Skip to content

Latest commit

 

History

History
95 lines (65 loc) · 2.9 KB

README.md

File metadata and controls

95 lines (65 loc) · 2.9 KB
Raw

Image Translation using Conditional Adversarial Network

Conditional GAN for pix2pix Aerial-to-Map images dataset.

This is an implementation of Condition GAN based on a paper by Isola et al. Link : https://arxiv.org/abs/1611.07004

Implemantated using Tensorflow framework for python. This implementation has a U-Net generator as proposed by Isola et al. and also a ResNet-9 generator, of which, any one can be chosen. Optionally, noise can also added to input of generator.

Results

The performance of each model was evaluated by compluting Mean Squared Error (MSE) between the predicted and expected output on test data.

Model MSE
U-Net 0.01834
U-Net with noise 0.01852
ResNet-9 0.01443
ResNet-9 with noise 0.01476

The study showed that ResNet-9 generator performed about 21.3% better as compared to U-Net generator. Although, addition of noise to generator input did not improve results.

See Project Report for details on results, model architecture, choice of hyperparameters, data preprocessing, etc.

Usage

Steps to use code

  1. Import required functions.
from initializeDataset import load_dataset
from c_gan import C_GAN, UNetGenerator, ResNet9Generator, PatchDiscriminator
  1. Create object of C_GAN class.
# for U-Net generator
gan = C_GAN(generator=UNetGenerator(), discriminator=PatchDiscriminator())

# for ResNet-9 generator
gan = C_GAN(generator=ResNet9Generator(), discriminator=PatchDiscriminator())

# for U-Net generator with noise added to inputs
gan = C_GAN(generator=UNetGenerator(noise=True), discriminator=PatchDiscriminator())

# for ResNet-9 generator with noise added to inputs
gan = C_GAN(generator=UNetGenerator(noise=True), discriminator=PatchDiscriminator())

# Default training rate is 2e-4 for both generator and discriminator
# Use 'generator_learning_rate' and 'discriminator_learning_rate' parameters to specify different learning rates
# Use 'ckpt_freq' to change frequency of saving TF checkpoint, default frequency is 20 epochs
  1. Train model.
EPOCHS = 200
gan.train(train_dataset, EPOCHS)
  1. Evaluate model (Calculate Mean Sqaured Error).
mse = gan.evaluate(test_dataset)
  1. Generate samples outputs.
NUM_SAMPLES = 10
for tar, img in test_dataset.take(NUM_SAMPLES):
    gan.generate_image(img, tar)
  1. Predict map image from a given aerial image.
# Input image can be of any size (square images are prefered)
# Outputs is a 3 x 256 x 256 RGB image of prediction (range of values [-1, 1])
gan.predict(input_image)
  1. Restoring model from checkpoint.
gan.restore_from_checkpoint()
  1. Output model summary
# Make sure model is compiled before this
gan.summary()