How to Version Model Weights: Best Practices and Examples

To version model weights, save them with clear versioned filenames or use tools like MLflow or Git LFS to track changes. Include metadata such as training date, model parameters, and performance metrics to keep versions organized and reproducible.

Syntax

Versioning model weights usually involves saving the model's parameters with a version identifier. This can be done by naming the file with a version number or timestamp.

Example syntax for saving weights in PyTorch:

• torch.save(model.state_dict(), 'model_v1.pth'): saves weights with version 1.
• torch.load('model_v1.pth'): loads the saved weights.

Using MLflow, you log the model with a version automatically managed by the tool.

python

import torch

# Save model weights with version
torch.save(model.state_dict(), 'model_v1.pth')

# Load model weights
model.load_state_dict(torch.load('model_v1.pth'))

Example

This example shows how to save and load model weights with versioning using PyTorch. It demonstrates saving weights with a versioned filename and loading them back.

python

import torch
import torch.nn as nn

# Define a simple model
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(2, 1)

    def forward(self, x):
        return self.linear(x)

model = SimpleModel()

# Save weights with version 1
torch.save(model.state_dict(), 'simple_model_v1.pth')

# Load weights into a new model instance
new_model = SimpleModel()
new_model.load_state_dict(torch.load('simple_model_v1.pth'))

# Check if weights are the same
same_weights = all(torch.equal(p1, p2) for p1, p2 in zip(model.parameters(), new_model.parameters()))
print(f'Weights match after loading: {same_weights}')

Output

Weights match after loading: True

Common Pitfalls

• Not including version info in filenames can cause confusion and overwrite important weights.
• Failing to save metadata like training parameters or metrics makes it hard to track model performance across versions.
• Using generic file names like model.pth without timestamps or version numbers risks losing previous versions.
• Not using a version control system or model registry can make collaboration and reproducibility difficult.

Always combine versioned filenames with metadata and consider tools like MLflow or DVC for better tracking.

python

import torch

# Wrong: overwriting without version
# torch.save(model.state_dict(), 'model.pth')

# Right: include version or timestamp
import datetime
version = datetime.datetime.now().strftime('%Y%m%d_%H%M%S')
torch.save(model.state_dict(), f'model_{version}.pth')

Quick Reference

Concept	Description	Example
Versioned Filenames	Save weights with version or timestamp in filename	'model_v1.pth', 'model_20240601.pth'
Metadata	Store training info and metrics alongside weights	JSON file or MLflow tags
Model Registry	Use tools to track versions and metadata	MLflow, DVC, or Git LFS
Loading Weights	Load specific version by filename	torch.load('model_v1.pth')
Avoid Overwrites	Never save without version info	Use timestamps or version numbers

Key Takeaways

Always save model weights with clear version identifiers in filenames or use a model registry.

Include metadata like training parameters and metrics to track model versions effectively.

Avoid overwriting weights by using timestamps or version numbers in filenames.

Use tools like MLflow or Git LFS for better version control and collaboration.

Load weights by specifying the exact version to ensure reproducibility.