Metadata-Version: 2.1
Name: pyvacy
Version: 0.0.31
Summary: Privacy preserving deep learning for PyTorch
Home-page: http://github.com/ChrisWaites/pyvacy
Author: Chris Waites
Author-email: cwaites10@gmail.com
License: UNKNOWN
Description: <div align="center">
            <img src="assets/logo.png" width="150px"><br>
            <h1>PyVacy: Privacy Algorithms for PyTorch</h1>
        </div>
        
        PyVacy provides custom PyTorch opimizers for conducting deep learning in a differentially private manner. Basically <a href="https://github.com/tensorflow/privacy">TensorFlow Privacy</a>, but for PyTorch.
        
        ## Getting Started
        
        ```bash
        pip install pyvacy
        ```
        
        ## Example Usage
        
        ```python
        import torch
        
        from pyvacy import optim
        from pyvacy.analysis import moments_accountant
        
        model = torch.nn.Sequential(...)
        
        optimizer = optim.DPSGD(
            l2_norm_clip=...,
            noise_multiplier=...,
            batch_size=...,
            lr=...,
            momentum=...,
        )
        
        epsilon = moments_accountant.epsilon(
            N=...,
            batch_size=...
            noise_multiplier=...,
            epochs=...,
            delta=...,
        )
        
        for epoch in range(epochs):
            # do training as usual...
        ```
        
        ## Tutorials
        
        ```python
        python tutorials/mnist.py
        
        Training procedure achieves (3.0, 0.00001)-DP
        [Epoch 1/60] [Batch 0/235] [Loss: 2.321049]
        [Epoch 1/60] [Batch 10/235] [Loss: 0.952795]
        [Epoch 1/60] [Batch 20/235] [Loss: 1.040896]
        ...
        ```
        
Platform: UNKNOWN
Description-Content-Type: text/markdown
