Repository files navigation

Recipe for simple GAN in Golang ecosystem viaGorgonia library

Generating f(x) = x^2	Generating f(x) = sin(x)

• About
• Why
• Instruments
• Usage
• ToDo
• Code explanation
• Support and contributing

This just a repository with simple example how to build GAN inGorgonia

What is GAN?

• Short: such networks are just two neural networks (Discriminator andGenerator) contesting each other.Generator must "cheat" onDiscriminator and the last one should detect lies.
• Long: Wiki-page about GAN's -https://en.wikipedia.org/wiki/Generative_adversarial_network

Note:although there is code with some wrappings/aliasing and helping abstractions and functions in repository, this does not pretend to be high-level machine learning framework

Note #2:By the way... Code is ugly since I've decided to handle errors instead of using panic(...) calls. Panicing is considered to be in main functions of examples only

Current examples folder contains limited set of layer types:

• Linear
• Convolutional
• Maxpool
• AvgPool
• Flatten
• Reshape
• Reshape
• Dropout
• Embedding
• LSTM

Just want to do that in Golang ecosystem.

Code is written on Golang -https://golang.org/

Used machine learning library -Gorgonia

Plotting library -gonum

• Get the repo

  git clone https://github.com/LdDl/gan-go.git

• Navigate to examples folder

  cd gan-gocd cmd/examples

• Pick one of examples. E.g. parabola:

  cd parabola

• Run example

  go run main.go

• Output

  After programm terminates there should be multiple files:

  1. Single image for reference function - reference_function.png
  2. Multiple images for generated data on N-th epoch - gen_reference_fun_%N-th epoch%.png
  3. Single image for generated data on last epoch - gen_reference_func_final.png

  Example for parabola:

  Actual reference function:

  

  Generated data on 0-th epoch:

  

  Generated data on 10-th epoch:

  

  Generated data on 60-th epoch:

  

  Generated data on 150-th epoch:

  

  Generated data on last epoch:

  

Current stage of TODO list for future releases:

• Reduce duplicating of code for.Fwd() method of each neural network type (GAN/Discriminator/Generator)
• SwitchLayer fromstruct tointerface or use other technique for building clean code
• Add basic layers: Linear, Convolutional, Maxpool, Flatten
• Deal with batch process
• More loss function
  • Cross Entropy
  • Binary Cross Entropy
  • L1
  • Huber (PSEUDO)
• Examples for text data generationWIP
• Simple LSTM
  • Proper layer types for RNNWIP
  • Examples
• RNN
• GRU
• Embedding

@TODO

If you have troubles or questions pleaseopen an issue.

If you want to improve this library / fix some bugs and etc. you canmake PR