Train on Google Colab

Introduction

Today, I was training my model for the CMI - Detect Behavior with Sensor Data competition, and I saw that my MacBook was becoming so hot. Also, I had other things to do with it simultaneously. I thought Google Colab gives us approximately 4 hours of GPU. Also, my models aren’t that deep and don’t require so many resources and time. So, I decided to train my model on Google Colab as much as I can, and if my free limit is over, I switch to Kaggle or my MacBook.

Setup a notebook

The first thing that I did was to make a notebook in my GitHub repository. I have already given access to Google Colab in order to see my private GitHub repositories. After that, I loaded the notebook into Google Colab. Now, it is time to clone the whole GitHub repository To do that, I first created a token in developors section in Github settings. Then, I made a secret in colab with the name of: GITHUB_TOKEN. After that, I used the code below to clone the repository and go to the main directory of it, also install the requirements that I need in Google Colab.

from google.colab import userdata

!git clone https://{userdata.get('GITHUB_TOKEN')}@github.com/LiterallyTheOne/cmi_sensor_kaggle

%cd cmi_sensor_kaggle

%pip install -r train_on_colab/requirements.txt -q

There are two packages that should be installed prior to the other packages that are already installed in Google Colab. These two packages are: pytorch-ignite and hydra-core Which I put them in the train_on_colab/requirements.txt

Then, to use tensorboard, I used the code below (I save tensorboard logs in tb-logger):

%load_ext tensorboard
%tensorboard --logdir=tb-logger

Then, I should run my training procedure. I already have them, the only thing that I should do is to set them up in a way that can be run on a notebook. At first, from the script that I already have, I import main, register_conifg like below:

from scripts.run_tof_all_normal_ignite import main, register_config

In my script, I used decorator in order to work with Hydra. But it is not supported in a notebook. So, I should make an instance of my hydra config as they suggested in their website:

from hydra import compose, initialize

register_config()

with initialize(version_base=None):
    cfg = compose(config_name="config")

Now, I can run my main function like below:

main(cfg)

Then, after everything is finished, I zip checkpoints and tensorboard logs to be able to download them on my laptop with the code below:

!zip -r combined.zip checkpoints tb-logger

Final Thoughts

Google Colab is a really great and powerful tool to train and test our models. It has so many great options, and the most useful deep learning packages are already installed on it. I am really grateful that Google gave us this masterpiece for free. Right now, my plan is to train my models on Google Colab. Then, download them and submit them to Kaggle.