anky209e / doteye-mlops

Image segmentation

Nuclie Image Segmentation

Download weights and put the tar file in webapp/seg_classes.

segmentation folder contains files related to all data presprocessing,model building and training process.

webapp folder contains django app deployment of model

You can also download weights from git lfs

Dataset used Data Scince Bowl 2018 kaggle

This Dataset contains medical images of cell containing nuclie and different masks of various part of Images.

Training Files

• model.py

  • This file containes custom UNET implemented from scratch using pytorch.

  • UNET is famous for segmentation of medical Imagery.

  • Research paper

• dataset.py

  • This file contains implementation of custom dataset class for training purpose.
  • We only used a little data augmentation.

• utils.py

  • As the name suggest this file contains all the utility function required.
    • save_checkpoint for saving parameters
    • load_checkpoint for loading parameters
    • iou_ for calculating intersection over union score
    • check_accuracy for checking pixel accuracy
    • And some little functions for saving mask and images

• train.py

  • This file handles the training process completly
  • To retrain model:

  # Hyperparameters
  LEARNING_RATE = 1e-3
  BATCH_SIZE = 16
  NUM_EPOCHS = 20
  NUM_WORKERS = 2
  PIN_MEMORY = True
  LOAD_MODEL = False
  DATA_PATH = "stage1_train"
  DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

  Run train.py after downloading datset

• This code already includes tensorboard integration to view graphs run:

  tensorboard --logdir runs

• This model was trained for only 20 epochs.It can be improved more using more training.These were the results after 5 epochs.

Django Webapp Files

• This folder is structured according to django for deployment of Model.

• webapp is the core django project containing django-app named segment handeling ML files.

• for more info run python manage.py help

Containerization and Orchestration

• This application is containerized using Docker.
• We are using docker-compose for orchestration of containers.
• For running webapp:

  docker compose -f docker-compose.yml up --build

• For stopping the aplication:

  docker compose -f docker/docker-compose.yml down

Note: Make sure you have installed both docker and docker-compose

WARNING: As this was developed in just 2 days.There can be bugs and erros.Please make sure to open a issue if you face a bug or error.

doteye-mlops's Issues

The model Weight link expired.

The model Weight link expired.
This is the link i have been trying https://drive.google.com/file/d/1-PkmPa8m6jVXFJkUOmO3Df8fmEmcwwKD/view?usp=sharing