How I use Google Colab

Notes and recipes from how I use Google Colab day to day — setup, data handling, and package installs. All snippets are meant to be run inside Colab cells — lines starting with ! are shell commands run from the notebook, and %config is an IPython magic.

💡 Why Google Colab

The choice of Colab as the default substrate is deliberate — the principles below explain why.

🌐 Data lives near the compute

Datasets for serious work are large — archives routinely run into tens or hundreds of GB. Downloading them to a laptop is bandwidth-bound and wastes local disk. On Colab the data is pulled directly into a cloud VM that sits next to Google’s network, so the same download is faster and free of local storage pressure.

⚡ Free GPU / TPU runtime

Foundation models and most modern deep learning pipelines need a GPU. Colab provides one for free, with TPU and higher-tier GPUs available on paid plans. There is no driver install, no CUDA mismatch — the runtime comes pre-configured.

🧰 Zero local setup

Python, PyTorch, JAX, NumPy, SciPy, scikit-learn, and most of the scientific stack ship pre-installed. The notebook environment is identical across machines, which removes the “works on my laptop” problem and makes recipes reproducible.

🔗 Shareable by default

A Colab notebook is a single Drive file. It can be opened, copied, or shared with a link, which makes collaboration with colleagues, students, and reviewers trivial — no local environment to replicate.

🫧 Ephemeral, by design

Runtimes are wiped on disconnect. This forces every notebook to be re-runnable from a clean state, which is exactly the discipline you want for scientific work. Persistent state goes to Google Drive; everything else is reproducible from the cell.

📂 Mount Google Drive

Drive is the only storage that survives a runtime disconnect, so mounting it is usually the first cell:

from google.colab import drive
drive.mount('/content/drive')

🗄️ Filesystem

/content is the default working directory and where you land when a notebook starts. It typically has around 100+ GB of free disk space (varies by runtime type). Anything here is ephemeral and gets wiped when the runtime disconnects.
Home directory /root is writable too, but /content is the convention.
/tmp is also writable and ephemeral, but smaller. Good for temporary scratch files.
/content/drive/MyDrive is available after you mount Google Drive. Files here persist across sessions since they live in your Drive (subject to your Drive quota).

!df -h /content

Filesystem      Size  Used Avail Use% Mounted on
overlay         108G   22G   86G  21% /

💾 Copy data to personal Google Drive

Downloaded datasets live in ephemeral /content and vanish with the runtime. Sync them to Drive once so the next session can skip the download:

!rsync -ah --info=progress2 --stats /content/data/ /content/drive/MyDrive/data/

🔑 Set environment variables

Some datasets and APIs are password-protected; libraries usually pick the credentials up from environment variables, so set them at the top of the notebook before anything else runs:

import os
os.environ["DATASET_PASSWORD"] = "xxx"

📦 Install packages

From PyPI (recommended — the latest tagged release):

%pip install --upgrade "package[all]"

From the GitHub main branch (for unreleased changes):

%pip install --upgrade "package[all] @ git+https://github.com/org/repo.git#subdirectory=package-repo"

Why `--upgrade`

Colab images come with a large set of pre-installed packages — NumPy 2.0.2, for example. Without --upgrade, pip’s resolver leaves any already-installed package in place as long as it nominally satisfies a requirement, even when the package you are installing and its transitive deps (SciPy, scikit-learn) were built against a newer NumPy. The mismatch surfaces deep in the import chain:

ImportError: cannot import name '_center' from 'numpy._core.umath'

With --upgrade, pip honors the version constraints declared in the package’s pyproject.toml and bumps NumPy (and anything else) to a compatible version. ⚠️ After the install finishes, restart the Colab runtime once so the kernel picks up the new binaries.

✨ Retina plots

One line at the top of the notebook makes matplotlib figures render crisp on high-DPI displays:

%config InlineBackend.figure_format = 'retina'

🧩 VS Code Colab extension

The Google Colab VS Code extension connects VS Code to Colab’s hosted Jupyter runtime, giving access to free GPUs and TPUs without leaving the editor.

🐛 Plotly charts not rendering

fig.show() can silently render nothing in the VS Code Colab extension. Plotly’s auto-detection picks its colab renderer whenever the google.colab module is importable — even outside web Colab — and that renderer’s output only displays in the Colab web app (plotly.py#5471).

Two fixes:

Upgrade plotly — from 6.8.0 detection uses the COLAB_NOTEBOOK_ID environment variable (set only in web Colab) instead of the module import (plotly.py#5473).
```
%pip install -U plotly
```
Or set the renderer explicitly at the top of the notebook (works on any version):
```
import plotly.io as pio
pio.renderers.default = "vscode"
```

This is a living list — I’ll keep adding recipes as my workflow evolves. 🌱