Metadata-Version: 2.4
Name: lpcf
Version: 0.1.1
Summary: Learning parametric convex functions
Author-email: Maximilian Schaller <mschall@stanford.edu>, Alberto Bemporad <alberto.bemporad@imtlucca.it>
License: Apache 2.0
Project-URL: Homepage, https://github.com/cvxgrp/lpcf
Requires-Python: <3.13,>=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: jax-sysid>=1.0.6
Requires-Dist: cvxpy>=1.6.0
Requires-Dist: numpy>=1.21.6
Dynamic: license-file

# LPCF
LPCF stands for *learning parametrized convex functions*.
A parametrized convex function, or PCF, depends on a variable and a parameter,
and is convex in the variable for any valid value of the parameter.  

LPCF is a framework for fitting a parametrized convex function to some given data 
that is compatible with *disciplined convex programming*.
This allows to fit a function that can be used in a convex optimization
formulation directly to observed or simulated data.

The PCF is represented as a simple
neural network whose architecture is designed
to ensure disciplined convexity in the variable, for any valid
parameter value. After fitting this neural network to triplets
of observed (or simulated) values of the function, the variable,
and the parameter, the learned PCF can be exported for use in optimization
frameworks like [CVXPY](https://www.cvxpy.org) or [JAX](https://docs.jax.dev/en/latest/index.html).

LPCF supports learning vector functions that depend on multiple variables and parameters.
An overview of LPCF can be found in our [manuscript](https://stanford.edu/~boyd/papers/lpcf.html).

## Installation
LPCF is available on PyPI, and can be installed with
```
pip install lpcf
```

LPCF has the following dependencies:

- Python >= 3.10, <3.13
- jax-sysid >= 1.0.6
- CVXPY >= 1.6.0
- NumPy >= 1.21.6

## Example
The following code fits a PCF to observed function values `Y`,
variable values `X`, and parameter values `Theta`, and
exports the result to CVXPY.

```python3
from lpcf.pcf import PCF

# observed data
Y = ...      # shape (N, d)
X = ...      # shape (N, n)
Theta = ...  # shape (N, p)

# fit PCF to data
pcf = PCF()
pcf.fit(Y, X, Theta)

# export PCF to CVXPY
x = cp.Variable((n, 1))
theta = cp.Parameter((p, 1))
pcf_cvxpy = pcf.tocvxpy(x=x, theta=theta)
```
The CVXPY expression `pcf_cvxpy`
might appear in the objective or the constraints of a CVXPY problem.


## Settings

### Neural network architecture
The function is approximated as an input-convex *main network* mapping variables to function values.
The weights of the main network are generated by another *parameter network*, whose inputs are the parameters.

When constructing the `PCF` object, we allow for a number of
customizations to the neural network architecture:

| Argument         | Description                                                            | Type       | Default         |
| ---------------- | ---------------------------------------------------------------------- | ---------- | --------------- |
| `widths`         | widths of the main network's hidden layers                             | array-like | `[2((n+d)//2), 2((n+d)//2)]` |
| `widths_psi`     | widths of the parameter network's hidden layers                        | array-like | `[2((p+m)//2), 2((p+m)//2)]` |
| `activation`     | activation function used in the main network                           | str        | `'relu'`        |
| `activation_psi` | activation function used in the parameter network                      | str        | `'relu'`        |
| `nonneg`         | Force the PCF to be nonnegative                                        | Bool       | `False`         |
| `increasing`     | Force the PCF to be increasing                                         | Bool       | `False`         |
| `decreasing`     | Force the PCF to be decreasing                                         | Bool       | `False`         |
| `quadratic`      | Include a convex quadratic term in the PCF                             | Bool       | `False`         |
| `quadratic_r`    | Include a quadratic term with low-rank + diagonal structure            | Bool       | `False`         |
| `classification` | Use the PCF to solve a classification problem                          | Bool       | `False`         |

Note that `d` is the number of components of the function, `n` the number of variables, `p` the
number of parameters, and `m` the number of outputs of the parameter network, i.e., the number of weights 
of the main network.

### Learning configuration
When fitting the `PCF` to data with its `.fit()` method, we provide
the following options:

| Argument         | Description                                                            | Type       | Default         |
| ---------------- | ---------------------------------------------------------------------- | ---------- | --------------- |
| `rho_th`         | regularization on the sum of squared weights of the parameter network  | float      | `1e-8`          |
| `tau_th`         | regularization on the sum of absolute weights of the parameter network | float      | `0`             |
| `zero_coeff`     | entries smaller (in abs value) than `zero_coeff` are zeroed            | float      | `1e-4`          |
| `cores`          | number of cores used for parallel training                             | int        | `4`             |
| `seeds`          | random seeds for training from multiple initial guesses                | array-like | `max(10, cores)`|
| `adam_epochs`    | number of epochs for running ADAM                                      | int        | `200`           |
| `lbfgs_epochs`   | number of epochs for running L-BFGS-B                                  | int        | `2000`          |
| `tune`           | auto-tune `tau_th`?                                                    | Bool       | `False`         |
| `n_folds`        | number of cross-validation folds when auto-tuning `tau_th`             | int        | `5`             |
| `warm_start`     | warm-start training?                                                   | Bool       | `False`         |


## Citing LPCF

<a name="ref1"></a>
Please cite the following paper if you use this software:

```
@article{SBB25,
    author={Maximilian Schaller and Alberto Bemporad and Stephen Boyd},
    title={Learning Parametric Convex Functions},
    note = {available on arXiv at \url{https://arxiv.org/pdf/2506.04183}},
    year=2025
}
```
