Participate
To get started and submit your first model, you will need to pass through the following steps.
Register
Creating an account on the CrunchDAO platform will allow you to be identified and get access to the competition dataset. Follow the link below to join the competition.
Submit
Two distinct formats of submission are accepted for the competitions:
Jupyter Notebook (
.ipynb
), which is a self-contained version of the codePython Script (
.py
), which allows more flexibility and to split the code into multiple files
Jupyter Notebook
Notebook users can use the Quickstarters provided by CrunchDAO to quickly experiment with a working solution that users can tinker with.
Setting the Environment
Before trying to execute any cell, users must set up their environment by copying the command available on the competition page:

Run the commands to set up your environment and download the data to be ready to go:
# Upgrade the Crunch-CLI to the latest version
%pip install crunch-cli --upgrade
# Authenticates yourself, it will downloads your last submission and the data
!crunch setup <competition name> <model name> --notebook --token <token>
Users can now load the data locally:
# Load the notebook, run me once
import crunch
crunch = crunch.load_notebook()
# Load the data, re-run me if you corrupt the dataframes
X_train, y_train, X_test = crunch.load_data()
Local Testing
When users are satisfied with their work, they can easily test their implementation:
# Run a local test
crunch.test()
Submitting your Notebook
After testing the code, users need to have access to the .ipynb
file.
If you are on Google Colab:
File
>Download
>Download .ipynb
If you are on Kaggle:
File
>Download Notebook
If you are on Jupyter Lab:
File
>Download
Then submit on the Submit a Notebook page:

Some model files can also be uploaded along with the notebook, which will be stored in the resources/
directory. Read more about the file selection dialog.
Global variables
If you want to use global variables in your notebook, put them in a class, this will improves the readability of your code:
class Constants:
TRAIN_DEPTH = 42
IMPORTANT_FEATURES = [ "a", "b", "c" ]
def infer():
print(Constants.TRAIN_DEPTH)
# 42
Automatic line commenting
The notebook is automatically converted into a Python script that only includes the functions, imports, and classes.
Everything else is commented out to prevent side effects when your code is loaded into the cloud environment. (e.g. when you're exploring the data, debugging your algorithm, or doing visualizating using Matplotlib, etc.)
You can prevent this behavior by using special comments to tell the system to keep part of your code:
To start a section that you want to keep, write:
@crunch/keep:on
To end the section, write:
@crunch/keep:off
# @crunch/keep:on
# keep global initialization
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# keep constants
TRAIN_DEPTH = 42
IMPORTANT_FEATURES = [ "a", "b", "c" ]
# @crunch/keep:off
# this will be ignored
x, y = crunch.load_data()
def train(...):
...
The result will be:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
TRAIN_DEPTH = 42
IMPORTANT_FEATURES = [ "a", "b", "c" ]
#x, y = crunch.load_data()
def train(...):
...
The command does not affect comments, functions, classes, or imports.
Specifying package versions
Since submitting a notebook does not include a requirements.txt
, users can instead specify the version of a package using import-level requirement specifiers in a comment on the same line.
# Valid statements
import pandas # == 1.3
import sklearn # >= 1.2, < 2.0
import tqdm # [foo, bar]
import scikit # ~= 1.4.2
from requests import Session # == 1.5
Specifying multiple times will cause the submission to be rejected if they are different.
# Inconsistant versions will be rejected
import pandas # == 1.3
import pandas # == 1.5
Specifying versions on standard libraries does nothing (but they will still be rejected if there is an inconsistent version).
# Will be ignored
import os # == 1.3
import sys # == 1.5
If an optional dependency is required for the code to work properly, an import statement must be added, even if the code does not use it directly.
import castle.algorithms
# Keep me, I am needed by castle
import torch
It is possible for multiple import names to resolve to different libraries on PyPI. If this happens, you must specify which one you want. If you do not want a specific version, you can use @latest
, as without this, we cannot distinguish between commented code and version specifiers.
# Prefer https://pypi.org/project/EMD-signal/
import pyemd # EMD-signal @latest
# Prefer https://pypi.org/project/pyemd/
import pyemd # pyemd @latest
Embed Files
Additional files can be embedded in cells to be submitted with the Notebook. In order for the system to recognize a cell as an Embed File, the following syntax must be followed:
---
file: <file_name>.md
---
<!-- File content goes here -->
Lorem ipsum dolor sit amet, consectetur adipiscing elit.
Aenean rutrum condimentum ornare.
Submitting multiple cells with the same file name will be rejected.
While the focus is on Markdown files, any text file will be accepted. Including but not limited to: .txt
, .yaml
, .json
, ...
Python Script
Script users can use the Quickstarters provided by CrunchDAO to know what the structure should be.
A mandatory main.py is required to have both functions (train
and infer
) in order for your code to run properly.
Setting the Environment
Before starting to work, users must setup their environment which will be similar to a git repository.

Run the commands to set up your environment and download the data to be ready to go:
# Upgrade the Crunch-CLI to the latest version
$ pip install crunch-cli --upgrade
# Authenticates yourself, it will downloads your last submission and the data
$ crunch setup <competition name> <model name> --token <token> [directory]
# Change the directory to the configured environment
$ cd <directory>
Read more about how setup tokens work and why it is safe to (accidentally) "leak" them.
Directory Layouts
# Example of a folder structure.
# The data files may change depending on the competition.
.
├── data/
│  ├── X_test.parquet
│  ├── X_train.parquet
│  └── y_train.parquet
├── main.py
├── requirements.txt
└── resources/
  └── model.joblib
data/
Directory containing the data of the competition, should never be modified by the user. Always kept up to date by the CLI.
main.py
Code entry point. Must contain the train()
and infer()
function. Can import other files if necessary. Learn more...
requirements.txt
List of packages used by your code. They are installed before your code is invoked.
resources/
Directory where your model should be stored. The content is persisted across runs during the transition between the Submission and Out-of-Sample phases.
Local Testing
When users are satisfied with their work, they can easily test their implementation:
# Run a local test using a shell command
$ crunch test
Pushing your Code
After the code has been tested, the submission needs to be uploaded to the server.
The message is optional and is just a label for users to know what they did.
$ crunch push --message "hello world"
Package version freezes
Before submitting, the CLI does a pip freeze
in the background to find out what version you are using locally. These versions are then used to freeze the requirements.txt
on the server.
This is to ensure that if your code works locally with the exact same versions, it should theoretically work the same on the server.
However, this behavior may result in your packages not being installed because:
you are using custom/externally installed package versions,
you have installed some packages by force, but PyPI considers them incompatible,
the architecture is different and the package is platform-specific.
If you want to disable this behavior, use the --no-pip-freeze
flag.
$ crunch push --no-pip-freeze --message "hello world"
Hybrid
For some complex setups, users may need to use the CLI to submit a Jupyter Notebook. This can happen if they want to submit with a large pre-trained model, or they want to include non-PyPI packages.
It will be very similar to the Python Script setup:
Setting the Environment, like for a Python Script.
Remove the
main.py
.Move your notebook to the project directory and name it
main.ipynb
.
If done correctly, before each crunch push, the CLI will first convert the notebook to a script file before sending it.
Package version specifiers will not work.
The requirements.txt
file must be updated manually.
Files
If you do not want to use the CLI and did not use a notebook to write your code, you can submit files directly. This is an advanced feature that requires preparation.
The directory layout must be the same as the CLI directory layout.
File Selection Dialog
To add files you can:
select multiple files by clicking on the "Add file(s)" button
or select the contents of an entire directory by clicking on the "Add directory" button

Once added, files can be disabled if you add too many, or renamed if the name is incorrect.

You need to submit the requirements.txt
yourself.
Setup Tokens
The site generates new tokens every minute, and each token can only be used once within a 3-minute timeframe.
This prevents any problems if your token is accidentally shared, as it will likely have already been used or expired. Even the team shares their expired tokens in Quickstarters.
This token allows the CLI to download the data and submit your submission on your behalf.
Run
Checking your submission

The system parses your work to retrieve the code of the interface functions (train()
and infer()
) and their dependencies. By clicking on the right arrow, you can access the contents of your submission.

Running in the Cloud
Once you've submitted, it's time to make sure your model can run in the cloud environment. Click on a submission and then click the Run in the Cloud button.

Your code is fed a standard epoch of data and the system simulates an inference.

Debugging with the logs
If your run crashes or you want to better understand how your code behaved, you can review the logs.


Due to abuse, only the first 1,500 lines of a user's code logs will be displayed.
Last updated