https://github.com/cyrildever/neural-networks-and-deep-learning-in-golang
Code adaptation of "Neural Networks and Deep Learning" book in Go
https://github.com/cyrildever/neural-networks-and-deep-learning-in-golang
deep-learning deep-neural-networks golang neural-networks
Last synced: about 2 months ago
JSON representation
Code adaptation of "Neural Networks and Deep Learning" book in Go
- Host: GitHub
- URL: https://github.com/cyrildever/neural-networks-and-deep-learning-in-golang
- Owner: cyrildever
- License: mit
- Created: 2020-06-21T10:47:54.000Z (almost 5 years ago)
- Default Branch: master
- Last Pushed: 2024-10-18T10:10:19.000Z (8 months ago)
- Last Synced: 2025-03-28T21:01:48.022Z (2 months ago)
- Topics: deep-learning, deep-neural-networks, golang, neural-networks
- Language: Go
- Homepage:
- Size: 32.3 MB
- Stars: 1
- Watchers: 1
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# neural-networks-and-deep-learning-in-golang
Code freely adaptated from Michael Nielsen's ["Neural Networks and Deep Learning"](http://neuralnetworksanddeeplearning.com/) book in Go.
_NB: Instead of using the 'mnist.pkl.gz' file, I used the MNIST set [in CSV](https://pjreddie.com/projects/mnist-in-csv/) as provided by Joseph Redmon and zipped the test and the train sets in the 'data' folder._
### Motivation
In order to practice neural networks in Go, I decided to make my own adaptation of Michael Nielsen's reference book on the subject.
Note that it's neither optimized nor production ready: it's just an adaptation to stay as close as possible to his Python code. Don't view this as some kind of Go reference on the subject, but just as a small exercise.
A few notes on the results:
* For the first implementation (eq. `network.py`), the best results were reached by using 300 instead of 30 neurons in the hidden layer, which seems more coherent to results found by other researchers1 in the field for the same kind of network;
* Results on the second implementation (eq. `network2.py`) finally went as high as the original implementation, but it took more epochs (rather 400 than 30), and a change of eta value (from 0.12 to 0.5) after the first 30 epochs, and still 300 neurons in the hidden layer;
* I didn't try to adapt the third implementation (id. `network3.py`) as it would have needed to reimplement Theano's lib in Go as well. Please let me know if you have so that I can point to it.
### Installation
```console
$ git clone https://github.com/cyrildever/neural-networks-and-deep-learning-in-golang.git
$ cd neural-networks-and-deep-learning-in-golang
$ go build
```
### Usage
```console
$ ./neuraldeep -n=1 -op=train -layers="784,300,10" -data=training -useMNIST=true -epochs=30 -size=10 -eta=3.0 -load=false -eval=true
```
```
Usage of ./neuraldeep:
-cost string
cost function: crossEntropy | quadratic (default "crossEntropy")
-data string
a single data set to feed the first layer (a comma-separated list of float64), or the name of the MNIST set (test | training | validation)
-epochs int
number of epochs (default 1)
-eta float
learning rate (default 0.1)
-eval true
set to true to add evaluation at each training epoch
-label string
the label/target of the passed value as a float64 number
-lambda float
the regularization parameter
-layers string
comma-separated list of number of neurons per layer (the first one being the size of the input layer)
-load true
set to true if you want to load an existing network
-n string
the network implementation to use: 1 | 2 | 3 (default "1")
-op string
operation to proceed: predict | test | train
-path string
path to the existing file (default "./data/saved/network/")
-size int
mini-batch size (default 10)
-src string
the source file to use as input data
-mnist
set to true to use MNIST dataset (the layers flag should start with 784 and end with 10)
```
## License
The code in Go is distributed under an [MIT license](LICENSE). \
Please see [Michael Nielsen's website](http://neuralnetworksanddeeplearning.com/) for his authorization.
© 2020-2024 Cyril Dever. All rights reserved.
---
[1](#footnote1): Stuart Russell, Peter Norvig, "Artificial Intelligence: A Modern Approach", 3rd Edition.