Use a free GPU in your favorite environment

I want to use GPU but I can't buy it because it's expensive ...

If you are studying machine learning and have a GPU, you can study faster and more efficiently. But it's too expensive to buy ...

There is also a Google Colaboratory, but I just can't become a Jupyter Notebook and want to work in my usual environment (a little selfish).

At that time, I found this article!

Access Colab with ssh and use GPU

Get auth token from ngrok

Go to ngrok and open LOGIN-> Authentication-> Your Auth token. Then, the auth token is displayed on the screen below, so copy it.

Run ngrok on Google Colabora try

Open Google Colab, open Runtime-> Runtime Type Settings, and set the hardware accelerator to GPU.

Paste the following code into a code block and execute it. At this time, paste the auth token you copied earlier in the place of YOUR AUTH TOKEN on the 13th line.

# Install useful stuff
! apt install --yes ssh screen nano htop ranger git > /dev/null
# SSH setting
! echo "root:password" | chpasswd
! echo "PasswordAuthentication yes" > /etc/ssh/sshd_config
! echo "PermitUserEnvironment yes" >> /etc/ssh/sshd_config
! echo "PermitRootLogin yes" >> /etc/ssh/sshd_config
! service ssh restart > /dev/null
# Download ngrok
! wget -q -c -nc https://bin.equinox.io/c/4VmDzA7iaHb/ngrok-stable-linux-amd64.zip
! unzip -qq -n ngrok-stable-linux-amd64.zip
# Run ngrok
authtoken = "YOUR AUTHTOKEN"
get_ipython().system_raw('./ngrok authtoken $authtoken && ./ngrok tcp 22 &')
! sleep 3
# Get the address for SSH
import requests
from re import sub
r = requests.get('http://localhost:4040/api/tunnels')
str_ssh = r.json()['tunnels'][0]['public_url']
str_ssh = sub("tcp://", "", str_ssh)
str_ssh = sub(":", " -p ", str_ssh)
str_ssh = "ssh root@" + str_ssh
print(str_ssh)

If the execution is successful, the following will be displayed.

WARNING: apt does not have a stable CLI interface. Use with caution in scripts.

ssh [email protected] -p XXXXX

Open it with VSCode and try it out

Try sshing to ssh [email protected] -p XXXXX using VS Code's Remote-SSH feature. You will be asked for a password, but you can type password specified in! Echo "root: password" | chpasswd.

Make main.py etc. appropriately and check that GPU can be used with the following code

import torch

print(torch.cuda.is_available())
# True

I was able to use it!

Verification

CIFAR10 I verified it with the task of learning. The code I actually used is here.

The following environment was used for comparison, and it was carried out with 5 epochs.

It was lazy that I couldn't match the torch version.

Result (never correct as it is only run once)

The speed is about 20% faster!

Finally

In Colab, environment construction such as Pytorch and Keras is pre-built, so that is also convenient!