A PyTorch library for fitting sensory neural responses with deep neural network models
Contact: Ulysse Rançon — @urancon, ulysse.rancon@uni-goettingen.de
deepSTRF is a community-oriented library for system identification of sensory neurons — predicting trial-resolved neural responses (spikes, calcium fluorescence, EEG, intracellular potential, ...) from naturalistic stimuli with PyTorch models. It bundles:
- Datasets. A growing zoo of publicly available recordings (auditory cortex, midbrain, songbird auditory pallium, EEG, ...) behind a single
NeuralDatasetAPI with consistent NaN-sentinel handling for missing trials, optionaldownload=Trueauto-download, and built-in selection / concatenation utilities. - Models. A unified four-slot template (
wav2spec → prefiltering → core → readout) with reference implementations of widely used encoders (Linear, 2D-CNN, StateNet, DNet, Transformer, NRF) and the AdapTrans module of ON/OFF auditory adaptation. - Pretrained checkpoints published on the Hugging Face Hub.
- Metrics. NaN-aware Pearson, FVE, Schoppe-normalized correlation, signal/noise power, coherence — all functional and
torch.compile-friendly. - Training utility. A thin opt-in
Fitter(~150 lines) for early-stopping + best-checkpoint training, on top of the canonical PyTorch loop.
📖 Full documentation: deepstrf.readthedocs.io
deepSTRF requires Python ≥ 3.10. Install the latest release from PyPI:
pip install deepSTRFFor a development checkout, install from source instead:
git clone https://github.com/urancon/deepSTRF
cd deepSTRF
pip install -e ".[dev]"# or `pip install -e .` for runtime onlyOptional extras: [docs], [allen] (Allen Brain Observatory tooling), [s4] (CUDA kernels for S4 layers), [eeg] (MNE for .fif parsing).
See the Installation guide for conda recipes and troubleshooting.
Load a published checkpoint, score it on the canonical NS1 ferret-A1 dataset:
fromtorch.utils.dataimportDataLoaderfromdeepSTRF.datasets.audio.ns1importNS1DatasetfromdeepSTRF.models.audioimportStateNetfromdeepSTRF.metricsimportcorrcoef, normalized_corrcoeffromdeepSTRF.utils.dataimportneural_collate# 1) Load a dataset (auto-downloads to a local cache the first time).ds=NS1Dataset(download=True, dt_ms=5)
loader=DataLoader(ds, batch_size=8, collate_fn=neural_collate)
# 2) Load a pretrained model from the Hugging Face Hub.model=StateNet.from_pretrained("urancon/deepSTRF-statenet-gru-ns1").eval()
# 3) Score it. Each batch is a dict: 'stims', 'responses', 'valid_mask', 'stim_meta'.batch=next(iter(loader))
responses=batch['responses']
pred=model(batch['stims']) # (B, N, R=1, T)psth=responses.nanmean(dim=2, keepdim=True)
cc=corrcoef(pred, psth, reduction='mean')
cc_norm=normalized_corrcoef(pred, responses, method='schoppe', reduction='mean')
print(f"CCraw = {cc:.3f} CCnorm = {cc_norm:.3f}")The opt-in Fitter wraps the canonical training loop (loss + early stop + best-checkpoint selection) in ~10 lines of user code:
fromtorch.optimimportAdamfromdeepSTRF.trainingimportFitterfromdeepSTRF.metricsimportmse_loss, normalized_corrcoeffitter=Fitter(
model=model,
optimizer=Adam(model.parameters(), lr=1e-3),
train_loader=train_loader,
val_loader=val_loader,
loss_fn=mse_loss,
val_metrics={
'cc_norm': lambdapred, resp: normalized_corrcoef(pred, resp, method='schoppe', reduction='mean'),
},
)
fitter.fit(num_epochs=50)For custom loops (mixed precision, multi-GPU, curricula, ...), the canonical three-line PyTorch loop documented in metrics_paradigm.md §7 stays a one-liner thanks to the metrics API. See the Fitter docs for hooks and design rationale.
Runnable notebooks live under examples/ — each opens in Colab in one click.
| Notebook | Focus |
|---|---|
crcns_aa_tutorial.ipynb | Start here. Load CRCNS AA1 / AA2 zebra finch data end-to-end. |
explore_nat4.ipynb | Inspect NAT4 ferret A1 / PEG recordings. |
dataset_concatenation.ipynb | Mix multiple datasets behind one DataLoader. |
fit_ns1_statenet.ipynb | Fit StateNet on NS1 from scratch. |
load_pretrained_statenet_ns1.ipynb | Reuse the published HF Hub checkpoint. |
alice_eeg_tutorial.ipynb | EEG (Brodbeck 2023, "Alice"). |
le_2025_baseline.ipynb | Zebra finch responses to occluded conspecific song. |
strf_parameterizations_ns1.ipynb | Parametric Gaussian-mixture STRFs. |
strf_gradmap_aa2.ipynb | Gradient-attribution receptive fields on AA2. |
adaptrans_transformer_aa1.ipynb | AdapTrans + Transformer on AA1 Field L. |
espejo_nat_nrf.ipynb | Network Receptive Field model on Espejo ferret A1. |
Current top model on each dataset. Want to claim the podium? Open a PR with a ready-to-deploy PyTorch class so others can reproduce.
| Dataset | Model | Remarks | Params / nrn | CCraw / CCnorm [%] | Paper |
|---|---|---|---|---|---|
| NS1 | StateNet | GRU, pop | 30,465 | 55.6 / 75.1 | Rançon et al. |
| NAT4 A1 | StateNet | LSTM, pop | 40,271 | 46.6 / 65.1 | Rançon et al. |
| NAT4 PEG | Transformer | pop | 28,437 | 39.7 / 55.5 | Rançon et al. |
| AA1 Field L | StateNet | GRU, pop | 24,900 | / 71.0 | Rançon et al. |
| AA1 MLd | StateNet | Mamba, pop | 32,334 | / 73.4 | Rançon et al. |
Note. The three CRCNS AC1 datasets (Wehr, Asari A1, Asari MGB) are single-unit fitting only and yield very different results depending on response preprocessing (detrending, spikes vs. raw potential, ...). Their benchmark will be reported separately.
deepSTRF wraps publicly available recordings — please cite the original authors when you use them. See each dataset page on the docs for details and download instructions.
- Auditory:NS1, NAT4, CRCNS AA1, CRCNS AA2, CRCNS AA4, CRCNS AC1 (Wehr + Asari), Espejo 2019, Alice EEG (Brodbeck 2023), Le 2025, Wingert 2026.
- Visual: under construction — see Status below.
The first release of deepSTRF focuses on auditory datasets and models. A working video API is on the roadmap but is not yet shipped on develop — the deepSTRF.datasets.video and deepSTRF.models.video namespaces currently expose only their base classes (VideoNeuralDataset, VideoEncodingModel). Earlier draft loaders (Allen Ophys / Ecephys, CRCNS PVC1 / PVC11 / MT1 / MT2 / VIM2, MICrONS, UW Neural Data Challenge) live on the archive/video-api-v0 branch and will be revived once rewritten against the modernized base class.
Pull requests are welcome — most useful drops are new datasets, new model backbones, and pretrained checkpoints. Please open an issue first so we can scope it together (most importantly to confirm the dataset's license allows redistribution).
If deepSTRF is useful for your work, please cite the relevant paper:
@article{rancon2024pcb,
title = {A general model unifying the adaptive, transient and sustained properties of ON and OFF auditory neural responses},
author = {Rançon, Ulysse and Masquelier, Timothée and Cottereau, Benoit R.},
journal = {PLOS Computational Biology},
year = {2024},
volume = {20},
number = {8},
pages = {1--32},
doi = {10.1371/journal.pcbi.1012288},
}
@article{rancon2025commbio,
title = {Temporal recurrence as a general mechanism to explain neural responses in the auditory system},
author = {Rançon, Ulysse and Masquelier, Timothée and Cottereau, Benoit R.},
journal = {Communications Biology},
year = {2025},
volume = {8},
number = {1},
pages = {1456},
doi = {10.1038/s42003-025-08858-3},
}A running list of papers that build on deepSTRF lives on the Publications docs page. PRs welcome to add yours.