Google Colab lets you run Python notebooks in the browser with zero setup, free GPUs on the entry tier, and easy sharing. If you're new to notebooks, session limits, or bringing in your own data, this guide gives you a clean, repeatable setup that works on both the Free and Pro plans.
Before you start
- A Google account you can use for Drive storage.
- An up‑to‑date Chromium‑based browser or Safari.
- Stable internet; Colab runs on remote VMs that can disconnect if idle.
- Optional: a small dataset in Drive (CSV/Parquet/images) to practice with.
- Optional: budget for Pro/Pro+ if you want faster GPUs, longer sessions, and higher priority compute.
1) Create your first notebook
Go to Colab, click New Notebook. You’ll see a code cell and a text cell (Markdown). Click the code cell and press Shift+Enter to run it. Add a text cell for notes with Ctrl+M then B (Mac: Cmd+M, then B).
2) Pick your hardware (CPU, GPU, or TPU)
Open Runtime → Change runtime type and choose Hardware accelerator: None (CPU), GPU, or TPU. Free users get variable availability; Pro tiers offer higher priority. After switching, the environment restarts.
Pro tip: If you need CUDA, verify it with !nvidia-smi right after selecting GPU. For TPU, use TensorFlow or JAX builds that support it.
3) Understand cells and execution
Each notebook is a linear list of cells. Code cells run top‑to‑bottom and share the same Python kernel state. Use Shift+Enter to run a cell and move, or click the play icon. Restart the runtime anytime via Runtime → Restart session to clear memory and variables.
4) Install the packages you need
Colab ships with many libraries, but you can add more per‑session. Use a leading exclamation point to call shell commands from cells:
!pip install polars scikit-learn plotly
!apt-get -qq update && apt-get -qq install -y libspatialindex-devNote: Installs live inside the temporary VM and disappear when the session ends. Save your requirements.txt to Drive and re‑install at the top of the notebook for reproducible runs.
5) Save your work reliably
Colab autosaves notebooks to your Drive. Use File → Save a copy in Drive to duplicate, or File → Save a copy to GitHub to back up in a repo. Export with File → Download to get .ipynb, .py, or .html.
Why use it: You keep a clean, versioned copy and avoid losing code when sessions recycle.
6) Bring in your data
You can upload files, mount Google Drive, or fetch from the web. For Drive, run:
from google.colab import drive
drive.mount('/content/drive') # Grants access to your Drive
import pandas as pd
df = pd.read_csv('/content/drive/MyDrive/data/sample.csv')
df.head()Pro tip: Keep a data/ folder at MyDrive/<project>/data and reference it with absolute paths so teammates can run your notebook without edits.
7) Keep project files organized
Create a minimal folder layout in Drive: data/, notebooks/, models/, outputs/. Save models and artifacts explicitly to Drive so they persist after the VM shuts down.
8) Watch your RAM, disk, and time
Click the RAM/Disk meter to see resources. Free tier sessions are time‑limited and can disconnect if idle or resource‑constrained. Pro tiers extend runtime length and improve stability, but long‑running jobs should still checkpoint outputs to Drive frequently.
9) Share and collaborate
Use the Share button to add viewers, commenters, or editors. Comments work like Docs. For safer collaboration, make an “execution” copy per teammate to avoid clobbering each other’s state.
10) Use forms and parameters
Turn cells into mini UIs with form fields. Add a line like #@param {type:"slider", min:1, max:128, step:1} above a variable to expose a control. This helps non‑technical teammates change inputs without touching code.
11) Connect to Git and datasets
You can clone repos with !git clone, pull updates with !git pull, and push back using a personal access token. For public datasets, download with !wget or Python requests. Always write the final copy to Drive, not just the VM.
12) Export results cleanly
Save outputs that matter: charts as .png, tables to CSV/Parquet, trained models to Drive. Then export your notebook as .ipynb for future editing and a .html snapshot for sharing.
13) Troubleshoot common issues
- “Session crashed” after install or heavy training: Restart the runtime, re‑order installs to the top, and checkpoint outputs to Drive.
- Can’t get a GPU: Switch to None temporarily, try again later, or upgrade to a paid tier for higher priority.
- Files disappear after reconnect: Anything outside Drive is ephemeral. Always write to
/content/drive/.... - Slow installs every run: Cache wheels in Drive and install from there, or build a setup cell that installs everything at once.
Tips
- Pro tip: Use
%pipand%condamagic where available to keep installs tied to the kernel. - Pro tip: For reproducibility, pin versions:
pandas==2.2.3. - Security: Only run notebooks you trust; code cells can execute shell commands.
- Speed: Move heavy preprocessing to GPU/TPU or vectorize with NumPy/Polars to reduce Python loops.
- Organization: Name notebooks like
YYYY‑MM‑DD_topic.ipynbso Drive sorts them automatically.
FAQs
Is Colab really free? Yes, the Free tier includes CPU and opportunistic access to GPUs/TPUs with shorter sessions. Paid tiers prioritize faster hardware and longer runtimes.
How long do sessions last? It varies by tier and activity. Expect automatic disconnects after idle periods or very long runs. Checkpoint to Drive frequently.
Can I use R or other languages? Python is first‑class. You can call system tools via !, and some community kernels support R, but support is primarily Python‑focused.
Can I run jobs in the background? No persistent background jobs; the VM pauses when the session ends. Use Drive to save outputs and consider moving production workloads to a managed service when you outgrow Colab.
What’s the difference between GPU and TPU? GPUs are general‑purpose accelerators great for PyTorch and TensorFlow. TPUs target TensorFlow/JAX with massive matrix math throughput for certain models.
Summary
- Create a new notebook and learn cell basics.
- Choose CPU/GPU/TPU under Runtime.
- Install required packages at the top.
- Mount Drive and load your datasets.
- Organize outputs in Drive folders.
- Monitor RAM/disk; restart when needed.
- Share safely and use forms for parameters.
- Export to
.ipynb,.py, and.htmlfor reuse and sharing.
Conclusion
Start in the browser, pick the right accelerator, mount Drive, and pin your package versions. With that base, you can train models, analyze data, and share results without local setup. Expect occasional restarts; your workflow stays safe if you consistently write data and artifacts to Drive. When you need steadier compute or advanced GPUs, upgrade to a paid tier or migrate mature workloads to a dedicated service.
Discussion (0)
Be the first to comment.