- Develop an algorithm while debugging visually with plots, while checking robustness & continuity to parameters change.
- Magically create a graphical interface to easily demonstrate a concept or simply tune your algorithm.
❤️ You do not need to learn anything about making a graphical user interface (GUI) ❤️
Science notebook | Toddler DIY Jukebox on a raspberry Pi |
---|---|
![]() |
![]() |
Sliders are added automatically in your jupyter notebook. This works on Google Collab and the code takes about 40 lines of code. No Widgets, event handlers or matplotlib knowledge required. | Plays some music when you touch the icon. Images can be generated using the OpenAI Dall-E API helpers . Caption added through the title mechanism. Music samples generated by prompting MusicGen |
Demo notebook on collab | jukebox.py demo code |
pip3 install interactive-pipe
- Demonstrate concepts by interacting with curves / images.
- Easy integration in Jupyter Notebooks (popular with Google Collab)
- You can also use the declarative nature of interactive pipe to make a graphical interface in a few lines of codes.
- For instance, it is possible to code a jukebox for a toddler on a RaspberryPi.
- While prototyping an algorithm or testing a neural network, you may be interested in making small experiments with visual checks. Instead of making a draft quick & dirty code that you'll never keep, you can use interactive pipe to show your team how your library works. A visual demo is always good, it shows that the algorithm is not buggy if anyone can play with it.
- Tune your algorithms with a graphical interface and save your parameters for later batch processing.
- Ready to batch under the hood, the processing engine can be ran without GUI (therefore allowing to use the same code for tuning & batch processing if needed).
- Do not spoil your production code with a huge amount of graphical interface code, keep your algorithms library untouched and simply decorate it.
Version 0.5.9
- Modular multi-image processing filters
- Declarative: Easily make graphical user interface without having to learn anything about pyQt or matplotlib
- Support in jupyter notebooks
- Tuning sliders & check buttons with a GUI
- Cache intermediate results in RAM for much faster processing
KeyboardControl
: no slider on UI but exactly the same internal mechanism, update on key press.- Support Curve plots (2D signals)
- 🔜 Scientific visual debugging: Display both colored images, heatmaps & graphs
- custom events on specific key press
- Display the execution graph of the pipeline
- thirdparty/music Play audio (Qt backend only). Play songs on spotify (linux only) when the spotify app is running.
- thirdparty/images_openai_api Generate images from prompt using OpenAI API image generation DALL-E Model (:dollar: paid service ~ 2cents/image)
Shortcuts while using the GUI (QT & matplotlib backends)
F1
to show the help shortcuts in the terminalF11
toggle fullscreen modeW
to write full resolution image to diskR
to reset parametersI
to print parameters dictionary in the command lineE
to export parameters dictionary to a yaml fileO
to import parameters dictionary from a yaml file (sliders will update)G
to export a pipeline diagram for your interactive pipe (requires graphviz)
- supported backends
- 🆗
gui='qt'
pyQt/pySide - 🆗
gui='mpl'
matplotlib - 🆗
gui='nb'
ipywidget for jupyter notebooks
- 🆗
- tested platforms
- 🆗 Linux (Ubuntu / KDE Neon)
- 🆗 RapsberryPi
- 🆗 On google collab (use
gui='nb'
)
GUI | Pipeline |
---|---|
![]() |
![]() |
ISP means image signal processor
Since ipywidgets in notebooks are supported, the tutorial is also available in a google collab notebook
Let's define 3 image processing very basic filters exposure
, black_and_white
& blend
.
By design:
- image buffers inputs are arguments
- keyword arguments are the parameters which can be later turned into interactive widgets.
- output buffers are simply returned like you'd do in a regular function.
We use the @interactive()
wrapper which will turn each keyword parameters initialized to a tuple/list into a graphical interactive widgets (slider, tick box, dropdown men).
The syntax to turn keyword arguments into sliders is pretty simple (default, [min, max], name)
will turn into a float slider for instance.
Finally, we need to the glue to combo these filters. This is where the sample_pipeline function comes in.
By decorating it with @interactive_pipeline(gui="qt")
, calling this function will magically turn into a GUI powered image processing pipeline.
from interactive_pipe import interactive, interactive_pipeline
import numpy as np
@interactive()
def exposure(img, coeff = (1., [0.5, 2.], "exposure"), bias=(0., [-0.2, 0.2])):
'''Applies a multiplication by coeff & adds a constant bias to the image'''
# In the GUI, the coeff will be labelled as "exposure".
# As the default tuple provided to bias does not end up with a string,
# the widget label will be "bias", simply named after the keyword arg.
return img*coeff + bias
@interactive()
def black_and_white(img, bnw=(True, "black and white")):
'''Averages the 3 color channels (Black & White) if bnw=True
'''
# Special mention for booleans: using a tuple like (True,) allows creating the tick box.
return np.repeat(np.expand_dims(np.average(img, axis=-1), -1), img.shape[-1], axis=-1) if bnw else img
@interactive()
def blend(img0, img1, blend_coeff=(0.5, [0., 1.])):
'''Blends between two image.
- when blend_coeff=0 -> image 0 [slider to the left ]
- when blend_coeff=1 -> image 1 [slider to the right]
'''
return (1-blend_coeff)*img0+ blend_coeff*img1
# you can change the backend to mpl instead of Qt here.
@interactive_pipeline(gui="qt", size="fullscreen")
def sample_pipeline(input_image):
exposed = exposure(input_image)
bnw_image = black_and_white(input_image)
blended = blend(exposed, bnw_image)
return exposed, blended, bnw_image
if __name__ == '__main__':
input_image = np.array([0., 0.5, 0.8])*np.ones((256, 512, 3))
sample_pipeline(input_image)
❤️ This code shall display you a GUI with three images. The middle one is the result of the blend
Notes:
- If you write
def blend(img0, img1, blend_coeff=0.5):
, blend_coeff will simply not be a slider on the GUI no more. - If you write
blend_coeff=[0., 1.]
, blend_coeff will be a slider initalized to 0.5 - If you write
bnw=(True, "black and white", "k")
, the checkbox will disappear and be replaced by a keypress event (pressk
to enable/disable black & white)
from interactive_pipe import interactive, interactive_pipeline
import numpy as np
COLOR_DICT = {"red": [1., 0., 0.], "green": [0., 1.,0.], "blue": [0., 0., 1.], "gray": [0.5, 0.5, 0.5]}
@interactive()
def generate_flat_colored_image(color_choice=["red", "green", "blue", "gray"], global_params={}):
'''Generate a constant colorful image
'''
flat_array = np.array(COLOR_DICT.get(color_choice)) * np.ones((64, 64, 3))
global_params["avg"] = np.average(flat_array)
return flat_array
- Note that you can also create filters which take no inputs and simply "generate" images.
- The
color_choice
list will be turnt into a nice dropdown menu. Default value here will be red as this is the first element of the list!
💡 Can filters communicate together?
Yes, using the special keyword argument global_params={}
.
- Check carefully how we stored the image average of the flat image in global_params.
- This value will be available to other filters.
special_image_slice
is going to use that value to set the half bottom image to dark in case the average is high.
@interactive()
def special_image_slice(img, global_params={}):
if global_params["avg"] > 0.4:
out_img[out_img.shape[0]//2:, ...] = 0.
return out_img
@interactive()
def switch_image(img1, img2, img3, image_index=(0, [0, 2], None, ["pagedown", "pageup", True])):
'''Switch between 3 images
'''
return [img1, img2, img3][image_index]
Note that you can create a filter to switch between several images. In ["pagedown", "pageup", True]
, True means that the image_index will wrap around. (it will return to 0 as soon as it goes above the maximum value of 2).
@interactive()
def black_top_image_slice(img, top_slice_black=(True, "special", "k"), global_params={}):
out_img = img.copy()
if top_slice_black:
out_img[:out_img.shape[0]//2, ...] = 0.
return out_img
@interactive_pipeline(gui="qt", size="fullscreen")
def sample_pipeline_generated_image():
flat_img = generate_flat_colored_image()
top_slice_modified = black_top_image_slice(flat_img)
bottom_slice_modified_image = special_image_slice(flat_img)
chosen = switch_image(flat_img, top_slice_modified, bottom_slice_modified_image)
return chosen
if __name__ == '__main__':
sample_pipeline_generated_image()
- Interactive pipe was initially developped by Balthazar Neveu as part of the irdrone project based on matplotlib.
- Later, more contributions were also made by Giuseppe Moschetti and Sylvain Leroy.
- Summer 2023: rewriting the whole core and supporting several graphical backends!
🐛 Want to contribute or interested in adding new features? Enter a new Github issue
🎁 Want to dig into the code? Take a look at code_architecture.md
- Backport previous features
- Image class support in interactive pipe (Heatmaps/Float images)
- Advanced feature
- Webcam based "slider" for dropdown menu (like "elephant" will trigget if an elephant is magically detected on the webcam)
- Animations/While loops/Video source (Time slider)
- Exploratory backends