Model Sharing

Model Sharing#

Guide for sharing your trained scXpand models with the community.

Upload Steps#

1. Prepare Your Model Archive#

# Required files: model.joblib/model.pt, data_format.json, parameters.json, *.npz, *.csv
zip -r logistic_model.zip model.joblib data_format.json parameters.json *.npz *.csv

2. Upload to File Hosting Platform#

Upload your ZIP archive to any file hosting platform that provides direct stable public download URLs like Figshare.

3. Get Direct Download URL#

4. Add to scXpand Registry#

# In scxpand/pretrained/model_registry.py
"my_tumor_classifier": PretrainedModelInfo(  # model_name: any descriptive identifier
    name="my_tumor_classifier",
    url="https://your-platform.com/direct-download-url.zip",
    version="1.0.0",
),

Important Notes:

  • model_name: Registry identifier - can be any descriptive string (e.g., “tumor_classifier”, “blood_model”)

  • Model type: Automatically detected from model_type.txt file in the model archive

Note: The system automatically handles files with numeric prefixes (e.g., 57702349_data_format.json) by normalizing them to standard names (data_format.json).

5. Test#

import scxpand

# Test download and inference
results = scxpand.run_inference(
    model_name="my_tumor_classifier",  # Use your model_name (registry identifier)
    data_path="test_data.h5ad",
    device=None  # Auto-detect best available device
)

Usage#

Users can access models in two ways:

Registry Models (Curated)#

# List curated models
scxpand list-models

# Use curated model
scxpand inference --data_path data.h5ad --model_name pan_cancer_logistic

Direct URL Access (Any External Model)#

# Use any external model directly via URL - no code changes needed!
scxpand inference --data_path data.h5ad --model_url "https://your-platform.com/model.zip"

# Python API also supports both approaches
import scxpand

# Registry model
results = scxpand.run_inference(
    model_name="pan_cancer_logistic",  # model_name: registry identifier
    data_path="data.h5ad",
    device=None  # Auto-detect device
)

# Direct URL (seamless model sharing!)
results = scxpand.run_inference(
    model_url="https://your-platform.com/model.zip",
    data_path="data.h5ad",
    device=None  # Auto-detect device
)
Contents