploy network will be created for you automatically. Once caffe implements their "all-in-one" nets feature, you will be able to specify which layers to include in the deploy file. Until then, these things are done automatically:
The last SoftmaxWithLoss layer is replaced with a Softmax layer.
All other loss layers and accuracy layers are removed.