Skip to content

Execute several jupyter cells simultaneously with beautiful output. Do not waste time waiting

License

Notifications You must be signed in to change notification settings

alexdremov/igogo

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

30 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

igogo 🐎🏎️

Execute several jupyter cells at the same time

Have you ever just sited and watched a long-running jupyter cell? Now, you can continue to work in the same notebook freely

gitdemo.mov

Use Cases

  1. You have a long-running cell, and you need to check something. You can just start the second cell without interrupting a long-running cell.

    Example: you run a machine learning train loop and want to immediately save the model's weights or check metrics. With igogo you can do so without interrupting the training.

  2. If you need to compare the score of some function with different parameters, you can run several functions at the same time and monitor results.

    Example: you have several sets of hyperparameters and want to compare them. You can start training two models, monitoring two loss graphs at the same time.

  3. Process data in chunks. Check processed data for validity

    Example: you do data processing in steps. With igogo you can execute several steps at the same time and process data from the first processing step in the second processing step in chunks. Also, you can quickly check that the first step produces the correct results

Install

Igogo is available through PyPi:

pip install igogo

Wait, isn't it just a background job? No.

  • No multithreading, no data races, no locks. You can freely operate with your notebook variables without the risk of corrupting them.
  • Beautiful output. When several cells execute in parallel, all printed data is displayed in the corresponding cell's output. No more twisted and messed out concurrent outputs.
  • Easily cancel jobs, wait for completion, and start the new ones.
  • Control execution of jobs through widgets.

Usage

At the core of igogo is collaborative execution. Jobs need to explicitly allow other jobs to execute through igogo.yielder(). Mind that regular cells also represent a job.

Placing igogo.yielder() in code that is not executed in igogo job is not a mistake. It will return immediately. So, you don't need to care about keeping igogo.yielder() only in igogo jobs. You can place it anywhere

To start an igogo job, you can use %%igogo cell magic or function decorator.

import igogo

@igogo.job
def hello_world(name):
    for i in range(3):
        print("Hello, world from", name)
        
        # allows other jobs to run while asleep
        # also can be `igogo.yielder()`
        igogo.sleep(1)  
    return name

Call function as usual to start a job:

hello_world('igogo'), hello_world('other igogo');
gitdemo.mov

Configure Jobs

Decorator @igogo.job has several useful parameters.

  • kind
    Allows to set how to render output. Possible options: text, markdown, html Default: text
  • displays
    As igogo job modify already executed cell, it needs to have spare placeholders for rich output. This parameter specifies how many spare displays to spawn. Default: 1
  • name
    User-friendly name of igogo job.
  • warn_rewrite
    Should warn rewriting older displays? Default: True
  • auto_display_figures
    Should display pyplot figures created inside igogo automatically? Default: True

Markdown example:

gitdemo.mov

Display Additional Data

Pyplot figures will be automatically displayed in igogo cell.

You can also use igogo.display inside a job to display any other content or several figures. Mind that displays must be pre-allocated by specifying displays number in igogo.job(displays=...)

import numpy as np
import matplotlib.pyplot as plt
import igogo

def experiment(name, f, i):
     x = np.linspace(0, i / 10, 100)
     fig = plt.figure()
     plt.plot(
         x,
         f(x)
     )
     plt.gca().set_title(name)
     igogo.display(fig)
     
     fig = plt.figure()
     plt.scatter(
         x,
         f(x)
     )
     plt.gca().set_title(name)
     igogo.display(fig)
     igogo.sleep(0.05)

As noted in "Configure jobs" section, igogo jobs have limited number of displays. If you try to display more objects than job has, warning will be shown and the oldest displays will be overwritten.

Cell Magic

The same way with %%igogo:

%load_ext igogo
%%igogo
name = 'igogo'
for i in range(3):
     print("Hello, world from", name)
     igogo.sleep(1)

Widgets

All executed igogo jobs spawn a widget that allows to kill them. Jobs are not affected by KeyboardInterrupt

Killing Jobs

Apart from killing through widgets, igogo jobs can be killed programmatically.

  • igogo.stop()
    Can be called inside igogo job to kill itself.
  • igogo.stop_all()
    Stops all running igogo jobs
  • igogo.stop_latest()
    Stops the latest igogo job. Can be executed several times.
  • igogo.stop_by_cell_id(cell_id)
    Kills all jobs that were launched in cell with cell_id (aka [5], cell_id=5).

Also, you can stop jobs of one specific function.

  • hello_world.stop_all()
    Stops all igogo jobs created by hello_world()

Supported Clients

Currently, igogo runs fully correct on:

  • Jupyter Lab
  • Jupyter

Runs but has problems with output from igogo jobs. Jobs are executed, but there could be problems with widgets and output:

  • VSCode. For some reason it does not update display data. Therefore, no output is produced.
  • DataSpell. It displays [object Object] and not output.
  • Colab. It does not support updating content of executed cells

More Examples

Check out pretty notebooks


Train model and check metrics

Screen.Recording.2023-03-24.at.14.14.03.mov

Also, you can modify training parameters, freeze/unfreeze layers, switch datasets, etc. All you need is to place igogo.yielder() in train loop.

Process data and montitor execution

import igogo
import numpy as np
from tqdm.auto import tqdm
%load_ext igogo

raw_data = np.random.randn(100000, 100)
result = []
def row_processor(row):
    return np.mean(row)
%%igogo
for i in tqdm(range(len(raw_data))):
    result.append(row_processor(raw_data[i]))
    igogo.yielder()
result[-1]

Process data in chunks

import igogo
import numpy as np
from tqdm.auto import tqdm
%load_ext igogo

raw_data = np.random.randn(5000000, 100)

igogo_yield_freq = 32
igogo_first_step_cache = []

result = []
%%igogo

for i in tqdm(range(len(raw_data))):
    processed = np.log(raw_data[i] * raw_data[i])
    igogo_first_step_cache.append(processed)
    
    if i > 0 and i % igogo_yield_freq == 0:
        igogo.yielder()  # allow other jobs to execute
%%igogo

for i in tqdm(range(len(raw_data))):
    while i >= len(igogo_first_step_cache):  # wait for producer to process data
        igogo.yielder()
    
    result.append(np.mean(igogo_first_step_cache[i]))
    
gitdemo.mov