Overview: How deepnet works

A deepnet model can be trained by creating the following three files and passing them as arguments to deepnet/trainer.py

1. Model : Specifies the model's structure and hyperparameters.
2. Trainer : Specifies the training algorithm, stopping conditions, data and output directories.
3. Evaluation : Specifies the evaluation method (to be used while training to see how well things are going).

For example,

cd deepnet
python trainer.py examples/ff/model.pbtxt examples/ff/train.pbtxt examples/ff/eval.pbtxt

These three files are written as human-readable protocol buffers (pbtxt). Protocol Buffers are a way of encoding structured data that is easy to extend and fast to (de)serialize. Both these features are important for doing experiments with deep learning models.

The result of training is written out periodically to disk as a serialized protocol buffer. This output file contains the model parameters - weights and biases for each layer.