Skip to content

title: Google Colab description: Free hosted notebooks — Google Colab essentials: GPUs/TPUs, drive & data access, secrets, magics, and sharing. category: 25-IDEs-Editors tags: [colab, notebooks, gpu, cloud] status: ready last_updated: 2026-09-06

Google Colab

Free Jupyter notebooks with hosted runtimes — including GPU and TPU access. Zero setup, shareable like a Google Doc, and good enough for many ML experiments and tutorials.

Table of Contents

Installation / Setup

  1. Go to colab.research.google.com → sign in.
  2. File → New notebook, or File → Upload notebook / open any GitHub .ipynb via URL.

Choose hardware: Runtime → Change runtime type:

Runtime Use for
CPU small data, pandas, plotting
T4 GPU deep learning (free tier)
A100 GPU larger training (paid tiers)
TPU specific TPU-compatible workloads 
import torch
print(torch.cuda.is_available())     # True on a GPU runtime
!nvidia-smi                          # see the GPU

Core Syntax

Environment & pre-installed packages

Colab ships with pandas, NumPy, scikit-learn, matplotlib, PyTorch, TensorFlow, and more. Install anything else per session:

!pip install xgboost           # shell commands with !
!pip install -q "transformers>=4.30"

Every session starts fresh — reinstalls and re-mounts are normal.

Data access

# 1. Upload a file (small; per-session)
from google.colab import files
files.upload()                          # browser dialog

# 2. Mount Google Drive (authorize once per session)
from google.colab import drive
drive.mount("/content/drive")
!ls /content/drive/MyDrive/data

# 3. Pull from the web (most reproducible)
!wget -q https://example.com/data.csv
!kaggle datasets download -d user/dataset   # needs kaggle.json

Secrets (API keys)

from google.colab import userdata
key = userdata.get("OPENAI_API_KEY")    # set via the key icon in the left panel

Never paste secrets into code cells — use the secrets manager so shared notebooks stay clean.

Common Operations

Save outputs back to Drive

df.to_csv("/content/drive/MyDrive/results/model_output.csv", index=False)
model.save("/content/drive/MyDrive/models/model.pt")

# And download to your computer
from google.colab import files
files.download("results/model_output.csv")

Notebook magics (same as Jupyter, plus):

%tensorflow_version 2.x        # legacy pinning (usually unnecessary now)
%%capture                      # silence noisy installs
!nvidia-smi                    # GPU status
!free -h                       # RAM status

Sharing & GitHub

  • Share: blue "Share" button → anyone with link (like Docs). Readers can copy.
  • Save a copy: File → Save a copy in Drive / as GitHub gist.
  • Open from GitHub: colab.research.google.com/github/<owner>/<repo>/blob/<branch>/<path>.ipynb
  • Download: File → Download .ipynb / .py.

Save to GitHub

# File -> Save a copy in GitHub — needs a linked GitHub account
# then: git clone + git pull like any repo (see git.md)

Practical Examples

1. Free GPU training

# Runtime -> Change runtime type -> T4 GPU
import torch
x = torch.randn(1000, 1000, device="cuda")
y = (x @ x.T).sum()
print(y.item())                    # works on GPU

2. Reproducible Colab start (a standard cell 1)

#@title Setup
!pip install -q pandas numpy matplotlib seaborn scikit-learn
!git clone https://github.com/you/project.git   # pull your own code
%cd project

3. Use Drive as a dataset store

from google.colab import drive
drive.mount("/content/drive")
import pandas as pd
df = pd.read_csv("/content/drive/MyDrive/datasets/train.csv")

Common Patterns

# Pin versions to avoid session drift
!pip install -q "pandas==2.2.2" "scikit-learn==1.4.0"

# Run long trainings with checkpointing to Drive (session may time out!)
model.train()
torch.save(model.state_dict(), "/content/drive/MyDrive/checkpoints/last.pt")

# Keep API keys out of the notebook via userdata

Best Practices

  • Treat sessions as ephemeral — files outside Drive/GitHub vanish when the runtime resets.
  • Pin package versions so today's notebook works next month.
  • Put data in Drive or GitHub, never re-upload by hand every session.
  • Use the secrets manager for keys; never hardcode or print them.
  • Save checkpoints to Drive for any run > ~30 minutes (free runtimes disconnect).
  • Start from GitHub URLs for versioned, shareable notebooks.
  • Commit the .py export (File → Download .py) alongside the .ipynb for review.
  • Respect free-tier limits: disconnect idle runtimes and keep Colab for prototyping, not production.

Common Mistakes

Mistake Consequence Fix
Forgetting to enable GPU Slow "deep learning" Runtime → Change runtime type → GPU
Uploading data each session Wasted time, lost state Drive mount / GitHub clone
Hardcoding API keys in cells Leaked secrets on share userdata.get(...)
No version pins Notebook breaks months later Pin with ==
Untracked long training Work lost on disconnect Checkpoint to Drive
%cd inside repo Path confusion Use absolute /content/... paths
Downloading huge files to /content Runtime disk fills Stream from Drive/cloud storage

Troubleshooting

Symptom Cause Fix
CUDA out of memory Batch too big for T4 (16 GB) Smaller batch, float16, gradient accumulation
Session expired mid-run ~12h idle/max limits Checkpoints to Drive; shorter runs
ModuleNotFoundError after restart Fresh runtime Re-run the setup cell (!pip install)
Drive permission error Mount authorization Re-run drive.mount and accept
Kaggle API auth fails kaggle.json location Place in ~/.kaggle/; chmod 600
Slow data loading Reading from Drive repeatedly Copy to /content first, or use Parquet

Official Documentation

Last Updated

2026-09-06 — Initial version.