/NeuralStyleTransfer-tensorflow

Implementation of Neural Style Transfer in Tensorflow

Primary LanguagePythonMIT LicenseMIT

NeuralStyleTransfer-tensorflow

Implementation of Neural Style Transfer from the paper A Neural Algorithm of Artistic Style in Tensorflow

Requirements

  • tensorflow-gpu==1.4.0
  • tqdm==4.23.0
  • scikit-image==0.13.1

Usage

  • Create folder named pretrained_model
  • Download pretrained VGG 19 and put it into pretrained_model folder
  • Open terminal and type the command
  • Command usage:
python nst.py [-h] -c CONTENT_IMG -s STYLE_IMG [-o OUTPUT_FOLDER]
              [-n N_ITERATIONS] [-e SAVE_EVERY_N_ITERATIONS] [-f OUTPUT_NAME]
              [-p PRETRAINED_MODEL] [-a ALPHA] [-b BETA] [-lr LEARNING_RATE]
              [-ht HEIGHT] [-w WIDTH] [-ch CHANNELS]
  • Command arguments:
  -h, --help            show this help message and exit
  -c CONTENT_IMG, --content_img CONTENT_IMG
                        path to content image
  -s STYLE_IMG, --style_img STYLE_IMG
                        path to style image
  -o OUTPUT_FOLDER, --output_folder OUTPUT_FOLDER
                        path to output folder
  -n N_ITERATIONS, --n_iterations N_ITERATIONS
                        number of iterations
  -e SAVE_EVERY_N_ITERATIONS, --save_every_n_iterations SAVE_EVERY_N_ITERATIONS
                        every n iterations the model will save an output image
  -f OUTPUT_NAME, --output_name OUTPUT_NAME
                        output image name
  -p PRETRAINED_MODEL, --pretrained_model PRETRAINED_MODEL
                        path to pretraned model
  -a ALPHA, --alpha ALPHA
                        importance of content cost
  -b BETA, --beta BETA  importance of style cost
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                        learning rate
  -ht HEIGHT, --height HEIGHT
                        height of image
  -w WIDTH, --width WIDTH
                        width of image
  -ch CHANNELS, --channels CHANNELS
                        channels of image

Example

Content image:

example/cat2.jpg

cat2.jpg

Style image:

example/starry_night.jpg

starry_night.jpg

Command line:

python nst.py -c example/cat2.jpg -s example/starry_night.jpg -n 400 -e 100

Result:

  • Iteration 0:

output/generated_0.png

generated_0

  • Iteration 100:

output/generated_100.png

generated_100

  • Iteration 200:

output/generated_200.png

generated_200

  • Iteration 300:

output/generated_300.png

generated_300

  • Final (Iteration 400):

output/generated.jpg

generated

References

Tensorflow神经风格迁移

用Tensorflow实现神经风格迁移， 论文：A Neural Algorithm of Artistic Style

环境要求

  • tensorflow-gpu==1.4.0
  • tqdm==4.23.0
  • scikit-image==0.13.1

使用方法

  • 创建文件夹 pretrained_model
  • 下载 训练过的 VGG 19，并放到 pretrained_model 文件夹里
  • 打开终端执行命令行
  • 命令行使用方法:
python nst.py [-h] -c CONTENT_IMG -s STYLE_IMG [-o OUTPUT_FOLDER]
              [-n N_ITERATIONS] [-e SAVE_EVERY_N_ITERATIONS] [-f OUTPUT_NAME]
              [-p PRETRAINED_MODEL] [-a ALPHA] [-b BETA] [-lr LEARNING_RATE]
              [-ht HEIGHT] [-w WIDTH] [-ch CHANNELS]
  • 命令行参数:
  -h, --help            显示帮助信息并退出
  -c CONTENT_IMG, --content_img CONTENT_IMG
                        内容图片文件路径
  -s STYLE_IMG, --style_img STYLE_IMG
                        风格图片文件路径
  -o OUTPUT_FOLDER, --output_folder OUTPUT_FOLDER
                        输出文件夹路径
  -n N_ITERATIONS, --n_iterations N_ITERATIONS
                        迭代次数
  -e SAVE_EVERY_N_ITERATIONS, --save_every_n_iterations SAVE_EVERY_N_ITERATIONS
                        保存生成图片的间隔次数
  -f OUTPUT_NAME, --output_name OUTPUT_NAME
                        输出图片名
  -p PRETRAINED_MODEL, --pretrained_model PRETRAINED_MODEL
                        预训练模型路径
  -a ALPHA, --alpha ALPHA
                        内容损失值权重
  -b BETA, --beta BETA  风格损失值权重
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                        学习率
  -ht HEIGHT, --height HEIGHT
                        图片高度
  -w WIDTH, --width WIDTH
                        图片宽度
  -ch CHANNELS, --channels CHANNELS
                        图片通道数

例子

内容图片:

example/cat2.jpg

cat2.jpg

风格图片:

example/starry_night.jpg

starry_night.jpg

命令行:

python nst.py -c example/cat2.jpg -s example/starry_night.jpg -n 400 -e 100

结果:

  • 迭代0次:

output/generated_0.png

generated_0

  • 迭代100次:

output/generated_100.png

generated_100

  • 迭代200次:

output/generated_200.png

generated_200

  • 迭代300次:

output/generated_300.png

generated_300

  • 最终结果 (迭代400次):

output/generated.jpg

generated

参考