https://github.com/bsamseth/eigennn
Neural Network implementation for use in Quantum Variational Monte Carlo
https://github.com/bsamseth/eigennn
Last synced: 3 months ago
JSON representation
Neural Network implementation for use in Quantum Variational Monte Carlo
- Host: GitHub
- URL: https://github.com/bsamseth/eigennn
- Owner: bsamseth
- Created: 2019-01-15T13:57:21.000Z (over 6 years ago)
- Default Branch: master
- Last Pushed: 2019-01-15T14:06:12.000Z (over 6 years ago)
- Last Synced: 2025-02-06T12:15:00.184Z (5 months ago)
- Language: C++
- Size: 95.7 KB
- Stars: 0
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Neural Network for Quantum Variational Monte Carlo
This is an implementation of a neural network for use in quantum VMC. The
reason for this, seeing as there are a multitude of NN implementations
available in Python, is mainly to provide efficient first _and_ second order
gradients. Especially the latter is needed to compute the __Laplacian of the network__, a
key component of VMC calculations. This has shown to be a technical
nightmare/extremely inefficient to compute with automatic differentiation tools
in e.g. TensorFlow.
The structure is modular and allows to add arbitrary layers and arbitrary
activation functions. It's built with Eigen in order to be as efficient as possible,
although not a lot of emphasis has been placed on optimizing.
__NOTE__: This package is not complete, both in functionality and
documentation. It is used within my Master's thesis project, over at
[bsamseth/thesis](github.com/bsamseth/thesis)
## Usage Example
The use of this is not seen by it self, but for reference, this is how the module could be used.
```python
import numpy as np
from EigenNN import Dnn
from EigenNN.layer import DenseLayer
from EigenNN.activation import identity, relu, sigmoid
dnn = Dnn()
dnn.add_layer(DenseLayer(2, 50, activation=sigmoid))
dnn.add_layer(DenseLayer(50, 25, activation=relu))
dnn.add_layer(DenseLayer(25, 1)) # Default is identity activation
x = np.random.randn(4, 2)
output = dnn.evaluate(x) # Shape (4, 1)
# Gradients are summed over the sample rows of x
parameter_gradients = dnn.parameter_gradient(x) # Shape (1451,)
gradient = dnn.gradient(x) # Shape (2,)
laplace = dnn.laplace(x) # Real number
```