The Human Protein Atlas Image Competition on Kaggle

-
Just before the RSNA Penumonia Competition drew to a close, another interesting competition opened up, namely the Human Protein Atlas image classification challenge. Link: kaggle.com/c/human-protein-atlas-image-classification

The goal of the competition is to detect cell features from an image taken by a microscope. Around 30,000 512x512 images are provided which is the training set and around 11,000 images are provided as a test set. The labeled data is provided in a csv file.

I started off with the data exploration kernel to have a feel for the data. Then through the discussion forums I found a useful starter code that used transfer learning. Link: https://www.kaggle.com/mathormad/inceptionv3-baseline-lb-0-379

Instead of inception network, I decided to use ResNet50 network with the imagenet weights. I also tested results on both augmented data and the original one. The augmented data was giving me a boost of at least 0.05-0.06 on the public leaderboard. 

An interesting concept I learned from that code was the "warming up" idea for the transfer model. When a model is initialized, the weights are usually initialized with random/Xavier/He initialization. However, in transfer learning, we want to keep the imagenet loaded weights safe from this initialization. Therefore we initially declare all except the base transfer model layers trainable, train it for some epochs, and then make all layers trainable and start training on that.

I initially trained all the data on a 224x224 input size since it was the default size for using imagenet weights. After training for multiple epochs with it, I decided to use the 512x512 sized images.

There was however an issue. The network I was using was not going to have the same number of parameter in the layers if I decided to use the 512 sized images with the weights trained on the 224px ones. The trick was to insert an additional conv2d layer in between the conv2d and the flatten layer so that the number of dense parameters remained the same. I also named the model layers and used the by_name = True flag in load_weights function so that the weights loaded to the right layers.

I wanted to use the k-fold technique to get better parameters for the model. However due to lack of time, I only used it a couple of times to make sure the model was working correctly and getting decent results.

I'm still working on this competition and I'll update my results and other findings when I'm done.

Comments

Post a Comment

Popular posts from this blog

First ever Kaggle competition - Pneumonia Detection

-
Image
Github here:  https://github.com/varunvprabhu/pneumonia_detection I'd never participated in a Kaggle competition before. So I decided to join one, namely, the RSNA Pneumonia Detection challenge . Briefly, the competition was about developing an AI algorithm that would assist radiologists in pneumonia detection. The dataset training and test images were provided by the competition organizers through Kaggle. Approximately 28000 training images and 1000 test images were provided. Some of the 28000 images had bounding boxes of the locations of pneumonia detections in chest x-rays.(Specifically 8964 images) The 1000 images were un-labeled and one had to post the results of the algorithm after the images were run through that. A score is calculated using the mean taken over the individual average precision of each image in the test dataset. There were 2 ways to go about this. One was through object detection and the other was through segmenta...
Read more

Densenet, Pneumonia Detection, Activation Maps and AUROC

-
Image
I came across C hexNe t when I was implementing the Pneumonia Detection model for the RSNA Kaggle challenge. Although I didn't implement Chexnet for the challenge, I was interested in doing that in the future. Once I was done with the Protein Detection challenge, I started looking into that. Chexnet is basically Densenet, implemented for detecting various pathologies in Chest X-rays. Hence the name Chexnet. Densenet is a popular neural network architecture, along the lines of ResNet, Inception etc. The dataset used for Chexnet was the NIH dataset .  The paper implementing Chexnet tested their model specifically on detecting Pneumonia like features. They compared the model's performance against some radiologists, the F1 score being the determinant of performance. Going by the F1 score alone, it seemed as though the model performed better than the radiologists. But the devil's is in the details. I came across this excellent post by a radiologist/ml engineer. H...
Read more

Simple GUI for Model Inferences

-
Image
Running inferences by loading models via scripts is cool but the end user may not be familiar with these methods of obtaining predictions.  If deep learning is to be useful as a tool to the user, then it should be available in the form of a GUI, either web based or desktop based.  Keeping that in mind I decided to see what tools are available for me to create a GUI on the desktop that will allow me to: 1) Load a model 2) Feed that model inputs via a file explorer 3) Obtain inference results and display it via labels. Display any additional data as required. Since I was programming on the Windows Desktop, I decided to check out the machine learning offerings available via the Visual C# IDE. I found out that Microsoft's CNTK was available on C# so I decided to give that a go. I already had a pretrained model that I'd used for pneumonia detection (I trained and tested it in my previous post). I now needed a way to use that model in C# via CNTK. Luckily, there ...
Read more