https://github.com/pandede/cyclebot1001
Transfer the painting style between real-life and Van Gogh's by using CycleGAN
https://github.com/pandede/cyclebot1001
deep-learning gan neural-network style-transfer
Last synced: 9 months ago
JSON representation
Transfer the painting style between real-life and Van Gogh's by using CycleGAN
- Host: GitHub
- URL: https://github.com/pandede/cyclebot1001
- Owner: Pandede
- Created: 2020-10-27T18:46:59.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2020-12-02T08:18:11.000Z (over 5 years ago)
- Last Synced: 2025-07-05T12:45:44.478Z (12 months ago)
- Topics: deep-learning, gan, neural-network, style-transfer
- Language: Python
- Homepage: https://www.instagram.com/cyclebot1001/
- Size: 765 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# CycleBot1001
Transfer the painting style between real-life and Van Gogh's by using CycleGAN. \
The instagram account: http://www.instagram.com/cyclebot1001/
## Preparation
1. Prepare the images of 2 domain for training, saved as following structure:
```
+-- Data
| +-- domain_a
| | +-- 0001.png
| | +-- 0002.png
| | +-- ...
| +-- domain_b
| | +-- 0001.png
| | +-- 0002.png
| | +-- ...
```
For the convenience, you may download the **image URLs** from the following sources, including Van Gogh's painting *(Domain A)* and landmarks in real life *(Domain B)*:
| Name | Size | Type | Link |
|----------|--------|------|--------|
| Van Gogh | ~107MB | CSV | [Google Drive](https://drive.google.com/file/d/1K51AtogQdSMnNEV5irpE1IAAVd-1DVCR/view?usp=sharing) |
| Landmark | ~119KB | CSV | [Google Drive](https://drive.google.com/file/d/1s4jAnVWdJ_vzbPWv83dx_v_YQS_RL4Kh/view?usp=sharing) |
Then, execute the following code in python console, for example, downloading images of Van Gogh *(Domain A)*:
```python
import pandas as pd
from streamer import ImageStreamer
src_list = pd.read_csv('./van_gogh.csv')['ImageURL']
s = ImageStreamer(1024)
s.parallel_download(src_list, './Data/domain_a/', num_workers=8, delay=1.)
```
- The argument in `ImageStreamer` denoted the image size of image. It can be `tuple`, `int` or `None`:
- the image will be resized in (1024, 256) if the argument is `tuple(1024, 256)`
- the image will be resized in (1024, 1024) if the argument is `int(1024)`
- the image will not be resized if the argment is `None`
- `num_workers` denoted the number of threads that requesting the content simultaneously, \
- `delay` denoted the time interval (in sec) between each request. The usage of `delay` avoids the `Error code 429: Too many requests`.
2. Edit `config.ini`
```ini
[default]
# Number of training epochs
epoch = 1000
# Save the model each n epochs
save_per_epoch = 20
# Sampling size of generating image each epoch
sample_size = 6
# The image size of generating immage
img_size = 1024
# The image size for random cropped
crop_size = 256
# The number of image channels (should be always RGB)
img_channel = 3
# Batch size
batch_size = 5
[path]
# The folder path of images of Domain A
domain_A_src = ./Data/domain_a/
# The folder path of images of Domain B
domain_B_src = ./Data/domain_b/
# The folder path of saved model
model_src = ./Pickle/
# The folder path for saving the generated samples
sample_src = ./Sample/ab/
[cuda]
# The device string (training on CPU if device=cpu or GPU if device=cuda)
device = cuda:0
[loss]
# The lambda of cycle consistency loss for Domain A
lambda_a = 10.0
# The lambda of cycle consistency loss for Domain B
lambda_b = 10.0
# The lambda of identity loss for both domains
lambda_idt = 0.5
[wgan]
# The clipping limit of weight (WGAN)
weight_clipping_limit = 0.01
```
## Run
`python main.py`
## Evaluation
**Images** or **Videos** can be rendered by `Stylizer` in `renderer.py`:
```python
from renderer import Stylizer
# Render a single image
file_name = './sample.jpg'
stylizer = Stylizer('./Pickle/cyclegan.pkl', domain='b', device='cuda:0')
stylizer.render_image(file_name, './Output/', num_workers=1)
# Render all images in folder
from glob import glob
file_name = glob('./Sample/*.jpg')
stylizer.render_image(file_name, './Output/', num_workers=8)
# Render video (Beta)
file_name = './sample.mp4'
stylizer.render_video(file_name, './Output/', num_workers=8)
```
## Result
Several filters are implemented in `rendered.py` for combining the original and rendered images
| Image | Description | Example |
|--------------------|----------------------------------------|---------------------------------------------------------------------------------------------------------------------|
| Original | The genuine image | 
|
| Rendered | The image which rendered by GAN | 
|
| **Slice** | Combine the images simply | 
|
| **Gradient Slice** | Combine the images with gradient blend | 
|