/fall-2019-research

Primary LanguagePythonBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

Human-in-the-Loop Training of Generative Models

This project is builds on Facebook's PyTorch GAN Zoo (https://github.com/facebookresearch/pytorch_GAN_zoo) using the CIFAR10 dataset

Requirements

- pytorch
- numpy
- torchvision
- scipy
- visdom (optional, for monitoring training)

Usage

Training the model (Progressive GAN, WGANGP loss)

  • Activate visdom server before training to monitor progress (optional):

    • python -m visdom.server

  • Train PGAN, load most recently saved model. Save every (S) iterations (minibatches, batch size 16). Run this in another shell.

    • python train.py PGAN -c config_cifar10.json -n cifar10 -s (S)

  • navigate to the url specified in visdom shell (usually http://localhost:8097) to monitor training

Visualization

  • generate (N) new feature map samples (size 200x200) without training the model
    • python train.py PGAN -c config_cifar10.json -n cifar10 -nS (N) -xT
  • Prep loss binary for visualization
    • python helpers.py
  • Activate the web server: '''python -m http.server'''
  • navigate to http://localhost:8000/pgan_vis/

screenshot

To do

  • 11-3 - 11-9

    • input plot;
    • colormap for all netG and netD layers based on input plot
    • discriminator layers; colormap;

  • 11-10 - 11-16:

    • interactivity:
      • brush select
      • image grid

  • 11-17 - 11-23:

    • brush
    • change sample layers
    • image grid

  • 11-24 - 11-30:

    • image grid

  • 12-15 - 12-21;

    • fix image grid
    • image scaling issue

  • model training error (RAM)

  • loss function for evaluating distance between samples