No custom image to build
pandas and scikit-learn are already installed, so there is no Dockerfile and no docker build.
JupyterLab 4.6.3 Data
Free credit available Limited time
PyTorch Jupyter Notebook torch, torchvision and torchaudio, installed before you signย in.
Selfhost.dev runs a PyTorch Jupyter notebook as a one-click template: torch, torchvision and torchaudio preinstalled in JupyterLab, on a dedicated server we provision, secure and run. It runs on the CPU and keeps your saved models.
Open a notebook and run import torch. Nothing to download first.
Jupyter + PyTorchone-click template ready
- librariestorch 2.14.0, vision, audio
- alsopandas, scikit-learn
- sign-inyour password, 12+
- computeCPU
- saved modelskept on restart
European 8 GB server$0.03/hr
In one line
What is a PyTorch Jupyter notebook on Selfhost.dev?
What it is
Jupyter + PyTorch is JupyterLab with torch, torchvision and torchaudio added to numpy, pandas and scikit-learn. It runs PyTorch notebooks on the CPU, on a server of your own.
Where to start. Open the console with Jupyter + PyTorch selected
Who it is for
People following a PyTorch course, researchers sketching a model and developers testing an idea in a notebook that stays up.
When to use it
When Jupyter keeps answering No module named torch and you would rather skip the environment work.
Server
PyTorch Jupyter notebook, with torch already installed
PyTorch's own steps for Jupyter are pip install torch torchvision torchaudio or a conda line with cpuonly, then ipykernel so Jupyter sees that environment. Skip one and the import fails with No module named torch. Here all three packages live in the kernel you open.
The stack, end to end
Callers
Your browser
Edge
HTTPS
Your project server
Container
JupyterLab
Kernel
Python
Library
torch + vision + audio
Kept
Home folder
To sign in
Your password
| PyTorch in Jupyter by hand | PyTorch Jupyter notebook on Selfhost.dev |
|---|---|
| pip install torch torchvision torchaudio | All three installed |
| Pick the CPU wheel with --index-url | The CPU-only image, chosen for you |
| python -m ipykernel install --user | The kernel you open has torch |
| Change kernel, then restart it | import torch works in cell 1 |
Only your password opens JupyterLab and its PyTorch kernel. Deploy the stack
Container
PyTorch Docker, already running in JupyterLab
PyTorch Docker guides pull pytorch/pytorch, add a -v mount so work survives, then write a custom Dockerfile to add pandas and scikit-learn. This template is that container finished: JupyterLab on HTTPS, behind your password, with both libraries in place.
No volume flag
Your home folder keeps notebooks and saved models across restarts without a -v mount.
No terminal to keep open
JupyterLab answers on its own HTTPS address, not on a port you map by hand.
What you need
PyTorch Jupyter notebook requirements
The console starts this template on an 8 GB server, on x86 or ARM. Project servers have no GPU, so PyTorch runs on the CPU, which suits coursework, small models and inference with saved ones. A domain is optional.
What it needs
- Server
- 8 GB, the console's pick; x86 or ARM
- Processor
- CPU only: courses, small models, inference
- Password
- At least 12 characters, chosen at deploy
- Domain
- Optional; a generated HTTPS address is ready first
Server size: 8 GB, enough for Jupyter + PyTorch. Deploy on an 8 GB server
Quickstart
How to start a PyTorch Jupyter notebook in 3 steps
Launch, sign in, import torch.
-
Launch the template
Launch opens Jupyter + PyTorch. Set a password of at least 12 characters and accept the no-mining terms. Before a first top-up, a platform admin approves it. The template deploys in about twenty minutes.
-
Sign in
Visit the HTTPS address the console gives you and type your password into the Password or token box; no username is needed.
-
Import torch
Start a Python notebook and run import torch. torchvision, torchaudio, pandas and scikit-learn are ready beside it.
Deploy Jupyter + PyTorchStep 1 starts in the console.
Launch Jupyter + PyTorch Billing
PyTorch Jupyter notebook cost
A PyTorch Jupyter notebook here is one server billed hourly from prepaid credits; torch, torchvision and torchaudio add nothing. On a European 8 GB server that is $0.03/hr, close to $23 a month.
| Item | Cost |
|---|---|
| The Jupyter + PyTorch template | $0 |
| Bandwidth to your notebooks | $0, included in the server price |
| Optional daily backups | +20% of the server price |
| Datasets and models you store | $0, on the server's own disk |
| Server | RAM, cores, disk | Where | Rate | About a month |
|---|---|---|---|---|
cx33 (recommended) | 8 GB, 4 cores, 80 GB | Europe | $0.03/hr | $23.36 |
cax21 (ARM) | 8 GB, 4 cores, 80 GB | Europe | $0.04/hr | $29.20 |
cx43 (room to grow) | 16 GB, 8 cores, 160 GB | Europe | $0.06/hr | $43.22 |
ccx13 | 8 GB, 2 cores, 80 GB | US | $0.16/hr | $119.28 |
cpx32 | 8 GB, 4 cores, 160 GB | Singapore | $0.19/hr | $135.63 |
Monthly figures assume 730 hours. The US and Singapore servers need a first top-up.
If your balance reaches zero
With a card on file or a past top-up, Jupyter + PyTorch runs for 5 more days after your balance reaches zero, then stops. Top up within 24 hours to restart it; after that it is deleted, with a 14-day copy you can restore. Welcome-credit-only servers are deleted at zero.
Price a 16 GB server for larger image datasets in the calculator, no signup.
Libraries
torchvision and torchaudio, in the same kernel
torchvision brings datasets, model architectures and transforms for images. torchaudio brings transforms for sound. Both sit in the same kernel as torch, ready to import.
Images
Load a dataset with torchvision.datasets and shape it with torchvision.transforms in one notebook.
Audio
Turn waveforms into spectrograms with torchaudio.transforms before they reach a model.
Classic machine learning
Use scikit-learn for a baseline and pandas for the tables around your tensors.
Status
Five checks before your first tensor
- Every service is health-checked every 60 seconds.
- A stalled deploy is retried automatically, without you starting over.
- Done means the URL answers, not just that a container started.
- Locked from the first request. The password you set at deploy is the only way in.
- You see the release it runs. The console flags Update available when a newer one ships.
~20 min to deploy
The console estimates about twenty minutes for the Jupyter + PyTorch template after the server is ready; first-time server provisioning adds time.
Notes
PyTorch Jupyter notebook limitations
- CPU, not GPU. Project servers have no graphics card. The template fits courses, small models and running saved ones, not training large networks.
- Shared by everyone with the password. There is one workspace and no user accounts. Give each person who needs privacy a separate deployment.
- The first deploy takes about twenty minutes. torch and its two companions make a large image, so the console allows the extra time.
Nothing here in your way? The template deploys in about twenty minutes.
Launch Jupyter + PyTorch Planning
Running a PyTorch course or lab?
Tell us how many notebooks run together and the size of your datasets. We reply with the server we would pick and its monthly cost.
Help
Frequently asked questions
How do I install PyTorch in a Jupyter Notebook?
On this template you do not need to. torch, torchvision and torchaudio are already installed in the Python kernel, so import torch works in the first cell. Elsewhere you install them with pip or conda into the kernel's environment, then register that environment with ipykernel.
Does Jupyter + PyTorch use a GPU?
No. Project servers have no GPU, so the template uses the CPU-only image and every model runs on the processor. That suits coursework, small networks and predictions from saved models. torch.cuda.is_available() returns False here; pick a bigger server for larger data.
Which packages come with Jupyter + PyTorch?
torch, torchvision and torchaudio, plus numpy, pandas and scikit-learn from the scientific Python stack. Add anything else from a notebook with pip install --user. Those packages land in your home folder, so they are still there after the server restarts.
Do saved PyTorch models survive a restart?
Yes. Save a model or its state_dict into your home folder with torch.save and it is still there after the server restarts, along with your notebooks. Keep another copy of any model that took hours to train, in a repository or a download.
How is Jupyter + PyTorch different from JupyterLab?
It is the same JupyterLab with torch, torchvision and torchaudio added to the scientific Python stack, so the image is larger and the console allows about twenty minutes. JupyterLab hosting is the lighter template without them; Jupyter TensorFlow is the sibling for Keras.
Why might a new Jupyter + PyTorch wait for approval?
A platform admin approves each new Jupyter + PyTorch on an organisation that has never topped up. The name is reserved while it waits; you hear the decision by email. A top-up removes the wait for new ones. Crypto mining is banned on every server. Deleting the service withdraws the request.
PyTorch, ready in JupyterLab. import torch and go.
- torch, torchvision and torchaudio in the Python kernel
- pandas and scikit-learn beside them
- A European 8 GB server, about $23 a month
About Selfhost.dev
Selfhost.dev is a managed self hosting platform for developers and small teams. Open source apps and databases on dedicated servers you control, billed by the hour, in six locations across three continents.
Best for: PyTorch notebooks next to the data and apps they serve, a vector database for embeddings and managed PostgreSQL, MySQL and Redis.