Paddy Doctor: Paddy Disease Classification

-

Introduction:

This time, I don't want to write the hurdles in fastbook and how to overcome them because once you move to notebook 6 or 7 it shows that you have the strong mindset to overcome by yourself and finish the book. Here after we have to practise what we learn by participating in Kaggle Competitions. 

Problem statement:

Identify the type of disease present in paddy leaf images.

Understanding the dataset

Labels are given in a separate csv file, train data set contains 10 folders each containing an image of paddy disease with respect to folder name, test data contain the images we have to predict.

Model Implementation

Once we Explore the dataset, the next step is to build a data block for the problem. We can create using two ways they are:

  1. We may map the label of the disease name to the image by csv file.

  2. Or we may label the data according to the folder name.

What I chose is mapping the label of the images with respect to folder names. Once the data loaders are ready we can check the train and validation set. If it shows no error, our data block is built well. For this problem I tried 3 different models, they are Efficient B0, Resnet34, Resnet18. I tried everything that I learned till now with these models, like discriminative learning rate rate, unfreezing and so on.

Once the model is working, predict the label with test data. And prepare the data for the submission file and submit it.

Problem faced and how I overcame

Building a model with resnet architecture comes handy for efficient net not being able to use. I searched through the fastbook and the notebook of great people of fastai. I got to know about the timm and the issue was solved.

Why I am saying this, because once you start working on a problem you may face some issues which you may never come across. This type of issue helps to build your knowledge and problem solving abilities.

Conclusion

The idea of the notebook is not to build a great model, it is to build an end to end model for the kaggle and implement what we learn till now. This type of solving problems in the Kaggle give us a mental boost to complete the book.

Note:

You can check out my notebook of this competition in Kaggle. And also my Github for more datascience related projects and works.

Comments

Post a Comment

Popular posts from this blog

Best practice to follow on git and github

-
Some basics: linux Note: <> just used to differentiate the file for understanding : ls -list the file mkdir <folder name> - creach the file name touch <file name> - create the file code . - take to vs code vi <filename >- open the vim editor - Don’t dive deep into vim editor There are two mode Edit mode and command mode You can press Esc to enter cmd mode to give command like insert delete update and so on  i - for insert a text in a file O, o move the cursor above and below :wq - write and save the file These are the basic about vim editor but don’t spend too much time on this use other editor for faster progress and so on We can directly open the file and made the changes Vim is fun to learn cat <file name > - show what are there in that file rm -rf <file/folder name> -remove the the file or folder from the current directory cd - change the directory(where we are working and our file path) cd .. go back the previous folder in a directory Git confi
Read more

Deep learning with TensorFlow

-
Image
After learning the Sci-kit learn library, I started to think What, Where,  How should I start learning deep learning which library should go on, I find myself Tensorflow is easy to begin with because TensorFlow is developed by google and they providing a stable and extensive documentation and so on... While learning Image classification I came to know the power of GPU, Everything comes in colab which is very much handy. Main thing I learned about deeplearning is You can do anything as long as you can convert into object into Tensors (Array of numbers) Classification of 101 image of food data image data is done with EfficientnetB0 model using transfer learning. There I, learned the power of Transfer Learning. TensorflowHub providing various Pretrained model, we can use based on our need.  And there is something called Tensorboard whwere we can the evaluate the various model, we build with deeplearning. I attached the sample image of Tensorboard.  And there are so many resource in the in
Read more

Image classification(Fastbook notebook_2)

-
Image
Hi Friends, What i am so excited about notebook2 is we are going to deploy a model, which helps to understand the flow of deep learning. I am going to discuss the problem I faced, may be you may face in your journey. First and foremost thing is you may not be able to get the bing api. If you are getting it, It's ok but if you face any difficulties means use DuckDuck go api which comes very much handy. Once you downloaded the dataset Everything worked fine until the deployment part.  What I would suggest is if you are using the colab like me means you may face difficulties in using viola. What I will suggest is to go for Gradio which is much easier. With an few line of code you may built a amazing application. My project is about classication of great ape(Chimpanzee, Bonobos and Gorilla) eg: In case you may face any difficulties go through my notebook . Feel free to ask the doubts.
Read more