🧱 Define processing blocks as regular functions.
Define each atomic processing block as a regular python function.
2 kinds of elements are fed into each processing block:
➡️ Arguments are signal buffers (numpy arrays).
➡️ Keyword arguments are default parameters (float, int, bool, str).
↪️ Function can return several signal buffers (numpy arrays).
import numpy as np
# --- Define processing blocks ---
def gen_color(green=1., decay=2):
t = np.linspace(0., 1., 256)
X, Y = np.meshgrid(np.exp(-t*decay), t)
return np.stack([X, green*Y, Y[::-1]], -1)
def flip_img(img, flip=True):
out = img[::-1] if flip else img.copy()
return out, img[:, ::-1]
def amplify(img, amp="L"):
ratio = {"S":4, "M":3, "L": 2, XL":1}[amp]
return 1/ratio*img
🔗 Define a pipeline and output layout.
🔀 Chain processing blocks to define a pipeline.
🪟 Output layout is specified by the return statement.
return [A, B, C] |
return [[A, B], [C, D]] |
return [[A], [B]] |
|---|---|---|
# --- Define pipeline ---
def pipe():
inp = gen_color()
out_flip, out_mirror = flip_img(inp)
out = amplify(out_mirror)
return [inp, out_flip, out]
# Display 3 images side by side
🕹️ Add controls to processing blocks.
🎚️ By decorating a processing block, Interactive pipe adds controls.
👉 Controls are automatically created for each keyword argument.
string |
boolean |
integer |
float |
|---|---|---|---|
|
Amplify |
Flip |
Color Decay 3 |
Green amount 0.5 |
import interactive_pipe as ip
# --- Decorate processing blocks ---
ip.interactive(
green=(1., [0., 1.], "Green amount"),
decay=(2 , [0 , 10], "Color decay")
)(gen_color)
ip.interactive(
amp=("L", ["S", "M", "L", "XL"], "Amplify")
)(amplify)
ip.interactive(flip=(True, "Flip"))(flip_img)
▶️ Launch interactive pipe.
🎯 Interactive Pipe deals with several graphical interfaces backends.
💫 No code changes needed: Same GUI, no matter the backend .
# --- Decorate pipeline & run ---
ip.interactive_pipeline(gui="qt")(pipe)()
🎁 Multiple backends support
Interactive Pipe offers a range of options for displaying your code, no code changes needed.
💻 PyQT
✅ Dynamic control of output layouts.
✅ Keyboard shortcuts, Full screen.
✅ Audio Playback.
Sample python code on github »
⌨️ Ideal during development, debug sessions and code reviews.
🤗 Gradio
✅ Share your pipeline with others.
✅ Easily host in Hugging Face spaces.
✅ Runs in a browser.
Example: Basic Image editing »
🧪 Ideal to share ideas inside a scientific team, ready for demos.
📒 Notebooks
✅ Annotate with markdown and equations.
✅ Easily host on Google Colab.
✅ Based on Jupyter IPywidget.
Example: Signal processing notebook » 
🧑🎓 Ideal for students projects (computer vision, machine learning).
🚀 Features
📈 📷 Images and Curves are supported by Interactive Pipe.
🎶 🎧 Live audio playback is supported using Qt and Gradio GUI.
🏍️ 💾 A cache mechanism allows skipping computation if the inputs and sliders have not changed.
✔️ Motivations
✊ Empower engineers and developers to showcase their expertise.
🔍 Visual debugging made intuitive: Traditional debugging tools often fall short for image and audio algorithms, where simple breakpoints aren't enough. Visual tools provide clearer insights.
🧑⚖️ Elevate code reviews: Go beyond unit tests—demonstrate your code's functionality visually, making it easier to convince peers that it works as expected.
🏭 Seamless production integration: Maintain a clean separation between production code and graphical interfaces. By decorating processing blocks in a dedicated file, your pipeline remain functional whether run in a GUI or headlessly via the command line.
🔬 Built by scientists, for scientists. Designed with research-driven workflows in mind.
💥 Real use cases
🧸 Tutorial is written as a comprehensive toy example (to understand the concept).
🏭 But Interactive Pipe can be used on real use cases.
🚀 Check by yourself with the examples provided on Image, Audio and Machine learning.
🎓 Tutorials
🔗 Many tutorials are available from basic to advanced features of interactive pipe.
🙏 Acknowledgement
- Main contributor: Balthazar Neveu
- Additional contributors: Emmanuel Chaboud, Giuseppe Moschetti, Sylvain Leroy
- Additional testers: Jamy Lafenetre, Mathilde Dupouy, Matthieu Dinot