Model Zoo

TODO this page has been generated by AI. needs checking before publishing.

OpenRetina provides a collection of pre-trained retinal models from published research. These models can be easily loaded and used for inference, analysis, or as starting points for further training.

Available Pre-trained Models

All models are automatically downloaded and cached when first used. They are hosted on Hugging Face.

Höfling et al., 2024 Models

Based on the paper "A chromatic feature detector in the retina signals visual context changes" (eLife 2024).

Model Name	Type	Input Shape	Description	Size
hoefling_2024_base_low_res	Core-Readout	(2, T, 16, 18)	Low-resolution model trained on mouse retina calcium imaging data	~50MB
hoefling_2024_base_high_res	Core-Readout	(2, T, 32, 36)	High-resolution version with larger spatial input	~80MB

Dataset: Mouse retina calcium imaging responses to natural scenes and artificial stimuli. TODO link

Usage Example:

from openretina.models import load_core_readout_from_remote

# Load low-resolution model
model = load_core_readout_from_remote("hoefling_2024_base_low_res", "cpu")

# Create sample input (batch_size=1, time_steps=50, height=16, width=18, channels=2)
stimulus = torch.rand(1, 50, 16, 18, 2)
responses = model(stimulus)

Karamanlis et al., 2024 Models

Based on the paper "Nonlinear receptive fields evoke redundant retinal coding of natural scenes" (Nature 2024).

Model Name	Type	Input Shape	Description	Size
karamanlis_2024_base	Core-Readout	(1, T, H, W)	Base convolutional model for primate retina	~60MB
karamanlis_2024_gru	GRU Core-Readout	(1, T, H, W)	Model with GRU temporal processing	~70MB

Dataset: Primate retina responses to natural scenes. TODO dataset link.

Usage Example:

# Load GRU-based model
model = load_core_readout_from_remote("karamanlis_2024_gru", "cuda")

# The exact input dimensions depend on the specific dataset preprocessing
stimulus_shape = model.stimulus_shape(time_steps=100)
stimulus = torch.rand(stimulus_shape)
responses = model(stimulus)

Maheswaranathan et al., 2023 Models

Based on the paper "Interpreting the retinal neural code for natural scenes: From computations to neurons" (Neuron 2023).

Model Name	Type	Input Shape	Description	Size
maheswaranathan_2023_base	Core-Readout	(1, T, H, W)	Base model for salamander retina	~45MB
maheswaranathan_2023_gru	GRU Core-Readout	(1, T, H, W)	Model with recurrent temporal dynamics	~55MB

Dataset: Salamander retina responses to natural movies. TODO dataset link.

Usage Example:

# Load base model
model = load_core_readout_from_remote("maheswaranathan_2023_base", "cpu")

# Get appropriate input shape
input_shape = model.stimulus_shape(time_steps=200, num_batches=4)
stimulus = torch.rand(input_shape)
responses = model(stimulus)

Loading and Using Models

Basic Loading

import torch
from openretina.models import load_core_readout_from_remote

# Load any available model
model = load_core_readout_from_remote("hoefling_2024_base_low_res", "cpu")

# Check model properties
print(f"Number of neurons: {model.readout.n_neurons}")
print(f"Input shape for 50 time steps: {model.stimulus_shape(time_steps=50)}")

Device Handling

# Load on GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model = load_core_readout_from_remote("karamanlis_2024_gru", device)

# Move existing model to different device
model = model.to("cuda")

Model Storage and Caching

Models are automatically cached in your local filesystem:

• Default cache location: ~/.cache/openretina/
• Custom cache location: Set via cache_directory_path parameter
• Manual cache management: Use openretina.utils.file_utils functions

from openretina.utils.file_utils import get_cache_directory

# Check cache location
cache_dir = get_cache_directory()
print(f"Models cached in: {cache_dir}")

# Load with custom cache location
model = load_core_readout_from_remote(
    "hoefling_2024_base_low_res", 
    "cpu", 
    cache_directory_path="/custom/path"
)

Troubleshooting

Common Issues

Model download fails: - Check internet connection - Verify cache directory permissions - Try different cache location

Out of memory errors: - Use CPU instead of GPU for inference - Reduce batch size or temporal length - Use lower resolution models

Input shape mismatches: - Use model.stimulus_shape() to get correct input dimensions - Check channel ordering (some models expect specific color channels) - Verify temporal length is appropriate

Getting Help

For model-specific issues: 1. Check the FAQ 2. Review original paper documentation 3. Open an issue on GitHub 4. Contact the model authors