I recently got scholarship from udacity in Facebook and Pytorch in Nano degree. In this post i will share some very useful details with links for you on solving images classification task for beginner using CNN in Pytorch.
In the lab challenge we are given flowers with 102 categories in total for classification with training and validation sets provided.
Before starting doing any deep learning project we need to select good platform for development with GPU enabled. There are many options GCP, AWS, paper space and crestle etc. for that. But they are paid but good for production based systems. As a beginner or learner if you may be sufficient with Google Colab providing you 1 GPU Instance completely free( A Good News Indeed.). But it has some limitations as well. I provides you session for working after which you may loose your unsaved data. Well it is still a good point to start from. Also having owned laptops or computers with GPUs also a good hand but you can work without it as well but need passion and with directions clear.
Now next step in solving problem is to solve first the lectures that is provided in lesson 4 and lesson 5 provided in Part 1 course.
Deep Learning with PyTorch
Those videos and material covered there surely will make you shine and provide you necessary confidence.
One important thing whenever you start this course make sure or it is suggested spend more time on lesson 4 and lesson 5.
After that you can start your lab 1 work.
First and foremost, do consult with documentation available specially https://pytorch.org/tutorials and pytorch forum https://discuss.pytorch.org that would golden eggs for you.
Course Students Community Resources:
As for the session that i attended included 10,000 students. So there are many passionate students shared there workings and helpful material for you to get help. I have organized few for you. They will surely would land you in safe zone.
As you are clear we solve this using CNN so there are many Architectures available i refer you here https://pytorch.org/docs/master/torchvision/models.html?highlight=resnet first. It is recommended to solve the problem using transfer learning a way of using pre trained neural networks.
There is a sheet available here where students mentioned their used architecture Architecture and Models Scores.
I started from resnet18 and it is good to start from here.
Next you should get help from Avinash’s medium post Tips for Cracking Challenge.
Also Jose Nieto shared his three blog posts for us. He shared his experience and good tips.
Now guys, i do not want it to be long post so lets conclude things. I want to share github repository and youtube videos sharing by some generous students. Here are few of those that i found useful.
https://github.com/udacity/deep-learning-v2-pytorch from Udacity
How to save and load models from one of our fellows https://www.youtube.com/watch?v=07obsF0bOUA here. Really worthy way of explaining.
Also remember you can use kaggle for solving the problem if you feel comfortable but on kaggle we have 1.x version of pytorch where on your own system or on colab as instructed by Avinash we used 0.4.x version. Also you feel easy to get quickly use with it. Voila! it is pleasure you read till here. But believe me you they are great sharing by my fellows and thanks to them once again, a wonderful community indeed.
If you enjoyed please comment and share it with your friends. Thanks a lot. Happy days Ahead.