DynVision — Trained Models and Evaluation Reports

This repository contains pre-trained model checkpoints and evaluation reports from the DynVision toolbox — a modular framework for constructing and evaluating recurrent convolutional neural networks (RCNNs) with biologically inspired temporal dynamics.

Model Description

DynVision implements visual network architectures where neural activity evolves according to continuous-time differential equations with realistic timescales. Recurrent connections within and across processing stages allow the network to integrate information over time, producing dynamics that can be aligned with properties of biological visual systems.

Architectures Included

Architecture	Type	Description
DyRCNNx8	Recurrent	8-layer dynamic RCNN with configurable recurrence types (full, self, depthwise, pointwise) and feedback/skip connections
CorNet-RT	Recurrent	CORnet model of the primate ventral visual stream with anatomically-inspired recurrent connections between areas V1, V2, V4, IT (Kubilius et al., 2018)
CordsNet	Recurrent	Scale-invariant contour integration with recurrent dynamics (Soo et al., 2024)

Training

• Framework: PyTorch Lightning with Snakemake workflow management
• Data loading: FFCV for high-throughput image loading
• Optimizer: Adam with cosine annealing
• Mixed precision: Automatic (AMP)
• Seeds: 7000–7003
• Datasets: Imagenette, ImageNet, CIFAR-10, MNIST

Evaluation Experiments

The reports directory contains evaluation results across multiple experimental paradigms designed to probe temporal response properties:

• contrast: 9 CSV files
• duration: 9 CSV files
• gaussiannoise: 16 CSV files
• gaussiannoiseffonly: 11 CSV files
• hundred: 5 CSV files
• interval: 9 CSV files
• response: 52 CSV files
• responseintermediate: 12 CSV files
• stability: 9 CSV files
• ten: 4 CSV files

Repository Contents

Item	Count
Model configs	190
Report files	136
Total size	20005 MB

models/
  {architecture}/
    {model_config}/
      {dataset}/
        trained-best.pt         # Best checkpoint weights
        trained.pt.config.yaml  # Training hyperparameters

reports/
  {experiment}/
    {model_config}/
      {data_group}/
        {run_hash}/
          test_data.csv         # Evaluation metrics

Usage

Loading a Pre-trained Model

from huggingface_hub import hf_hub_download
from dynvision.utils.model_utils import load_model_and_weights

# Download a specific checkpoint
checkpoint_path = hf_hub_download(
    repo_id="neurograce/DynVision",
    filename="models/DyRCNNx8/DyRCNNx8:tsteps=20+dt=2+...+_7000/imagenette/trained-best.pt",
    repo_type="model",
)

# Load weights into the model architecture
model = load_model_and_weights(
    model_name="DyRCNNx8",
    weights_path=checkpoint_path,
)

Accessing Evaluation Reports

import pandas as pd
from huggingface_hub import hf_hub_download

# Download a report CSV
report_path = hf_hub_download(
    repo_id="neurograce/DynVision",
    filename="reports/contrast/DyRCNNx8:tsteps=20+...+_7000/imagenette:all_trained-best/test_data.csv",
    repo_type="model",
)

df = pd.read_csv(report_path)
print(df.head())

Downloading the Full Model Directory

from huggingface_hub import snapshot_download

# Download all files for offline use
local_path = snapshot_download(
    repo_id="neurograce/DynVision",
    repo_type="model",
    local_dir="./dynvision_models",
)

Intended Uses

• Reproducing experimental results from the DynVision manuscript
• Benchmarking biologically-inspired recurrent architectures
• Studying temporal dynamics and response properties of visual networks
• Transfer learning and feature extraction with recurrence-augmented backbones

Limitations

• Models were trained on specific image datasets (Imagenette, ImageNet, CIFAR-10, MNIST) and may not generalize to other domains without fine-tuning.
• Recurrent dynamics are sensitive to temporal parameters (τ, dt, t_recurrence); inference with different timestep configurations may produce unexpected behavior.
• Per-epoch checkpoints capture training trajectories but not all epochs are guaranteed to be stable minima.

Citation

If you use these models or the DynVision toolbox in your research, please cite:

@article{gutzen2025dynvision,
  title={DynVision: A Modular Toolbox for Biologically Plausible Recurrent Visual Networks},
  author={Gutzen, Robin and others},
  journal={bioRxiv},
  year={2025},
  doi={10.1101/2025.08.11.669756}
}

License

This repository is distributed under the MIT License.