Human Protein Atlas Image Classification (kaggle)

Overview

Implementing human protein atlas image classification from kaggle using tensorflow.

Environment

conda create -n protein python=3.6 anaconda

activate protein
# or for unix
source activate protein

conda install tensorflow-gpu
conda install keras
pip install kaggle
conda update --all

Place the API credintials instructed here. https://github.com/Kaggle/kaggle-api

Go to the working directory you decide.

git clone https://github.com/meitetsu3/HumanProtain.git
cd HumanProtein
kaggle competitions download -c human-protein-atlas-image-classification

Extract the downloaded zip file.

Create input folder and move the extracted folders and csv files under input folder.

Preparing data

go to code folder which is assumed working directory. Open TFRecord.py with your favorite IDE / editor.

In this project, we are converting the image files to TFRecord files.

create input_tf folder under the working directory. Copy sample_submission.csv under the input folder, and name it test.csv. The run the TFRecord.py.

You will get 12 Train-?.tfrecord files containing more than 31k training images(each stacked array of 4 gray scale images with different filters) and lables (multi-lables). It also creates Test-.tfrecord file which contains all the test images.

Based on your system and preference, change the number of files you create.

Run CheckTFRecord.py to visualize some of the images from the tfrecord files. It shows 2 images with the first 3 channel (R,G,B) and another image swapping the 3rd channel B with Y.

image here:

First training

Reviewing the model performance

Predicting