For users without local GPU resources, Colab is an available solution. It could be transformed into a GPU instance with full SSH access.
This document refers to this tutorial.
First of all, create a Colab notebook using your Google account and use the GPU runtime mode.
Ngrok is used to make ssh forwarding. Before this, please create a password. Then, let's start sshd
.
Let's setup and run sshd
.
! apt-get install -qq -o=Dpkg::Use-Pty=0 openssh-server pwgen > /dev/null
! echo root:$password | chpasswd
! mkdir -p /var/run/sshd
! echo "PermitRootLogin yes" >> /etc/ssh/sshd_config
! echo "PasswordAuthentication yes" >> /etc/ssh/sshd_config
! echo "ClientAliveInterval 30" >> /etc/ssh/sshd_config
! echo "ClientAliveCountMax 5" >> /etc/ssh/sshd_config
get_ipython().system_raw('/usr/sbin/sshd -D &')
After that, we can install and run Ngrok, which creates a TCP tunnel. You can get authtoken from https://dashboard.ngrok.com/auth".
! wget -q -c -nc https://bin.equinox.io/c/4VmDzA7iaHb/ngrok-stable-linux-amd64.zip
! unzip -qq -n ngrok-stable-linux-amd64.zip
import getpass
authtoken = getpass.getpass()
get_ipython().system_raw('./ngrok authtoken $authtoken && ./ngrok tcp 22 &')
Now, You can access your server through ssh
command in your local machine.
The TCP address and port number can be found through the Ngrok interface https://dashboard.ngrok.com/status.
ssh root@[tcp_address] -p [port_number]
These experiments are run under the Colab default setting(T4).
-
PyGaggle: Baselines on MS MARCO Document Retrieval: It takes about 4 hours to re-rank each half subset.
-
PyGaggle: Neural Ranking Baselines on MS MARCO Passage Retrieval:
- Re-Ranking with monoBERT: It takes about 50 mins to re-rank.
- Re-Ranking with monoT5: It takes about 40 mins to re-rank.