See the report for an introduction the this project. The master thesis discussing SpiDNN in detail can be found here.
First, in order to compile the binaries for the SpiNNaker toolchain and for SpiDNN, you need to download the GNU ARM Embedded Toolchain and make it available in your path.
Afterwards, download and install the SpiNNaker toolchain with the graph front-end:
git clone https://github.com/SpiNNakerManchester/SupportScripts
bash SupportScripts/install.sh gfe
bash SupportScritps/setup.sh
bash SupportScripts/automatic_make.shOther dependecies that must be installed for SpiDNN:
numpy
pip install numpytensorflow
pip install tensorflowkeras
pip install keras
Afterwards, enter the SpiDNN/ directory and execute source scripts/env.sh. This adds SpiDNN to your python-path variable. Also from the SpiDNN/ directory, execute make, which will build the binaries for SpiNNaker. In the SpiDNN/tests/test_inference.py and SpiDNN/tests/test_training.py files, you will find examples showing how SpiDNN is used.
- custom injectors which support partitions
- pong receive without payload (less network traffic, more complexity in neurons)
- split neurons (bigger layers)
- quantization (speed and memory size)
- populations or some other mechanism to combine neurons (saving cores)
- optimizer interface
- loss into perceptron (saving cores but higher complexity)
- progress interface with epoch loss (loss computed by host?)
- validation interface and more metrics, which could be a nice way for early stopping (then definetly loss computed on host (error still on the board) (see previous bullet point))
- monitor/sync core for each layer (simplifying interface)
- enable multiple successive sessions (currently LiveEventConnection does some weird stuff)
- export/import weights to/from file (fancy HDF5 or json which is less sexy)
- nicer injection of weights into machine_vertices (different interface than just using init_neurons)