ProperModelVault

Description

ProperModelVault is a secure, file-system based registry for machine learning models. It focuses on data integrity and safe storage using atomic writes and SHA-256 checksums. By default, it uses safetensors to avoid the security risks of pickle while ensuring fast loading speeds (the original ModelVault used pickle!)

Features

• Atomic Saves: Prevents data corruption by writing to temp files first.
• Integrity Checks: Automatically verifies SHA-256 hashes on every load.
• Safe Serialization: Uses safetensors by default (no code execution risks).
• Version Control: Automatic versioning with metadata tracking.
• Pluggable Backend: Easy to swap serializers via Protocol interface.

Why not pickle?

Pickle is unsafe because it can execute arbitrary code during loading. ModelVault defaults to safetensors to keep your environment secure and your loads fast.

Requirements

pip install torch safetensors

Quick Start

import torch.nn as nn
from modelvault import ModelVault

# define your model architecture
class MyNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.layer = nn.Linear(10, 1)

# initialize vault
vault = ModelVault("./models")

# save a model instance
model = MyNet()
version = vault.save("classifier", model, metadata={"accuracy": 0.92, "epoch": 10})
print(f"Saved classifier as version {version}")

# load the latest version
empty_model = MyNet()
loaded_model = vault.load_latest("classifier", model=empty_model, device="cuda", strict=False)