Get started
Choose a model
Choose a neuron model for sequence prediction or connected population simulation.
For single-neuron traces, choose GRU or Mamba and load a trained checkpoint. Both predict spike logits and somatic voltage from the same native input channels. For connected populations, choose Lite, whose compact state supports recurrent routing and inference-time adaptation across many neurons.
Model catalog
AxoSim–GRU
Recurrent neuron models for spike and voltage prediction. Choose Small for a compact model or Medium for the report's strongest spike Mean F1.
- Profiles
axosim-gru-smallaxosim-gru-medium- Create
create_axosim_profile- Predict
(batch, time, 2)
Spike logit and soma voltage
AxoSim–Mamba
Selective state-space neuron models with a shared dendritic front end. Choose among three capacities and use the checkpoint's temporal backend.
- Profiles
axosim-mamba-smallaxosim-mamba-mediumaxosim-mamba-large- Create
create_axosim_profile- Predict
(batch, time, 2)
Spike logit and soma voltage- Use it
- Train a named model →
AxoSim–Lite
Compact neuron dynamics for connected populations. Adapt behavior, morphology, and contact efficacies while retaining the temporal gradient.
- Connect
AxoSimPopulation- Predict
- Four causal forecast steps
Explicit recurrent state
Choose a capacity and backend
A model profile selects architecture parameters; a checkpoint supplies fitted weights. Compare candidate checkpoints using the AxoBench metrics relevant to your application: Voltage SERA, spike Mean F1, the subthreshold voltage RMSE, and Response SERA, evaluated as described in evaluate models. The technical report compares fidelity and inference cost across the trained models with the earlier SERA definition and Dynamics SERA.
Architecture recipes
| Profile ID | model_dim |
state_dim |
num_layers |
head_dim |
|---|---|---|---|---|
axosim-gru-small |
32 | 8 | 2 | 32 |
axosim-gru-medium |
128 | 32 | 2 | 128 |
axosim-mamba-small |
32 | 8 | 2 | 32 |
axosim-mamba-medium |
176 | 44 | 4 | 88 |
axosim-mamba-large |
314 | 80 | 20 | 80 |
These are the exact fields in the profile registry. The Large profile also sets mamba_update_normalization="fixed_rms", which normalizes the branch updates of its 20 layers. For GRU, the recurrent hidden width equals model_dim; state_dim describes the common backbone configuration rather than the GRU hidden state.
Use the profile factory when starting a new training run. This example selects the compact GRU architecture and a PyTorch fallback backend so you can inspect it on a machine without fused CUDA kernels:
from axosim import AXOSIM_MODEL_PROFILES, create_axosim_profile
profile = AXOSIM_MODEL_PROFILES["axosim-gru-small"]
model = create_axosim_profile(
profile.profile_id,
use_pytorch_fallback=True,
)
print(profile.public_name)
print(type(model).__name__)
AxoSim-GRU Small
AxoTemporalModel
The factory returns the named GRU implementation, AxoTemporalModel. Its weights are initialized for training; to make trained predictions, load a checkpoint. Preserve the backend choice when saving and restoring a model, and use the fused backend for checkpoints trained with it.
Load trained weights
For inference, restore the checkpoint’s configuration instead of reconstructing a model from its display name:
from axosim.checkpoint import load_checkpoint
model, metadata = load_checkpoint("runs/axosim-mamba.pt", map_location="cpu")
model.eval()
print(type(model).__name__)
Use a trusted checkpoint from your training run or a verified release. load_checkpoint restores the saved model class, weights, and configuration, including supported historical formats. The checkpoint reference describes the persistence contract; inference demonstrates input construction, morphology IDs, output conversion, and streaming.
The top-level AxoSimGRU export names the compatible block-forecast class CausalBlockForecastModel. Use create_axosim_profile for the named GRU profiles above, or consult block-forecast models when constructing that class directly.
Next steps
Run single-neuron inference, train a model, or construct a Lite population. For a connected application, the fly-connectome article follows population activity from visual input to low-dimensional motion representations.