Awesome

pytorch-pix2pix

Pytorch implementation of pix2pix [1] for various datasets.

• you can download datasets: https://people.eecs.berkeley.edu/~tinghuiz/projects/pix2pix/datasets/
• you can see more information for network architecture and training details in https://arxiv.org/pdf/1611.07004.pdf

dataset

• cityscapes
  • 2,975 training images, 200 train epochs, 1 batch size, inverse order: True
• facades
  • 400 training images, 200 train epochs, 1 batch size, inverse order: True
• maps
  • 1,096 training images, 200 train epochs, 1 batch size, inverse order: True
• edges2shoes
  • 50k training images, 15 train epochs, 4 batch size, inverse order: False
• edges2handbags
  • 137k training images, 15 train epochs, 4 batch size, inverse order: False

Resutls

cityscapes

• cityscapes after 200 epochs
  • First column: input, second column: output, third column: ground truth

• Generate animation for fixed inputs

• Learning Time
  • cityscapes pix2pix - Avg. per epoch: 332.08 sec; Total 200 epochs: 66,846.58 sec

facades

• facades after 200 epochs
  • First column: input, second column: output, third column: ground truth

• Generate animation for fixed inputs

• Learning Time
  • facades pix2pix - Avg. per epoch: 44.94 sec; Total 200 epochs: 9,282.64 sec

edges2handbags

• edges2handbags after 15 epochs
  • First column: input, second column: output, third column: ground truth

• Generate animation for fixed inputs

• Learning Time
  • edges2handbags pix2pix - Avg. per epoch: 10,228.08 sec; Total 15 epochs: 153,443.62 sec

Development Environment

• Ubuntu 14.04 LTS
• NVIDIA GTX 1080 ti
• cuda 8.0
• Python 2.7.6
• pytorch 0.1.12
• matplotlib 1.3.1
• imageio 2.2.0
• scipy 0.19.1

Reference

[1] Isola, Phillip, et al. "Image-to-image translation with conditional adversarial networks." arXiv preprint arXiv:1611.07004 (2016).

(Full paper: https://arxiv.org/pdf/1611.07004.pdf)