Hi everyone, after lesson 2 I created my own image classifier to differentiate between 10 different medications: Allopurinol, Atenolol, Ciprofloxacin, Levothyroxine, Metformin, Olanzapine, Omeprazole, Oxybutynin, Prednisone, Rosuvastatin. The specific strength of each med is noted in the notebook
As a former pharmacist turned software developer, I thought it would be interesting to see how a ML model would perform.
I sourced images from US National Library of Medicine’s Pillbox and google images. As you can tell, google images included quite a bit of junk images.
I built a simple language model based on whatsapp chat data for part 1. I decided to go back, make some improvements and write a medium post about it all.
I’ve now completed part 2 so I wanted to see if I could go back and use what I’ve learned to add a custom rule to the tokenizer, as well as clearly explain what’s going on with the language_model_learner. With a bit of digging through the documentation and source code I was able to do it!
Here’s a writeup I did for a competition I participated in on the Zindi competitive data science platform. The object of the contest was to use remote sensing imagery from different timepoints to classify what types of crops were growing in fields with given boundaries. I used a U-Net style approach and I found a really nice library called eo-learn to help with the processing pipeline for the remote sensing data.
This is my second pass (had done V1 last year partially). Being and Indian, I had to do something with cricket. While we use both types of balls to play cricket, thought of asking machine to learn the differences (while we ignore those during the cricket session )
A big shout out to @melonkernelCHRISTOFFER BJÖRKSKOG and his post for cleaning up the wrong images and generating URL list via a bookmarklet
Did this using Google Collab
So… was able achieve 98% accuracy even before unfreezing the laters… so just sharing it here.
Just finished the second lecture. I made a model for classifying between East Asian traditional dresses - Kimono, Hanbok and Qipao. I was able to achieve 95% accuracy . On retraining the models several times, I kept on getting different Learning rate curves. I am not sure why that happens but I will be asking that on the forums after this post.
I used the tutorial on how to download pictures from google images to assemble a dataset, to train a so-called “Baby Monitor”. Its purpose is to take a picture of a baby in the crib and predict if the baby is sleeping, sitting or standing.
As you can see in the picture I got a 13% error rate, which is pretty good considering that I only used the defaults (the model was a pretrained resnet34 architecture) and my training set is around 170 images for three categories.
I have two questions:
In my experiments most of the time the validation loss was lower than the train loss. Why is this happening and is it that bad?
nicely done! Also nice medium article. I have never tried doing that, let me try that. I am also experimenting with Lesson 2 & 3 currently.
Few queries
Which tool did you use for images? How large was the final training/validation dataset? My experience was that out of 500 Google Images URLs for each class I only got 50-70 images which are valid, rest were deleted as invalid.
I am trying to make my model tell me “I don’t know” when someone gives an irrelevant image. Would you want to collaborate? (as you can see your deployment can also help from that so it can tell something is not even a dress )
Hey Dimitris
This happens with me as well. After searching the net for answers, I found that there are 4 main reasons for validation loss being lower than training loss as listed below. Usually this happens when you have a small dataset or you train for too few epochs as random error is high in such cases. We can expect a small margin of error due to reasons 1 and 2 also as mentioned below. My understanding is that if the validation loss is equal to or slightly more than training loss then it is the best scenario.
1: Regularization is applied during training, but not during validation/testing.
2: Training loss is measured during each epoch while validation loss is measured after each epoch. On average, the training loss is measured 1/2 an epoch earlier.
3: Your validation set may be easier than your training set or there is a leak in your data/bug in your code. Make sure your validation set is reasonably large and is sampled from the same distribution (and difficulty) as your training set.
4: You may be over-regularizing your model. Try reducing your regularization constraints, including increasing your model capacity (i.e., making it deeper with more parameters), reducing dropout, reducing L2 weight decay strength, etc.
Hi @jeremy,
I want do deploy Image Classification on Android using Pytorch Android API.
It requires .pt file, but using learn.save() i am getting only .pth file.
So how can i convert .pth file to .pt
The reason for the imbalance is that the dataset is actually 1/3 Normal, 1/3 Bacterial Pneumonia and 1/3 Viral Pneumonia. I recently tried classifying viral vs bacterial, parsing the filenames within the PNEUMINIA folder to get the labels.
Unlike NORMAL vs PNUEMONIA, where many have achieved accuracies in the high 90’s, I was only able to achieve around 77% accuracy for bacterial vs viral. I would be interested to learn if anyone has been more successful.
I used the Fatkun image downloader. It has been mentioned in the forums in another topic. My dataset consisted of 670 images.
Thanks for showing me the importance if having an ‘I don’t know’ class. I feel there are two ways of making the Neural Network say “I don’t know” :
Create a new class labeled as ‘unknown’ and fill it up with images of everything that is not related to your problem. However, this is very unrealistic.
Try to extract the features from the Neural network and pass the features through a Machine Learning classifier (Like a decision tree or SVM). This method however requires us to get some knowledge on Neural Network theory.
Hi everyone! After 5-6 months of intensive studies I deployed my first Deep Learning app. It is a personal project that I have worked on and felt it would be cool to deploy it. The app allows users to turn random pictures into paintings of 3 old masters Van Gogh, Monet and Cezanne. You can check out the app at photo2painting.tech
The app is a demonstration of how CycleGan models work in production and is deployed on a Digital Ocean droplet (free with Github as I am a student). The project is still on development, so I am eager for your feedback. Feel free to contact me if you want to collaborate on the project. If you guys want to check the code, I have open-sourced it here
Your app is marvelous, you have done a great job and have created a good application of the course content.
The app shows your 5-6 months of intensive study have been well spent!
)
