Examples ======== Runnable example scripts live in the `examples `__ directory. The top-level scripts are grouped by filename prefix: - ``0xx-`` for spiking neural network examples - ``1xx-`` for rate-based recurrent neural network examples The directory also includes two tutorial-style example series: - ``examples/pp_prop/`` for ``braintrace.pp_prop`` / ``ES_D_RTRL`` examples - ``examples/drtrl/`` for ``braintrace.D_RTRL`` examples Most scripts can be run directly. For example: .. code-block:: bash python examples/001-gif-snn-for-dms.py Spiking Neural Networks ----------------------- - **000-lif-snn-for-nmnist.py** - Trains a recurrent LIF-delta SNN on the N-MNIST dataset, using framed event inputs from ``tonic`` and the shared ``OnlineTrainer`` implementation based on ``IODimVjpAlgorithm``. - **001-gif-snn-for-dms.py** - Trains a GIF recurrent SNN on a delayed matching-to-sample task generated by ``DMSDataset``, using the shared ``OnlineTrainer`` based on ``IODimVjpAlgorithm`` for online learning. - **002-coba-ei-rsnn.py** - Trains an excitatory/inhibitory recurrent SNN on an evidence-accumulation task, with configurable CUBA or COBA synapses and online learning via ``ES_D_RTRL`` or ``D_RTRL``. - **003-snn-memory-and-speed-evaluation-all.py** - Runs a synthetic benchmark over multiple sequence lengths and compares BPTT, ``ES_D_RTRL``, and ``D_RTRL`` in terms of training time and memory usage. - **003-snn-memory-and-speed-evaluation-batched.py** - Trains and benchmarks an SNN on the DVSGesture dataset with the online algorithms running in ``brainstate.mixin.Batching()`` mode, reporting accuracy, speed, and memory. - **003-snn-memory-and-speed-evaluation-vmap.py** - Trains and benchmarks the same DVSGesture setup while keeping per-sample states under ``vmap``, reporting accuracy, speed, and memory. - **004-feedforward-conv-snn.py** - Builds a feedforward convolutional SNN with ``braintrace.nn.Conv2d``, ``LayerNorm``, IF neurons, and ``LeakyRateReadout``. By default it trains on N-MNIST, and the script also includes SHD data loading plus both online and BPTT trainers. Rate-Based Recurrent Neural Networks ------------------------------------ - **100-gru-on-copying-task.py** - Trains a GRU on the copying task and compares online learning with ``ParamDimVjpAlgorithm`` against BPTT. - **101-integrator-rnn.py** - Trains a ``MiniGRU``-based RNN to integrate a white-noise input signal and compares online ``D_RTRL`` with BPTT while plotting predictions and training losses. Tutorial Series --------------- - **pp_prop/** - A tutorial-linear example suite for ``braintrace.pp_prop`` (``ES_D_RTRL`` / ``IODimVjpAlgorithm``). The series covers neuron-model variants, batching with ``vmap`` versus batched primitives, single-step versus multi-step VJPs, sparse/LoRA/convolutional operators, Poisson-MNIST classification, and ``decay_or_rank`` sweeps. See ``examples/pp_prop/README.md`` and ``docs/tutorials/pp_prop.md``. - **drtrl/** - A tutorial-linear example suite for ``braintrace.D_RTRL``. The series covers integrator basics, batching modes, VJP variants, LoRA and convolutional operators, row-scan MNIST classification, toy character-level language modeling, and the ``fast_solve`` plus ``normalize_matrix_spectrum`` knobs. See ``examples/drtrl/README.md`` and ``docs/tutorials/drtrl.md``. Supporting Files ---------------- - **README.md** - Brief overview of the examples directory and its extra Python dependencies. - **snn_models.py** - Shared SNN model and trainer definitions used by the ``000`` and ``001`` examples. - **utils.py** - Shared command-line parser utilities used by the ``003-*`` benchmark scripts.