neuromap.network¶

`neuromap.network` ¶

High-level Network builder for Neuromap neuromorphic chips.

A :class:Network is the central object in the Neuromap SDK. It wraps a :class:~neuromap._internal.dynamic_snn.DynamicSNN model with the :class:~neuromap.chip.ChipSpec that describes the target hardware, providing convenience methods for inference, streaming, persistence and introspection.

Example::

from neuromap import Network, chips

net = Network(chips.NEUROSOC_V1)
print(net.summary())

`Network` ¶

Chip-aware SNN wrapper.

A :class:Network owns a :class:DynamicSNN model and carries the :class:ChipSpec (or equivalent topology metadata) used to build it.

Parameters:

Name	Type	Description	Default
`chip`	`ChipSpec`	The chip specification whose topology and neuron parameters will be used to construct the underlying SNN.	required
`use_decoder`	`bool`	Append a trainable linear decoder on the output layer. Useful for continuous regression tasks.	`True`
`membrane_readout`	`bool`	If `True`, the last SNN layer outputs continuous pre-spike membrane potentials instead of binary spikes. Essential for regression tasks (e.g. denoising) where the network must produce continuous-valued output.	`False`

`chip` `property` ¶

The chip specification backing this network.

`model` `property` ¶

The underlying PyTorch :class:DynamicSNN module.

`layers` `property` ¶

Layer sizes as a list.

`input_size` `property` ¶

Number of input features.

`output_size` `property` ¶

Number of output features.

`device` `property` ¶

Device of the first model parameter (or cpu if empty).

`from_topology(layers, *, weight_bits=4, neuron_params=None, name='custom', use_decoder=True, membrane_readout=False)` `classmethod` ¶

Build a :class:Network from an explicit layer topology.

This is the escape-hatch for simulation-only experiments that are not constrained to a specific physical chip.

Parameters:

Name	Type	Description	Default
`layers`	`list[int] \| tuple[int, ...]`	Sequence of layer sizes (e.g. `[32, 16, 8]`).	required
`weight_bits`	`int`	Bit-width for quantization metadata.	`4`
`neuron_params`	`Mapping[str, Any] \| None`	Optional dict of LIF neuron parameters.	`None`
`name`	`str`	Name for the synthetic chip spec.	`'custom'`
`use_decoder`	`bool`	Whether to include the linear decoder.	`True`
`membrane_readout`	`bool`	Use membrane potential readout on the last layer.	`False`

Returns:

Type	Description
`Network`	A new :class:`Network` instance.

`to(device)` ¶

Move the model to device and return self.

Parameters:

Name	Type	Description	Default
`device`	`str \| device`	Target device (e.g. `"cpu"` or `"cuda"`).	required

Returns:

Name	Type	Description
`This`	`Network`	class:`Network` instance for chaining.

`infer(x, *, state=None, return_state=False)` ¶

Run rate-coded inference.

Parameters:

Name	Type	Description	Default
`x`	`Tensor`	Input tensor `(batch, time_steps, input_size)`.	required
`state`	`Mapping[str, Mapping[str, Any]] \| None`	Optional carry-over state for streaming.	`None`
`return_state`	`bool`	If `True`, also return the next state dict.	`False`

Returns:

Type	Description
`Tensor \| tuple[Tensor, dict[str, Any]]`	Output tensor `(batch, output_size)`, or a tuple
`Tensor \| tuple[Tensor, dict[str, Any]]`	`(output, next_state)` when return_state is set.

`infer_sequence(x)` ¶

Run per-frame sequence inference.

Parameters:

Name	Type	Description	Default
`x`	`Tensor`	Input tensor `(batch, frames, input_size)`.	required

Returns:

Type	Description
`Tensor`	Output tensor `(batch, frames, output_size)`.

`infer_stream(x_chunk, state=None)` ¶

Stateful streaming inference - process one chunk at a time.

Parameters:

Name	Type	Description	Default
`x_chunk`	`Tensor`	Input chunk `(batch, time_steps, input_size)`.	required
`state`	`Mapping[str, Mapping[str, Any]] \| None`	State from the previous call, or `None` for the first chunk.	`None`

Returns:

Type	Description
`tuple[Tensor, dict[str, Any]]`	Tuple of `(output, next_state)`.

`summary()` ¶

Return a human-readable summary of the network.

`save(path)` ¶

Save the network (model weights + chip spec) to a .pt file.

Parameters:

Name	Type	Description	Default
`path`	`str \| Path`	Destination file path.	required

`load(path, *, device='cpu')` `classmethod` ¶

Load a network previously saved with :meth:save.

Parameters:

Name	Type	Description	Default
`path`	`str \| Path`	Path to the `.pt` checkpoint.	required
`device`	`str \| device`	Device to load the model onto.	`'cpu'`

Returns:

Type	Description
`Network`	A restored :class:`Network` instance.

neuromap.network¶